The scratch of the FP32 prompt attention is batch x chunks x n_head x 258 floats (about 0.41 GB at 32 and 1.65 GB at 128 on
Flash-Next), independent of the prompt chunk. Prefill::init allocated it from STRATA_ATTN_BATCH (128 since 09524eb47) while the
estimate that sizes the expert cache and the prompt path's loan (bytes_needed_impl) still counted 32, so about 1.2 GB was never
reserved and a single 32 GB Arc Pro B65 with the config setup writes failed its first request ("device buffers for a chunk of 512
tokens do not fit (2098 of 32656 MiB free)"). Both now use one function (attn_batch_for), and the batch is capped at the chunk's
token count, since a chunk never runs more queries than it has.
The default goes back to 32: with the budget honest, 128 would take 1.2 GB from the expert cache on every Intel card for a
1.1-1.3% faster prompt on two B65s. STRATA_ATTN_BATCH still sets it (1..1024), and is now counted.
INTEL_ARC.md: one bullet with the numbers (2x Arc Pro B65: decode +17% to +23% over the serial --spec 4 setup, 5 pairs of 5),
what stays as before, and the two switches; one row in the tested cards table. MULTI_GPU.md: the Default line of
"Pipeline windows" now names the Intel Arc case beside AMD RDNA4 and NVIDIA.
On two Battlemage cards the engine pipelines the two cards' verify windows by default, and then answers as a 2-token window
whatever --spec says (pipelined --spec 4 equals serial --spec 2 in 100 of 100 agent requests; serial --spec 4 differs in 3).
So the generated config says --spec 2 there: the same answers, and about 20% faster than the old --spec 4.
strata_control/intel_spec.py holds the rule: the cards are found by PCI device id (the xe driver and an id in 0xE200-0xE2FF,
the id the engine's architecture enum comes from), never by name. It changes only setup's own default "--spec 4", only
with the draft layer on, and only for exactly two such cards. A --spec in "user_args", any other value, one card, an
A-series card in the pair, three cards and "mtp": false are left alone. Applied in three places: sycl/setup_intel.py when it
writes the run config of a layer split (the copy in "args_written" follows, so a rerun does not take it for a hand edit),
the launcher's derive_config (before the person's --set options, which win), and the Intel server's engine_args at every
start (a config an older setup wrote). Tests in tests/sycl/test_intel_two_card_spec.py.
When --pipeline-windows is not given, the engine turns it on for --serve with a layer split of exactly two stages whose
GPUs are both Battlemage (Xe2) cards by architecture id (strata::intel_gpu_gen), with the MTP drafter (--spec 2 or more),
no --batch, no --peer-device and no helper caches. A single card, an Arc A-series card, a mix, three or more stages
and any other vendor stay serial. --pipeline-windows N and STRATA_PIPELINE_AUTO=0 override it, and one log line says it
is on. The condition is a host-only function (strata/pipeline_auto.hpp) with a test that needs no GPU.
Measured on 2x Arc Pro B65: a 1 h mixed serve (1,082 requests) with 0 errors, 0 token differences against the serial run
of the same requests and flat VRAM; decode +17% to +28% over the serial --spec 4 default. The pipelined windows answer as
a 2-token window, so setup and the launcher write --spec 2 for such a config (next commit).
Which builds take 16 rows and why, how the 8-row build stays main's code, what was tested (real 16-row windows,
identity on seven boxes), the four-card profile with its measured effect per setting, and the opt-out.
For exactly four Radeon AI PRO R9700 cards (at least 28 GB each, hip, no --peer-device) running Qwen3.8-Flash-Next
IQ3_S, setup.py and the launcher fill in what the person has not chosen: parallel 32 with --batch-groups 4, the layer
split 13,26,39 (the output head costs about four layers on the last card), --prefill 2048 (in place of the "auto" setup
writes), and STRATA_QSA_SLOTS, STRATA_BATCH_SELF_COMMIT and STRATA_SLOT_COPY_DIRECT in the run config's env. A setting
of the person's own always wins; STRATA_FOUR_CARD_PROFILE=0 or "four_card_profile": false turns it off. The prompt cache
stays on (agent follow-up turns wait 0.8 s for the first token instead of 1.5 s).
Through the launcher's configuration on the 4 x R9700 box (total tok/s, warm round, parallel 4 before): 1 request 66.6 ->
67.9, 8 at once 156.6 -> 261.2, 32 at once 155.4 -> 369.6, four 7.9K prompts out 24.3 -> 35.0 and in 1892 -> 2706.
batch_test with the profile's flags: 32 of 32 slots identical to solo; a 25 minute soak of 32 clients: 0 errors.
serve/server.py no longer says "at most 8 requests" for a layer split (each pipelined group holds its own slots).
Tests: tests/terminal/test_amd_four_card.py.
A pipelined group of a layer split is its own window, so 32 slots on four cards were four groups of 8 and nothing
larger could run. The window limit is now a build option (STRATA_WINDOW_MAX_T = auto | 8 | 16; auto is 16 for HIP,
8 for CUDA; SYCL and Metal never set it): include/strata/kernels/window_limits.hpp.
A window of up to 8 rows runs exactly the kernels it ran before (HIP gfx1201 window 8 against main: 3096 of 3097
kernels identical, the other a debug variant; CUDA sm_86 and sm_60: identical or renamed but for a few prefill kernels
NVVM compiles differently, with identical results). The code for 9 to 16 rows is only reached by a wider window:
- per-row arrays of the verifier, the expert source tables (128 / 256 entries) and the remote / peer expert tables follow
the window; CPU expert jobs stay at 8 tokens (a 9..16-row group is two jobs over the same expert);
- the hyper-connection read is made in slices of 8 rows (its Q8_0 launch would otherwise skip rows);
- the grouped expert kernels run a second pass (SKIP template argument) instead of a 16-entry shared array (51 KB,
over the 48 KB static limit of CUDA);
- resident_plan runs on 128 threads (up to 8 rows, as it was) or 256; MMVQ, bf16 GEMV, shared expert and MoE combine
take 9..16 columns only in a wide build; the sm_60 packed path runs wide calls in passes of 8;
- STRATA_P4_TIMING and STRATA_P4_TRACE (include/strata/program/pipe_probe.hpp): opt-in timing and trace of the
pipelined windows, off unless set.
Tests: src/core/window_limits_test.cpp (target window_limits_test). Real 16-row windows (--batch 64 on four stages,
--batch 16 --batch-groups 1) decode the tokens they decode alone.
The engine's host threads have no autorelease pool, and Metal parks one extra
reference per new MTLBuffer, command buffer and encoder in the calling thread's
pool. Nothing drained it, so every buffer Context::alloc made (a 16 KB job
array per expert matvec, ~90K a request) and every stream command buffer and
encoder stayed alive after free(): +1.3 GB, +90K Metal regions and a slower
queue per request, until a big Mac slowed to a crawl.
- Context::alloc and the fallback wrap: create the buffer inside an
autoreleasepool (DS4 runtime only; the Qwen build is untouched).
- strata::gpu runtime: a pool around command-buffer, compute-encoder and blit
creation; the compute encoder is retained and released in end_encoder.
- Event marks: an event recorded again, or destroyed, before it was waited on
now drops its command buffer token (ds4_ctx_drop / Context::drop_mark).
- route, combine and sparse_attn free the stand-in buffers they allocate for
absent inputs.
On an Arc A750 the lane-cell score phase read a 4,106-token prompt 0.7% slower than the unchanged kernel (three
interleaved runs against the published 0.1.52 engine; with STRATA_ATTN_LANECELL=0 the branch matched it), while it is
+5% on the Arc Pro B65. The default now follows strata::intel_gpu_gen(): the runtime's architecture enum first, the
device name only as a fallback. STRATA_ATTN_LANECELL=1/0 still forces either way on any device.
Problem: the timing lambda measured milliseconds per iteration and returned them, while the printout labels the value
"us" and derives its GB/s as `bytes / us / 1e3` - so every number this bench has printed was 1000x too small in time
and 1000x too large in bandwidth. A launch that moves 19.7 MB in ~140 us was reported as 0.1 us and 148,713 GB/s, which
is why the tool looked broken rather than mislabelled. The wall clock always agreed with the real work (200 iterations
618 ms, 2,000 iterations 3,566 ms).
Result, 2x Arc Pro B65, shard 1 (Qwen3.8-Flash-Next IQ3_S experts, 10 groups), before and after the one-line fix:
| case | before | after |
| --- | ---: | ---: |
| 2 tokens | 0.1 "us", 148,713 "GB/s" | **143.4 us: gate/up 84.5 us (148 GB/s), down 59.4 us (78 GB/s)** |
| 8 tokens | 0.3 "us" | **340.8 us: gate/up 217.4 us (58 GB/s), down 126.0 us (37 GB/s)** |
and the model's own layers at two tokens: layer 30 iq3_xxs/q2_0 143.6 us, 31 iq3_xxs/iq4_nl 140.4 us, 32 iq2_s/iq4_nl
135.5 us, 33 iq3_xxs/iq4_nl 140.2 us - gate/up 131-149 GB/s, down 78-168 GB/s, results bitwise equal to the reference in
every case.
Approach: return microseconds (`ms * 1e3 / iters`), which is what the printout, the minimum-of-three logic and the
bandwidth formula already assume. No kernel, no timing path and no measurement logic changed.
Validation: the same run before and after; the wall-clock scaling that exposed the bug; bitwise equality retained in
all four model layers and both token counts. Not run: CUDA/HIP (this file is the port's own), other shards.
Limits: this fixes the units only. The numbers are per-call costs with the bench's own synthetic groups, so a rate taken
from it still has to be matched to the real window (48 calls per layer step).
Problem: the bench fills its output with 0xff and then compares the mode against the reference, so a launch that
returns without enqueueing leaves both sides bitwise equal on an untouched buffer and the tool reports "0 differing"
and a 0.1 us timing - a measurement of nothing that looks like a very fast kernel. The 0xff fill is NaN as a float, so
"no finite value in the reference window" is exactly that case.
Result, 2x Arc Pro B65, shard 1 layer 20 (`iq3_xxs/q2_0`, 10 groups, 2 tokens): the guard does *not* fire - the window
does write finite values, so this defect is not what makes the timing wrong. The timing is still 0.1 us per launch
against a 19.7 MB working set (0.3 us at 8 tokens), which scales with work but is 20-60x too fast for the launch to be
waited for; the remaining cause is in the timing loop's wait, not in the launch. That stays an open port item.
Approach: the guard only refuses and reports; it changes no timing path and no kernel.
Validation: builds and runs on the box; the run above is the only one taken. Not run: CUDA/HIP (this file is the port's
own), and the timing itself is still unproven - do not use its numbers.
Limits: this is a diagnostic, not the fix. `native_expert_bench` numbers remain unusable until the wait is found.
Problem: `--mtp-q4 head` builds the Q4_0 draft head by gathering the subset and dequantizing it row by row, but with
`STRATA_Q6_PLANAR` (the SYCL port's default for Q6_K heads) the source is stored as 16-byte-aligned planes and the
gather copies planes as if they were rows. The result is a draft head that proposes nothing at all, with no error:
decode falls from 73.8-75.6 tok/s to 35.3 tok/s (the non-speculative rate) and the DONE line reports "0 of 0" drafts.
Result, 2x Arc Pro B65, Qwen3.8-Flash-Next IQ3_S, 128 tokens of a 483-token chat prompt, one warm request each:
| configuration | decode | drafts |
| --- | ---: | --- |
| `--mtp-q4 head`, planar head (before) | 35.3 tok/s | 0 of 0 |
| `--mtp-q4 head`, planar head (after) | refused with "needs the packed Q6_K head layout (STRATA_Q6_PLANAR=0)" | - |
| `--mtp-q4 head` with `STRATA_Q6_PLANAR=0` | 69.5 tok/s | 54 of 63 |
| default head (no `--mtp-q4`) | 74.9 / 75.6 tok/s | 55 of 63 |
Approach: a guard in `make_q4_head` that refuses the option when the head really is planar, so the engine says why
instead of running with speculation silently disabled. The packed-layout path is unchanged and still works, which is
what the measurement above shows.
Validation: the guard message appears in the engine's own output; `--mtp-q4 head` with the packed layout still produces
54 of 63 drafts; the default path's tokens are identical to the pre-change reference (`compare_tokens.py`, no
difference) and its rate is unchanged. Not run: CUDA/HIP (the file is the port's own copy and the guard is behind
`STRATA_NATIVE_EXPERTS`), a non-Q6_K head, `--batch-mtp`.
Limits: this makes a broken opt-in fail loudly; it does not make the 4-bit head fast. With the packed layout it is
*slower* than the Q6_K head (69.5 against 74.9), so the option is still not worth enabling.
Problem: the prompt path's per-layer grouping tables - `[ids | slot | src]`, T x K int32s each, ~245 KB per layer at a
2,048-token chunk - live in mapped pinned host memory, and on the mapped arm the GPU reads them there, once per layer.
On a PCIe-attached pair that is 48 rounds of host-memory reads per chunk, and the port already carried the A/B arm to
copy them instead (`STRATA_GROUP_COPY=1`) without a measurement deciding between them.
Result, 2x Arc Pro B65, Qwen3.8-Flash-Next IQ3_S, `--pipeline-windows 2`, three warm repetitions, same build, back to
back:
| case | mapped arm | copy arm | gain |
| --- | ---: | ---: | ---: |
| prefill 4,106 tokens (AOT) | 1,303.2-1,303.5 tok/s | **1,332.8-1,337.5** | +2.5% |
| prefill 16,380 tokens (AOT) | 1,617.8-1,622.7 | **1,657.2-1,665.2** | +2.5% |
| prefill 10,938 chat prompt (AOT) | 1,575.6-1,578.4 | **1,613.8-1,616.8** | +2.5% |
| decode, chat prompt, 128 tokens | 74.9 / 75.6 | 74.9 / 75.6 | unchanged |
Approach: the default flips to the copy arm for this port; `STRATA_GROUP_COPY=0` restores the mapped one. Both arms
allocate nothing new when the buffer is already big enough, and the copy arm reuses the host vectors the code already
keeps for exactly this.
Validation: the token stream is identical between the two arms at both prompt sizes and on the decode run
(`scripts/compare_tokens.py`: first difference none), and identical to the pre-change reference (the rc-p16k run of
round 42). Not run: CUDA/HIP (the file is the port's own copy), a host-shared-memory device (an iGPU, where the mapped
arm may be the better one - hence the switch), a 32K prompt.
Limits: +2.5% on prefill only; the decode path does not use these tables.
Problem: `--pipeline-windows 2` (the two-card window overlap) ran but produced different tokens from
`--pipeline-windows 1`. Ten rounds of switches, splits and modes narrowed it to the two alternating verifiers, and a
state dump (STRATA_PIPELINE_DUMP: a hash of the first stage's GDN recurrence and conv history after every window)
found the moment: identical after the prompt, different after the first decode window.
Root cause: the serial path lets a one-token window advance the GDN state inside the window and skips the commit
graph for it (`record_window` + `commit`). That is safe there because a serial window never runs before its
predecessor's verdict. A *pipelined* window may: stage 0 runs window K+1 speculatively while window K's verdict is
still open, so a self-commit lands on state that is not final yet - and the first decode window of every request is a
single token.
Fix: a per-verifier switch (`set_self_commit_off`) that the pipelined verifiers set. With it `record_window` does not
self-commit and `pl_commit_async` always queues the commit graph - the configuration that already measured
token-identical (`STRATA_ONE_TOKEN_COMMIT=0`) but kept a global switch.
Result on 2x Arc Pro B65, Qwen3.8-Flash-Next IQ3_S, same build, back to back:
| case | pw 1 | pw 2 | tokens |
| --- | ---: | ---: | --- |
| short prompt, 256 tokens, spec 4 | 56.3 / 56.1 tok/s | **59.6 / 61.4** | identical |
| 4K context, 128 tokens | 55.6 | **60.3** | identical |
| `--spec 2 --spec-min-p 0`, 64 tokens | 49.2 / 50.3 | **59.6 / 63.2** | identical |
| 512 tokens, one run | - | **60.6** | - |
| prefill 4,106 tokens | 1,244.8 / 1,242.0 | 1,253.8 / 1,251.3 | identical |
| prefill 16,380 tokens | 1,548.1 | 1,561.1 | - |
So decode gains 7-9% at the default window and 21-26% at two-token windows, and the prompt path is unchanged or
marginally faster.
Validation: tokens identical in all four decode cases above (`scripts/compare_tokens.py`); the serial path's tokens are
unchanged against the pre-change reference. The state dump still reads one window behind in the pipelined arm (an
ordering artifact of where it copies), so the token comparisons are the evidence, not the dump.
Not run: CUDA/HIP (the files are the port's own), a 32K-context prompt, an overnight soak.
Limits: `--pipeline-windows 2` stays opt-in. It costs 278 MiB of card 0's expert cache (the second verifier and the
GDN snapshots) and its gain depends on the window size, so the default remains `--pipeline-windows 1`.
The pipelined state dump showed the first multi-token window leaving the GDN state one row ahead of the serial path,
so the next question was whether the commit count was wrong. With STRATA_PIPELINE_DUMP set, `pl_commit_async` now
prints `n_keep`, `T`, `pos0` and whether the one-token skip applies, which answers it: the counts are right.
```
commit n_keep=1 T=1 pos0=18 self_skip=1 (stage 0)
commit n_keep=1 T=1 pos0=19 self_skip=1
verdict p=19 T=1 a=0 B none
commit n_keep=1 T=1 pos0=19 self_skip=1 (stage 1)
commit n_keep=2 T=2 pos0=20 self_skip=0 (stage 0)
commit n_keep=2 T=2 pos0=20 self_skip=0 (stage 1)
verdict p=20 T=2 a=1 B launched ON
```
So the multi-token window commits two rows in both stages, exactly as the verdict says, and the extra row must come
from the *window's own* self-commit: `record_window` self-commits when `T == 1 && !batch_rec_ && !g_qfuse() &&
one_token_self_commit()` (verify.cpp:844), and the transition from a self-committed single-token window to the first
graph-committed window is where the state gains the row. That is the next thing to diff - the dump currently hashes
the recurrence and the conv history together.
Debug only: the print needs a state dump to be requested and changes no behaviour.
Problem: the serial commit knows that a one-token window advances the GDN state itself (`record_window`) and skips the
commit graph for it; `pl_commit_async`, which the pipelined windows and the pipelined prompt reader use, ran the graph
anyway. A one-token window therefore advanced the state twice. The first decode window of a request is always one
token, which is why `--pipeline-windows 2` produced different tokens from `--pipeline-windows 1`.
Evidence: the state dump added in the previous commit (STRATA_PIPELINE_DUMP, hash of the first stage's GDN state after
every window) was identical for both arms after the prompt (`pos 18 sum 1952.13774 hash b97305876cf73d2c`) and
different after the first decode window (`pos 19` 3222.063 in the pipelined arm against 2815.93065 serial). With
`STRATA_ONE_TOKEN_COMMIT=0` - which makes the serial path take the graph as well - the two arms produced **identical
tokens**, which pinned the asymmetry to this one branch.
Result: with the skip, the two paths' states agree after the first and second windows (`pos 19` and `pos 20` hashes
equal) and the tokens are identical up to and including token 5; the first difference moved from token 4 to token 6.
Validation: `--pipeline-windows 1` tokens are unchanged (identical to the reference), and the prompt path, which also
commits through `pl_commit_async`, did not move (4,106 tokens 1,239.7-1,248.9 tok/s, 16,380 tokens 1,552.1, the same
as before the change). Not run: CUDA/HIP (the file is the port's own copy), a 32K-context prompt.
Limits: this does not make `--pipeline-windows 2` token-exact. A second defect remains: after the first *multi-token*
window the pipelined state is one row ahead of the serial one (`pos 22` in the pipelined dump equals the serial
`pos 23`), so the commit count and the window's own advance still disagree somewhere for T > 1. The path stays
opt-in and the shipped default is `--pipeline-windows 1`.
`--pipeline-windows 2` runs, is faster (57.7 tok/s against 51.3; 68.1 against 45.9 with a forced window size) and
produces different tokens. Rounds 16-21 eliminated every switch and mode: the speculation (THETA), the prestaging,
the agreement chain, the snapshot overlap, a fixed window size, the split point (a one-layer stage 1 still differs)
and the elastic KV (off by default). What is left is the two alternating verifiers themselves.
This adds the instrument for the next step: STRATA_PIPELINE_DUMP=<file> appends a hash and the sum of the first
stage's GDN state (recurrence state plus conv history) after every window. It is debug-only - it waits for the
stage's stream - and off unless the variable is set.
It already shows one defect on its own: in the pipelined path the same position of a second request does not hold the
same state as the first request's (`pos 18 windows 2 sum 3669.18189 hash 5f718405585185d9` against
`pos 18 windows 1 sum 1952.13774 hash b97305876cf73d2c`), so a request does not start from a reset sequence state.
That is separate from the token divergence, which reproduces on the very first request of a fresh process
(`compare_tokens.py`: first difference at token 4).
The serial path needs the same dump before the two can be diffed element by element; that is the next step, and
`--pipeline-windows 1` (the shipped path) is unaffected by this commit.
Problem: `attn_batch` was 32, a value chosen for the shuffle-reduction score phase's vector-matrix kernel. The
lane-cell phase is now the prompt default, and it wants more queries per launch.
Result, same build, back to back, 3 warm repetitions each (2x Arc Pro B65, Qwen3.8-Flash-Next IQ3_S, 2,048-token
chunks):
| prompt | batch 32 | batch 128 | gain |
| ---: | ---: | ---: | ---: |
| 4,106 tokens | 1,227.9-1,235.1 tok/s | 1,240.9-1,247.6 | +0.9% |
| 16,380 tokens | 1,582.2-1,584.2 | 1,604.9-1,605.8 | +1.3% |
Batch 512 is another 0.2% on the 16K prompt but takes four times the scratch (400 MB per stage instead of 100 MB),
so 128 is the trade. The expert cache keeps its slots: 10,253 either way at 4K.
Approach: the default of the `STRATA_ATTN_BATCH` switch (added earlier this branch) moves from 32 to 128; the
variable still overrides it in 1..1024.
Validation: 64 generated tokens are identical to the serial reference (`scripts/compare_tokens.py`, first
difference: none). Not run: CUDA/HIP (untouched - the switch is the SYCL port's), a card small enough for the
100 MB of extra scratch to cost expert-cache slots.
Limits: the gain is attention-bound and small; it is the tuning of the phase that landed in the previous commit, not
a new path.
Problem: the prompt path runs `qsa_decode_attn_batch` (the tensor-core kernel is opt-in and 2.9x slower here), and
attention is 21% of a 4,106-token prompt's GPU timeline and 24% of a 16,380-token one. The port already had the
`STRATA_ATTN_LANECELL` score phase - one cell per thread, no shuffle reductions, scores bit-identical to the
shuffle reduction - but only behind an environment variable, so nobody got it.
Result, same build, same box, back to back, 3 warm repetitions each (2x Arc Pro B65, Qwen3.8-Flash-Next IQ3_S,
`--spec 4`, 2,048-token chunks, nothing else on the GPUs):
| prompt | shuffle reduction | lane-cell | gain |
| ---: | ---: | ---: | ---: |
| 4,106 tokens | 1,167.4-1,171.9 tok/s | 1,226.3-1,235.3 | +5.2% |
| 16,380 tokens | 1,456.3-1,457.4 tok/s | 1,524.3-1,526.2 | +4.7% |
| decode (256 tokens, 18-token prompt) | 56.5 / 57.0 | 56.6 / 57.0 | unchanged |
Approach: the choice follows the query count. The prompt path hands the kernel a batch of queries (`attn_batch`, 32
by default), the decode window hands it the window's few rows (at most eight per call), and the few-query case is
1.4% slower with the lane-cell phase (55.9-56.3 against 56.6-57.0 tok/s), so `n_q >= 16` takes it.
`STRATA_ATTN_LANECELL=1/0` forces either way, so the old path is one variable away.
Validation: 64 generated tokens from the gated build are identical to the same prompt's tokens from the serial
reference (`scripts/compare_tokens.py`, first difference: none); the lane-cell phase's scores are the shuffle
reduction's bits by construction (one cell per thread, same order). Not run: CUDA/HIP (untouched - the change is in
`sycl/src/kernels/cuda/qsa_decode_attn.dp.cpp`, the port's own copy), the parity bench on the gated build (it forces
its own arms), a 32K-context prompt.
Limits: measured on two Arc Pro B65 with a fully resident expert cache; the gain is attention-bound, so a prompt
short enough to be dominated by the expert GEMMs will show less.
`--batch N` (the concurrent serving slots) loads the OpenCL adapter as well, so on a two-card `xe` box
dpct::device_count() then reports five devices: the two Level-Zero GPUs, the CPU, and the OpenCL view of the same
two cards. An automatic split took every index above 0, so it tried to put a stage on the CPU and the engine died
before serving:
layer split across 5 GPUs: CUDA0, then CUDA1,2,3,4 (split auto)
layer split, CUDA2 weights: the weight arena does not fit (1466 MiB needed, 0 MiB of 160876 MiB free on that card)
The auto-filled split now only takes GPUs of the primary device's backend, so `--batch 2` with `--layer-split auto`
starts on the same 25/23 split as a plain run. An explicit `--layer-split K` / `--split-device` is unchanged.
Measured on 2x Arc Pro B65, Qwen3.8-Flash-Next IQ3_S, 64 new tokens per slot:
| slots | window | rows/window | aggregate |
| ---: | ---: | ---: | ---: |
| 1 (speculative stream) | 44.4 ms | 2.35 | 57 tok/s |
| 2 | 40.6 ms | 2.00 | 49 rows/s |
| 4 | 58.1 ms | 4.00 | 69 rows/s |
| 8 | 88.7-89.3 ms | 7.00 | 78-79 rows/s |
So a batch window costs about 18 ms + 10 ms per row, and eight slots (the most that fit the VRAM) reach ~79
tokens/s of aggregate decode - above the single stream's 57, below the 100 the two-card budget would need.
The prompt path runs the decode-shaped FP32 attention kernel `attn_batch` queries per launch (the tensor-core
kernel is opt-in and 2.9x slower here, see qsa_prompt_attn_parity). 32 was chosen for a vector-matrix kernel;
on 2x Arc Pro B65 more queries per launch give it more to overlap.
Measured on Qwen3.8-Flash-Next IQ3_S, 16,380-token prompt, second request in the process (STRATA_PREFILL_TIMING):
| attn_batch | qsa attn | prefill |
| ---: | ---: | ---: |
| 32 (default) | 2,572 ms (23.0%) | 1,294 tok/s |
| 64 | 2,492 ms | 1,309 |
| 128 | 2,448 ms | 1,316 |
| 256 | 2,422 ms | 1,315 |
| 512 | 2,409 ms (22.0%) | 1,319 |
So the switch is worth ~2% and the default stays 32 (the difference is inside the run-to-run spread on one pair of
runs); STRATA_ATTN_BATCH=512 is the measured best for a long prompt. The kernel itself is the problem: it is
1-4% of the card's fp16 peak at long context, and the XMX port is correct (2e-6 against FP64 in
qsa_prompt_attn_parity) but 3x slower, using 122,880 of 131,072 bytes of local memory for its 64-cell layout
(mode 32 halves that to 86,784 and is slower still).
The bench's queue is created with an exception handler, which swallows a failing kernel: the timing then reads
"0.3 us per group of 10 experts" because nothing ran, and the bit-for-bit comparison passes trivially (both
buffers stay at the 0xff fill). Every wait now goes through wait_and_throw() and prints the what() of a failure.
Verified on the Arc box: a run with real geometry prints dn=10 and non-zero outputs, so the kernel does run; the
per-mode timing of a single small group is still under one expert's memory traffic and needs the mode/phase
semantics checked before its GB/s numbers are quoted.
sycl/src/kernels/native_expert_bench.cpp was in the tree without a target, so it never compiled - and it still
carried raw CUDA calls (cudaMalloc, cudaMemcpy, cudaMemsetAsync, cudaFree, cudaEventRecord/ElapsedTime). It is now
a target under STRATA_NATIVE_EXPERTS and uses USM allocations, queue copies and a host clock on its own in-order
queue, so the grouped native expert modes can be timed and compared bit for bit on real GGUF rows at the decode
window's shape (G experts x T tokens).
Run: native_expert_bench <shard1.gguf> <layer> <G> <T> <ref mode> <mode> [iters]
Measured on the Arc box (layer 0 of the IQ3_S pack = iq3_xxs gate/up, iq4_nl down): every mode is bitwise equal to
the reference, G=480 x T=6 reports gu 7.7 us and down 5.3 us for the whole group - the absolute numbers still need
the mode's own semantics checked (a single group's 0.3 us is below one expert's memory traffic, so that path is
returning early).
STRATA_VERIFY_PROFILE collects GPU timestamps per layer step into device memory, but the readback ran on the
device's default queue: neither ordered after the window's work on cs_ nor necessarily this stage's device. That
is why the stage table was only ever printed in the eager mode, which stamps from the host after draining the
device - so its numbers include launch overhead and gaps and cannot attribute kernel time.
* collect_profile() reads the stamps on cs_ (under OnDevice(device_)) and waits there.
* The window prints the table from the GPU's own clock every N windows:
STRATA_VERIFY_PROFILE=1 STRATA_VERIFY_PROFILE_EVERY=32, without the eager drain.
Measured on 2x Arc Pro B65, Qwen3.8-Flash-Next IQ3_S, decode window T=4, per stage (24 layers) and per window:
total 20-24 ms; of that the routed experts ("VRAM hits") 5.0-6.6 ms in the GDN layers and 1.3-3.0 ms in the QSA
layers, the dense q8/qkv gemv 2.0, hc-read1+router 1.9, shared+quant 0.96, out-proj 1.1, z 0.97. The eager table
used to overstate hc-read0 and the experts by 2-4x because it timed the host's waits.
The two-card window overlap never ran on the Intel engine: the first poll of
every pipelined window returned "verify: error".
* Verifier::done() lost CUDA's cudaErrorNotReady branch. The port's
sycl_event_query() returns 1 for "still running"; the migrated code treated
every non-zero value as an error. A pending window is not an error now:
flush the queue every 2 ms and time out after 20 s, as the CUDA original does.
* A drafter completion event was used as a barrier on the other card's queue.
SYCL rejects a cross-context dependency ("Cannot submit to a queue with a
dependency from a graph that is associated with a different context"). On two
cards the host waits for the event; the same card keeps the barrier.
* The first card's second verifier (ver_b) took whatever device was current at
init; it is created under OnDevice(0) now.
Diagnostics: queue_wait_diag() names the queue, its device and the SYCL message
instead of "error"; an unqueryable completion event says so; and with
STRATA_VERIFY_DEBUG=1 every verifier prints its device and streams.
Verified on 2x Arc Pro B65, Qwen3.8-Flash-Next IQ3_S:
* --pipeline-windows 1: tokens identical to the serial path (64 tokens on a
short prompt, 32 after a 4K prompt, two runs each); a short prompt read drops
from 206 ms to 118 ms (-43%).
* --pipeline-windows 2: runs now (it crashed before), ~8-20% faster decode, but
it does NOT produce the serial path's tokens yet - the decode overlap and its
GDN rollback still need work on this backend. Not a default.
Service took the ids of "<|im_end|>" and "<|endoftext|>" as its stop ids whatever the tokenizer made of them. DeepSeek-V4.1's tokenizer
has neither as one token: it cuts them into '<', '|', 'im', '_end', '|', '>' and 'end', 'of', 'text'. A server run without the family
named (a hand-written config, a gateway) therefore ended the reply in front of the first "im", "|", ">", "end", "of" or "text" it
wrote, and the real end token (id 1) did not end it. Nerazim is N, er, az, im: the reply stopped at 'im' and the client got "Neraz".
The engine and its --eos 1 were fine; nothing was cut by the stop-string hold-back (StopMatcher gives its held tail back at the end).
end_token_ids() keeps a marker only where it is a single token, and falls back to the tokenizer's own end token (special_ids "eos").
Qwen's tokenizers are unchanged. The family guess from the previous commit still names the right parser and DSML for DeepSeek.
serve/test_stop_holdback.py: the end ids of both tokenizers; StopMatcher over every split of the text and one character at a time
(Qwen and DeepSeek stop lists, words ending like the start of a stop string); the Detokenizer plus StopMatcher byte by byte; and the
OpenAI and /v1/messages routes, streamed and whole, with and without the family named: no stop string (whole reply), a stop string that
never comes (the held tail is given back), a stop string (cut exactly there).
The tool-call forms the model drifts into (<parameter=NAME> as the opener, JSON inside <tool_call>, a bare <function=NAME>, a second call in one
wrapper) are read as the call when the tool is declared; anything else stays text. On the 1,462-turn drift corpus 98.3% of turns read as intended
with it and 92.9% without; the specimens of the tool corpus read the same with it on or off.
- Service.tool_call_recovery starts True; tool_call_recovery_setting(cfg) is the config rule (absent = on, a non-bool still stops the start).
- Tests: a fresh Service reads the drifted call without being told; the config rule; the old on/off switch test is unchanged.
- docs/DETAILS.md: the config row says on since 0.2.8.
0.2.7 turned STRATA_ROUTE_PRIOR=1.5 on by itself on a discrete NVIDIA card of 13 GiB or less. Reports and our own 0.2.7 agent check (4.5 of 10
tasks with it on, 5.6 with it off) tie it to repetition loops and weaker results in long agent sessions, so the owner took the default back
(2026-10-29). With nothing set the engine now routes exactly as 0.2.6 and as STRATA_ROUTE_PRIOR=0 did; STRATA_DEFAULTS=026 is unchanged.
- sm120_series.hpp: route_prior_default() answers 0 on every card; route_prior_resolve() is the pure rule the engine reads (a set value wins on every
card); kRoutePriorSmallCard stays the 1.5 opt-in value.
- verify.cpp: g_route_prior() uses it; the "on by default on this card" start line is gone.
- Expert settings / Terminal: "Faster answers on 12 GB NVIDIA cards" is now an Off-by-default row that writes STRATA_ROUTE_PRIOR=1.5; the words
say it is faster on 12 GB cards but can repeat itself in long agent sessions (nine languages). A 0.2.7 file with STRATA_ROUTE_PRIOR=0 reads as Off.
- Tests: arch_defaults_test (default table over 12 cards x hip x integrated x 026, explicit opt-in values), speed-switches.test.mjs and
tests/terminal/test_speed_switches.py.
- docs/DETAILS.md: both places that described the 0.2.7 default.
serve/server.py run on its own with a hand-written config (no "arch": "deepseek41") gave DeepSeek-V4.1 the Qwen defaults: the stop ids
of <|im_end|> and <|endoftext|>. The DeepSeek tokenizer splits those strings into pieces, so a "|" or ">" in a reply ended it
(finish_reason stop at token 19 in front of a Markdown table, or right after "## 3. Recom"), and the real end token (id 1) was not a
stop id, so it came out as the text <｜end▁of▁sentence｜>. Reproduced on 2 x RTX A6000 with the reporter's request: the config with the
arch answers 730-1161 tokens six times out of six; without it, stops at 192-436 tokens, one of them right after "## 3. Recom".
The tokenizer's own pre-tokenizer name ("deepseek3") now picks the family when the config names none; a named arch is never changed.
(cherry picked from commit b18292d702)
--expert-cache auto sized the cache from the VRAM free after the weights, less --vram-reserve-mib, the draft head and the other priced
buffers. What the load allocates after the cache and no setting prices - the CUDA libraries the prompt path and the verifier warm up
(cuBLAS handles and workspaces, lazily loaded modules) and the verifier's small buffers - came out of the reserve, so the card ended
about 240 MiB under it: 459 MiB free at --vram-reserve-mib 700 on an RTX A4000 16 GB, 457 MiB on an RTX A6000 48 GB, under the 512 MiB line
the serve check warns about at every request (#2065). Measured with STRATA_TRACE=1: -164 MiB at the prompt path and -80 MiB at the verifier
beyond the draft head, the same with --prefill 512, 2048 and auto, on both cards, and the same on v0.2.6 (so not new in 0.2.7).
auto_cache_slots() (include/strata/core/auto_cache.hpp, tested in src/core/auto_cache_test.cpp) now leaves kLateAllocMib = 256 for it,
unless the card could not keep min_slots (#496) after it, and not under STRATA_DEFAULTS=026. An explicit --expert-cache N and
--vram-reserve-mib N are untouched; the cache is ~100 slots (about 1.8% on a 16 GB card) smaller.
On a two-GPU layer split, every prompt on the chunked path stalled: a device's wait kernel spun for a
hand-off while that device's thread blocked in its first cuBLAS call, and the other device's thread too.
host_wait_at_least polls the same host-mapped flag (the kernels' signed flag - seq test) before the four
cross-group waits of the chunk path, so the kernel finds the flag up. The GPU work and its order do not
change.
The server ends an engine an earlier run lost (strata_control/procs.py) before it
starts the next one, and waited 10 s after the kill. An engine with tens of GB
of pinned memory takes longer to give it back, so the new engine's memory guard
saw 50 GiB still in use and loaded a 40 GiB arena into 10 GiB. The wait is now
up to 120 s, with a progress line every 10 s; one still there after that is said
in plain words (what it is, why starting now is risky, what to do) and the start
goes on: nothing is refused, the engine waits for RAM itself and warns.
Also documents what the parser repairs (#4078#4427#4077) and the arena RAM wait
in docs/DETAILS.md. Tests: serve/test_leftover_engines.py and a fake-clock case
in tests/terminal/test_procs.py.
An arena load with 10 GiB available for a 40 GiB arena sent a 60 GiB host into
memory reclaim until it was reset. The check that is there warns only (a
cgroup limit below the arena still stops the start), and its own example is an
engine that is still exiting. Before the warning, the load now waits for the
RAM to come back, re-reading it every 0.25 s for up to 20 s, and says so in the
log every 5 s (STRATA_ARENA_RAM_WAIT_S changes the seconds, 0 loads at once).
If it is still short the load goes on as before, with the numbers; below half
of the arena the warning says the whole PC can stall in reclaim and names a
systemd MemoryHigh= as the way to keep one program from taking all the RAM.
Nothing is refused or capped.
detail::wait_for_ram is the loop with its sampler and sleeper passed in, tested
in file_expert_source_test (test_wait_for_ram). expert_source.cpp passes
g++ -std=c++20 -fsyntax-only against the CUDA 13.1 headers; the engine itself
was not built or run for this change.
serve/server.py run on its own with a hand-written config (no "arch": "deepseek41") gave DeepSeek-V4.1 the Qwen defaults: the stop ids
of <|im_end|> and <|endoftext|>. The DeepSeek tokenizer splits those strings into pieces, so a "|" or ">" in a reply ended it
(finish_reason stop at token 19 in front of a Markdown table, or right after "## 3. Recom"), and the real end token (id 1) was not a
stop id, so it came out as the text <｜end▁of▁sentence｜>. Reproduced on 2 x RTX A6000 with the reporter's request: the config with the
arch answers 730-1161 tokens six times out of six; without it, stops at 192-436 tokens, one of them right after "## 3. Recom".
The tokenizer's own pre-tokenizer name ("deepseek3") now picks the family when the config names none; a named arch is never changed.
#4427: tool calls came back as text or with a wrong argument.
- A JSON array or object argument with a slip in it (a trailing comma, single
quotes, a bare string between two members, brackets left open) is read as
written instead of reaching the client as a string that fails its schema; the
reporter's todo_write array with "pending", "in_progress" in one object is the
test. Valid JSON is untouched; text that is no array or object stays text.
The repair is counted (/metrics strata:tool_args_repaired_total) and logged.
- A parameter's </parameter> followed by </tool_call> ends the value: the call
whose </function> was left out is a call.
- The same holds for a call written inside the thinking, and one whose output
ends on </function> inside the thinking; a following opener confirms an
omitted </tool_call> there too, as in the answer.
- The call announced while streaming and finished by the end of the output
carried empty arguments in its final event (read by MCP and the Responses
assembler): it now carries the parsed ones.
#4077: after the reasoning ended at a closing think tag, one more at the start
of a line in the answer is the stray: the model quoted the tag while thinking.
It is dropped and the answer's code tracking starts over, so a call that follows
is not taken for a quote by a fence or backtick the reasoning left open. The
reasoning after the stray is still sent as the answer (a stream cannot take it
back); the call arrives. Logged and counted.
Adds serve/test_tool_corpus.py (49 specimens: the reporters' examples, the
Qwen3.8-Flash-Next form, the drift shapes, DeepSeek DSML; each whole, in pieces
of 1, 3 and 7 bytes and cut in two at every byte, streaming on and off) and
serve/test_tool_routes.py (Chat, Messages and Responses, whole and streamed).
setup.py said the CUDA 12 engine 'works with driver 528 or newer' (525 on Linux), NVIDIA's minor-version line, while the signed feed's cuda12
packages (built with CUDA 12.9) carry min_driver 575.51 / 576.02 and strata update refuses them below that: a V100 PC on driver 556-560 was told
the engine would work and then that the feed had no engine for it.
strata_control/driver_floor.py reads the floor from the installed cuda12 package, else the signed feed (once, 10 s at most, only in a setup run
that needs the CUDA 12 engine), else the figure the published packages carry. Setup warns with that number (recommend, never force: a locally
built engine may still run); the only stop left is NVIDIA's own line, under which no CUDA 12 program starts. docs/OLDER_GPUS.md says the same.
source_version() read CMakeLists.txt, which the packaged Terminal does not carry: prebuilt_bases() raised FileNotFoundError and ended the setup.
It answers '0' now when there is no source tree, prebuilt_bases() then looks at the latest release only, and an installed engine is not called
stale for want of a version to compare with.
strata install passed its tail to setup.py through argparse.REMAINDER, which does not take a tail that starts with an option the parser does
not know (--source, --family): 'unrecognized arguments', or the value split from its option. The options of setup.py now go behind a --
before the parse; --update and --help stay strata's own.
When the runtime refused the start because another host still owned the profile ('already has a live managed runtime'), no engine existed
and no log was written, but the error ended with 'The engine's log has the reason: <path>' and the screen showed an empty folder. A busy
profile now gets its own sentence (the other server, and strata launch --stop), and a log that is not on disk is no longer named, here or
in the sentence after an engine that stopped while loading.
The 0.2.7 policy locks the 26.8 GiB PLE table in RAM at start when table + expert arena + 8 GiB fit. What the rest of the load takes
afterwards (KV pools, pinned staging, host buffers of two GPUs) is not known then: on a 128 GB two-GPU PC it left under 5 GiB and the
automatic conversation cache turned itself off, a regression against 0.2.6, which kept 8192 MiB.
When the cache is automatic and the lock came from the policy (not --ple-io ram), the cache is sized first: if unlocking the table
would give the cache more room, PleTable::unlock() hands the lock back (the pages stay cached, the OS may reclaim them), the log says
so, and the cache is sized from the RAM then free. --conversation-cache-mib N and an explicit --ple-io ram are untouched.
ple_lock_blocks_cache() is the decision, unit-tested in src/core/ple_residency_test.cpp.
slot_load() selected the lead layer's device once and then copied into every member of a tensor-parallel group; the members' buffers
sit on their own GPUs, so on two identical cards (tp 2) the second member's copy ran with the wrong device current. A failed copy
makes slot_load() fail, and slot_take() then deletes the saved conversation. Every other per-member loop (restore, reset) already
selects the member's device; slot_load() now does too.
A batch slot's chain holds one checkpoint (the one its admission copied in). park_slot() went through park_source(), whose
drop_superseded() (#342) removes a parked entry when the outgoing chain holds that entry's deepest checkpoint; the main session's
copy of the same conversation qualifies, and the root checkpoint at the end of an agent's system prompt went with it. The next
overlapping request then read the whole 5 K-token prefix again (323 ms in 0.2.6, 4130 ms in 0.2.7).
ConversationCache::supersedes_richer() says when the park would drop an entry that has a checkpoint the slot's chain lacks; park_slot()
then leaves the slot unparked (it is overwritten, as in 0.2.6) and says so. A slot with no such parked copy is parked as before.
The predicate and its test come from the report (shenmintao, #4420).
store_gb_now() took any drop of 4 GB or more in the free RAM between the first estimate and the end of the weight upload for a
GPU that shares the PC's RAM, and subtracted a context-sized reserve (27 GB at 256K). On a discrete pair (2 x RTX 3090) the
pinned staging buffers and the page cache cause the same drop, so a 256K context shrank the store from 48 GB to 15-18 GB and
decode fell from 9-12 to 3-4 tokens a second (a regression since 0.2.4.2).
The decision is now store_gpu_shares_ram() (ds4_experts_rule.hpp): the RAM must have dropped AND a device must report
cudaDevAttrIntegrated. Strix Halo, the DGX Spark's GB10 and the Apple path keep the old rule; a discrete card never counts.
Unit test in tests/dsv41/experts_rule_test.cpp.
The Qwen engine's --vision stands alone, the DeepSeek engine's is a pair (--vision on|off). Taking out only the name left the value behind
(--eos 1 off --dspark auto: unknown option off, exit 2). A DeepSeek model keeps its pair whole now.
The Settings field 'GPU memory for other apps' wrote --vram-reserve-mib and the resource preference wrote --pool-workers into a DeepSeek
model's arguments; strata-dsv41 takes --reserve-mib and has no --pool-workers. Before a load, the arguments and the person's saved options
of a DeepSeek model are built from the engine's option list (src/dsv41/ds4_main.cpp, checked by a test): the reserve is renamed, anything
else is left out with its value and said in the load's log. The tune operation no longer writes the resource preference into a DeepSeek file.
The app's ceiling probe read one file. The second symptom of this branch - `direct runtime intent has an invalid timeout`,
on a build whose CLI validator was already fixed - is the reason it now reads the operation record too: a start wait is
bounded by management, by the record written before the start runs, and by the runtime, and a Control can raise any of
the three on its own. The app asks for more than 600 s only when a Control's own sources hand both of the first two to
the runtime's ceiling, so a half-raised Control keeps the wait it will accept instead of failing the load. Anything the
app cannot read, or a line that still spells out 600, keeps today's 600 s.
Checked against the Control folders on this PC and against the released sources in git: v0.2.5 and v0.2.7 both carry
`MAX_START_WAIT = 1800` with both gates at 600, so the app asks them for 600 s - which is what they take; the patched
tree and the staged 0.1.52 component ask 1720 s for a 180 GB model; the installed 0.1.49.2 component asks 600 s. The
marker for the record is `"wait_s": (` rather than `wait_s`, because the record names that key in its field lists long
before its range table, and the first match there would have pinned the app to 600 s forever.
The packaged test drives the four Control shapes (nothing, runtime only, CLI only, all three) so each half-raised
combination is a case, not a guess.
A local build of the previous commit got further and said `direct runtime intent has an invalid timeout`. A start wait is
checked three times before the engine is reached, and 27cbbfe6 raised only the runtime's: management.runtime_action (fixed in
the previous commit), journal._validate_direct_intent, and runtime.start itself. The record is written before the start runs,
so its own (0, 600) for runtime.start and (1, 600) for startup_wait_s stopped a long wait that the layer above had just
accepted.
The record now reads both ceilings from the runtime that runs the wait, the way the CLI validator does. A wait above them is
still refused, and stop and verify keep their own 600 s, which is what runtime.stop and runtime.verify take.
The tests are the reason this was found twice by hand instead of once here: the CLI test from the previous commit faked
runtime.start, so it stopped at the first gate and passed while the record still refused the wait. One test now walks
management -> journal -> runtime.start with only the engine itself faked, at 300, 600, 601 and MAX_START_WAIT, and says which
wait failed to arrive; another pins each kind's ceiling on the record directly. Reverting the journal change fails three of
them with the quoted message; with it, all four pass. StartWaitLimit keeps its records in a scratch folder.
The version gate of the first commit was wrong, and a local build of it said so on the first load of a big model:
`Start wait must be finite seconds in 0..600`. Control validates a start wait twice and the two ceilings disagree.
`runtime.start` takes `MAX_START_WAIT` (1800, 27cbbfe6), but `strata runtime start --wait` - the command the app uses -
checks the wait in `management.runtime_action`, where 27cbbfe6 left it at 600. It refuses the start before the engine
is reached, so no released Control (v0.2.4 through v0.2.7, Control 0.1.45 to 0.1.52) takes a wait above 600 s however
it is numbered: a version gate could not tell which Control would take it.
`management.runtime_action` now takes the ceiling from the runtime it hands the wait to, for both start and restart,
which is what 27cbbfe6 had already done for the direct `runtime.restart` path. The default stays 300 s and every wait
at or below 600 s is unchanged. The app reads the ceiling out of the Control it found - the same source-read it does
for the settings fields and /v1/responses - rather than guessing from a version, so it keeps 600 s on the Control
people have today and takes the longer one as soon as that Control is updated. A Control it cannot read falls back to
600 s, which is what the app asks for now.
Tests: the CLI path gets the ceiling test that only existed for `runtime.start` before, and the packaged load test
drives the three Control shapes (no ceiling, runtime 1800 with a CLI stuck at 600, CLI following the runtime) instead
of a version string.
The app waited a flat 300 s for any model, so a big model on a slow disk ended with `startup_wait_expired` while the engine
went on reading: #4191 reports it at about 180 GB and says editing 300 to 600 in this file fixed it on 0.2.1 but needs
several files now, and #3459 is the same stop on Windows after the tuning stage. Main solved the Terminal's side for
itself (27cbbfe6, #3949) and left this file with 300 s, or the config's engine_ready_s clamped to 600 s.
That clamp was written the same day as `MAX_START_WAIT = 1800`, on a branch whose Control still refused more than 600 s —
its message says "at most the 600 s Control allows". Control now takes 1800 s, so the clamp's reason has gone: it is the
only reason the app cannot use the size figure the Terminal already uses.
- startWaitFor(config,entry,model) takes the model's own size where the app has it (the installed model's bytes on disk,
else the catalog's download figure) and applies main's shape: ten minutes, about eight more seconds per GB above 40 GB.
A model nobody can size keeps the 300 s it always had, and a config's engine_ready_s is a floor, never shortened.
- The ceiling comes from the Control this app talks to, not from a guess: 600 s before Control 0.1.50, 1800 s from there.
Control rejects a wait above its own ceiling before it starts anything, so an older Control keeps today's behaviour.
- Where the start now waits longer than the server's own engine wait (900 s), the run config gets `engine_ready_s` above
it, as `strata launch` does: the server ends the engine when its wait runs out, so a start that waits longer than the
server can never win. It is only ever raised, and only in the launch's own copy of the config.
No new variable, no change to Control or the Terminal, and nothing to set: the wait follows the model, as main chose for
the Terminal. The app's outer five-minute health loop is not the load wait in packaged mode (Control has already reported
ready), so it is left alone.
Validation: Linux, Node v22.23.1, against upstream/main 5e8a4cca on the same machine. `npm test` in apps/host: 1227 tests,
1120 pass, the same 2 failures as on upstream/main (component-updates "an older host left running" and timeskip-real
"Stop ends the job"). The two new tests cover 180 GB at 600 s on Control 0.1.45, 1720 s on 0.1.50 with the server readying
after it, a small model keeping 300 s, the 1800 s ceiling, and setup's 3600 s for DeepSeek being honoured and never
shortened. The 11 version and size cases were checked against the formula directly, including an unknown Control version,
which falls back to the conservative 600 s. Not run: a real load on a real PC, Windows, macOS.
A bare `<function=` in prose or code (no name, or a name not followed by a
parameter or close) counted as a quoted call, so every later </parameter>
was skipped: the call came out as content, or never finished when streaming.
Adds the review's two values to test_quoted_call_values.
A parameter close inside a call that the value itself opened (more
<function= than </function> before it in the value), or one followed by the
same parameter name again, is part of the value. Prose between two calls
still ends the first call. Test: serve/test_quoted_call_values.py.
discover_run_config picked the first installed model by id. With a DeepSeek and a Qwen model installed, "deepseek-..." sorted first, and the
test start ran the engine under test (strata.exe) with DeepSeek's options: "unknown argument: --eos", server_exited, engine 0.1.50 refused
("did not start correctly on this computer"). A DeepSeek model runs on strata-dsv41, which has its own package, so it is left out; with only a
DeepSeek model installed the update has no model to test with, as on a first run.
strata_cli/launch/settings.py saves the model into a preset/choices file lower-case ("--model iq3_xxs"),
and LaunchRunner.setup_args() forwarded it unchanged. setup.py's --model choices are the MODELS table's own
keys ("IQ3_XXS", "Gyro-S"), so argparse refused the argument: starting a saved preset with any choice that
differs from the installed config (--vision, --context, --kv) reran setup and stopped there. Typing the same
name on the command line worked, because cmd.py normalises it through Catalog.resolve().
setup_args() now resolves the pair through the same Catalog.resolve() and keeps the old value when it
cannot. The except is deliberately broad: the resolver reads the model table, and a table that cannot be
read here must not turn a launch that worked into a traceback - the values then go on to setup unchanged,
exactly as before. The new test drives that case with a resolver that raises.
Only the engine / Control number: the app, the Terminal and the Metal engine keep their versions here; the release bump of the other files
is the release's own commit.
nvidia-smi gives a GB10's memory.total as [N/A]. tools/strata_mcp.py's own NVIDIA probe (the fallback of the launch hardware page)
dropped the card; it now counts it as setup.py's gpus() does (arm64 only). strata doctor --live accepted no card with that figure; on arm64
it lists the card with no memory figure (x86 still refuses it). The DeepSeek fit plan for a unified-memory GPU with no BIOS carve-out
(GB10) no longer speaks of Windows, a BIOS and a Radeon, and quotes the numbers measured on the Spark. The Windows Strix Halo plan text
is unchanged (tests).
tools/dgx_bench.py, dgx_warm.py, dgx_compare.py, dgx_quality.py, dgx_quality24.py, dgx_needle.py, dgx_turns.py, dgx_longctx.py, dgx_128k.py,
dgx_kvresident*.py, dgx_concurrent.py and dgx_smoke.py (the acceptance gate in one command; its data folder is now an option and its scratch
files live in a folder of their own) and docs/DGX_SPARK_LINUX.md.
- strata_control/targets.py: host_platform() names Linux ARM64 'linux-aarch64' (as uname and Tauri do; a Mac stays macos-arm64) and the
cuda backend is committed to there (nothing else is: no HIP, SYCL or CPU engine for arm64 Linux)
- strata_control/feed.py: linux-aarch64 is a feed platform for engine, dsv41 and control packages (and a Tauri target name, no app is built)
- packaging/package_terminal.py: the Terminal for linux-aarch64 with python-build-standalone 3.12.15 aarch64-unknown-linux-gnu (SHA-256 from
the release's SHA256SUMS); packaging/lock_wheels.py + locks/runtime-linux-aarch64-cp312.txt: its hash-pinned wheels (the same 23 packages
as the x86_64 lock; platformdirs resolved to 4.13.6 against 4.13.0 there)
- packaging/make_local_feed.py: --dsv41 for the DeepSeek packages, linux-aarch64 terminals; packaging/install.sh installs on Linux ARM64
and says the engine is for the DGX Spark (GB10) only, instead of refusing
- packaging/build_cuda_linux_aarch64.sh: the release recipe (strata + strata-dsv41 + the reader, the engine package and the separate
DeepSeek package, sm_121 + compute_121, driver 580.65, build id naming the commit)
- tests: test_linux_aarch64.py, the Spark rows in test_release_matrix.py
Nothing changes for another platform: every edit is a new table row, a new branch for aarch64 Linux or a new file.
- gpus(): a GPU whose memory nvidia-smi does not report (GB10) on arm64 is counted as Strix Halo's is: the RAM less the OS share
(UMA_OS_LEFT_GB), uma True, dedicated_gb 0, so the low-RAM and fit rules do not count the same memory twice. #3622 listed the whole RAM
as VRAM beside the RAM.
- cpu_info(): an arm64 CPU is named from lscpu (Cortex-X925 + Cortex-A725), not 'unknown CPU'. The PCIe probe is skipped for a
unified-memory GPU (nvidia-smi's 1.0 x1 there means nothing).
- get_prebuilt(): on arm64 Linux the ready-made engine is the linux-aarch64 package of the signed feed (engine_from_feed); without one,
setup compiles as #3622 had it do always. setup no longer forces --build on arm64.
- Step 7 on a GB10 (is_gb10: uma, sm_121, arm64 Linux, nothing else): --mmap-experts (one copy of the experts, all 24,576 cache
slots), --prefill auto:32768, --mtp-q4 all, --mtp-hnorm stream, each measured on one GB10 with the answers unchanged; --draft-vocab
en is printed as a tip, not set.
- The CUDA repository key URL is sbsa on arm64; the DeepSeek-engine-missing message names the Spark build script on arm64.
Tests: test_setup_arm64.py (unified-memory shape, is_gb10, lscpu name); test_setup_pins.py skips its x86-zip Engine tests on arm64 Linux.
nv_defaults.cpp gets an sm_121 table beside sm_86 and sm_60: STRATA_PF_FUSED=1 (fused IQ prompt kernels), STRATA_PREFILL_NVFP4=w4a4
(NVFP4 prompt MMQ with FP4 activations), STRATA_HC_UPMIX=1 (fused hyper-connection read), STRATA_SPEC_COST=1 (cost-aware MTP window).
Measured on one GB10 with paired A/Bs (docs/DGX_SPARK_LINUX.md); a switch the user set is kept. STRATA_SM121_PERF=0 turns the table
off, and so does the 0.2.7 master switch STRATA_DEFAULTS=026. Unlike the sm_86 / sm_60 tables these are not bit-exact (the fused prompt
path and the FP4 activations round differently; 24 of 24 greedy answers were identical), so the start line does not call them exact.
moe_fused_iq.cu: on a GB10 the work-item rule never reached its 128-row tile (the launch average is above 112 rows an expert at every
prompt size), so pick_ww takes the 128-row tile from 65,536 routed rows a launch up (+3% prompt at 16K and 32K; 4K keeps 64).
STRATA_PF_TILE_DEBUG=1 prints the choice. Other parts take the line they took before.
- ds4_gg_cpu.cpp: amd_family() on a non-x86 CPU read no variable the rest of the function uses; it now reports no AMD family
- exl3_cpu.cpp: kGemmK moved above the AVX-512 section, as exl3_cpu_gemm_scratch() sizes by it on every CPU
- ds4_model.cu: uma_free(): on an integrated NVIDIA GPU on Linux cudaMemGetInfo counts only MemFree, while the model files' page cache
holds most of what the kernel can still give back. The placement, the CPU-lane warm-up and the vision tower size their room from
MemAvailable less DS4_UMA_HEADROOM_GIB (default 6) when that is larger. A discrete GPU, HIP and Windows are not touched.
On a GB10: 12,163-12,226 of 15,744 experts in graphics memory where cudaMemGetInfo's free placed 11,961.
#3622 made the strata target compile on aarch64. This adds what a Spark build needs beyond that:
- the trellis dispatcher (tq_cpu.cpp, its scalar rows) is part of the arm64 kernel library, as it is in the Metal build, so the
DeepSeek engine's CPU lane links; ds4_tq_cpu.cpp names the AVX units on x86-64 only
- a STRATA_PORTABLE build on arm64 sets ggml-cpu's -march to armv8.2-a+dotprod+fp16+i8mm (STRATA_ARM64_ARCH) instead of the build
host's; GGML_NATIVE OFF alone would leave plain armv8-a and no NEON dot products
- the x86 kernels' parity programs and benches are not configured on arm64; the test suite defaults off there (-DSTRATA_BUILD_TESTS=ON
still builds what builds)
- spark/build_engine.sh: the reproducible build of strata, strata-dsv41 and the picture reader into an engine folder, with the CUDA
libraries beside them (RUNPATH $ORIGIN)
Every change is behind STRATA_CPU_X86 being OFF or inside a new file; the x86 and Metal builds do not take it.
The 0.2.7 DeepSeek changes (persistent-grid IQ kernels, int8 rounding batches, hc_mix_split with a block counter, head columns,
chunk-path knobs) are kept; the Metal engine's hooks sit on top of them: every kernel launch is strata_ds4_launch(), the rounding
batch ends in the Metal multi-quantiser dispatch, the head stays at 4 columns a pass on Metal (8 moves the KL), and the macOS
parallel expert prefetch is macOS-only.
Same 15 files as the 0.2.6 bump (issue template, AGENTS.md, CMakeLists.txt 0.1.52, desktop and host package files, Cargo.toml and the strata-desktop entry in Cargo.lock, SettingsUpdates comment, qa1-storage test, metal 0.1.52-metal, build_hip_linux.sh PKGVER 0.1.52, strata_cli PRODUCT_VERSION, the two version tests), plus the Strata-0.2.7-<file> names in the md files.
The engine's own tower (319 MiB) through strata launch and setup (--vision off still turns it off); setup fetches
the pinned 972 MB mmproj; the server sends pictures to the tower. Checked on 4x R9700, 1x R9700 and a real
launch on 2x A6000 (download, pack, picture answer over HTTP); 0/10 stalls, KL unchanged. Also: a newer shard 1
from --model-path (ISTA's 2026-10-26 chat template) is accepted after its own size and SHA-256 check.
Pipelined groups beat the serial default at 2 and 4 streams on 2x A6000 (+17/+22%) and 2x A4000 (+7..12/+17..20%);
one stream equal, -4% on A4000 IQ3_XXS. STRATA_BATCH_GROUPS_NV_SERIAL=1 or STRATA_DEFAULTS=026 gives serial.
ISTA re-published (and patch_chat_template.py rewrites) the first shard of each set with a larger chat template, so a shard 1 on disk can be 11 KB
larger than the pinned revision's. setup refused it (size mismatch; the GGUF preflight does not read DeepSeek tensors, so it was not taken as a
whole own file). The pinned revision stays the download; the newer shard 1 is accepted by its own size and SHA-256 with --gguf-dir.
native_arithmetic_parity failed on an Intel i9-13900K (AVX2 + AVX-VNNI, ggml built native)
in five of its eight configurations: "gu=16 down=20 ff=96 nt=2 token=0". Not the VNNI row
path: STRATA_NO_AVXVNNI=1 changes nothing, STRATA_NO_IQ4NL=1 passes. The IQ4_NL down
reference (iq4nl_reference_rows) differs from that ggml's one-token dot on the scalar tail
block of a row with an odd number of blocks (ff=96 is three): that ggml does not contract
`sumf += d * sumi` into an FMA, the reference lets the compiler do it. On the other
machines tested the two agree, so it is a property of the ggml build, not of the CPU.
With an unfused tail in the reference all eight configurations pass on that machine.
The invariant path now checks each reference once against the linked ggml-cpu on synthetic
blocks before using it (iq4nl_reference_ok for IQ4_NL with 2, 3, 5 and 20 blocks;
iq512_reference_ok for types 18, 21 and 22 with SwiGLU, on AVX-512 CPUs). On a mismatch
it prints one line and the invariant path uses ggml's own one-token dot per token: slower
on such a build, never different. The Default path is not touched; only the opt-in
STRATA_SPEC_PROB=1 decode reaches this code.
Tests: native_arithmetic_parity, eight configurations, on the i9-13900K (4,866,048
comparisons, with the fallback line), z440 and the owner PC (see the series notes).
0.2.1 made a layer split over exactly two NVIDIA cards with --batch and no --batch-groups run one serial group (measured
then on 2x RTX A4000, IQ3_XXS: +12..16%). With the 0.2.7 code the pipelined groups are faster at 2 and 4 streams on
every box and quant measured, so the default goes back to the pipeline.
Measured (tools/serve_load.py, --batch 4, 256 tokens per request, interleaved pairs, total tok/s, pipelined / serial):
- 2x RTX A6000, IQ3_S, all experts in VRAM: 1 stream 101.7 / 101.7; 2 streams 125.7 / 107.4 (+17%); 4 streams 171.4 / 141.0
(+22%). The v0.2.6 binary on the same box: serial default 104.9 and 137.2, --batch-groups auto 124.6 and 169.6.
- 2x RTX A4000, Q2_0 pack (17037 of 24576 pairs cached, 4 clean pairs): 2 streams 80.2-82.5 / 75.0-75.5 (+7..9%, 4 of 4);
4 streams 110.8-111.5 / 94.8-95.3 (+17%, 4 of 4); 1 stream 63.6-68.7 / 68.2-70.3.
- 2x RTX A4000, IQ3_XXS (13056 of 24576 pairs cached, 3 pairs): 2 streams 73.2-73.7 / 65.8-66.5 (+12%, 3 of 3); 4 streams
100.5-100.7 / 82.9-83.6 (+20%, 3 of 3); 1 stream 58.4-60.5 / 60.7-62.6 (serial about 4% faster, 3 of 3).
The single stream runs on the solo path in both; only the A4000s show the 4%.
What changed: batch_groups.hpp has a pure rule nv_serial_wanted(env value, STRATA_DEFAULTS=026). STRATA_BATCH_GROUPS_NV_SERIAL
set to anything but 0 asks for the serial group, an explicit 0 wins over STRATA_DEFAULTS=026, and with the variable unset or
empty STRATA_DEFAULTS=026 gives v0.2.6's serial group. An explicit --batch-groups (a number or auto) still wins. Three or more
cards, AMD builds and --batch-mtp/--adapt-async handling are unchanged. The serial message names the opt-in.
Tests: batch_groups_test (the rule over the env values and 026, two cards pipeline, the existing serial cases). Real engine on
2x A6000 (IQ3_S), no --batch-groups: the engine says "2 groups of 2 slots" / "2 groups of 1 slot"; with 4 and 2 slots, greedy,
greedy+seeded sampled, slots finishing at different lengths and a cancel in the middle of a draft, every slot's tokens equal its
solo decode and the pipelined build's. A long-prompt slot differs from the pre-merge base build in the same way with the flag
and without it (it is the parent's prompt path, not this change).
Not covered: no run on AMD or on three cards; the 4% single-stream cost on the A4000s is measured on one box.
THIS COMMIT ALONE MAKES THE WALK A DEFAULT. Drop it and the layer-major prefill stays
an opt-in (STRATA_PREFILL_LAYER_MAJOR=1) with the chunk-major path as the default.
Rule (include/strata/core/sm120_series.hpp, pure, tested in src/core/arch_defaults_test.cpp):
- prefill_layer_major_default(hip, stages, v026): on for two or more stages on CUDA,
off on one GPU, in a HIP build, and under STRATA_DEFAULTS=026 (qwen_defaults_026());
- prefill_layer_major_resolve(env, ...): STRATA_PREFILL_LAYER_MAJOR=0 / 1 (any text
atoi reads) wins in both directions, over the rule and over STRATA_DEFAULTS=026.
Every decline of the earlier commits stays and runs in lm_prepare as before (HIP,
image prompts, DFlash2 taps, KVA, stop_at_chunk, a pending force_chunk, a peer GPU,
APR / NVFP4 / WMMA / fused layers, STRATA_EXPERT_NIBBLE, no VRAM, no pinned RAM):
such a prompt reads chunk-major with nothing but a log line. When the walk is on by
default (nothing set) the VRAM margin is 512 MiB instead of 96 and the super-chunk
target 32768 tokens (the size measured) instead of 8192; "a short prompt" is not logged.
STRATA_PREFILL_LM_TOKENS and STRATA_PREFILL_LM_MARGIN_MIB still win.
Measured (z440, 2x RTX A4000 16 GB, Qwen3.8-Flash-Next IQ3_XXS, experts streamed, 5
interleaved pairs, time to first token, medians): 8K 2.8x, 20K 4.8-5.7x, 32K 5.1-5.5x
faster (141 s to 26-28 s); decode 38.6 vs 38.4 tok/s; needles 12 of 12 found in both
arms. The speed-up is larger than the report's 2x because --prefill auto picks 512-token
chunks there. The first-token KL numbers are in the gate notes (3976-kl.sh).
Not covered: one GPU (stays opt-in), HIP, SYCL, Metal, three or more stages, a card
with more VRAM than the pack needs (where the experts are VRAM-resident and the
chunk-major path has little to re-read).
Problem. #3948: with STRATA_QFUSE=1 the verified text changes and MTP acceptance drops, on CUDA. The QFUSE kernels are bit-exact
(gr_parity, gdn_parity, 10,500 random windows each on sm_86 and sm_120, after 91a4b8a67), yet the engine's greedy text still differed
from QFUSE off on every prompt tried. Model run, l620 (RTX PRO 5000, sm_120), Qwen3.8-Flash-Next IQ3_S, MTP --spec 6, identity-gate
flags, 4 prompts of 4K and 12K tokens, 96 greedy tokens: QFUSE on differed from off on 4 of 4 (the off-vs-off control was identical).
Cause. A window of one token advances the GDN state and the indexer tail itself ("self-commit") so that commit() needs no commit
graph. Verifier::record_window decides this with
`T == 1 && !always_publish_ && !batch_rec_ && !g_qfuse() && one_token_self_commit()`: with STRATA_QFUSE the fused GDN output
kernel takes the q8_1 destination and the window does not self-commit. Verifier::commit() skipped the commit graph with
`last_t_ == 1 && !always_publish_ && one_token_self_commit()`, without the QFUSE term. So with QFUSE on, every accepted one-token
window left the recurrent state (and the indexer tail) one token behind. Windows of one token are the ordinary case when the drafter
offers nothing, which matches the report: fewer drafts offered, lower acceptance, different text, and an unchanged text where
no window of one token occurs. On HIP (where QFUSE is the default on gfx1151) one_token_self_commit() is off unless
STRATA_ONE_TOKEN_COMMIT=1, so the two sides agreed there.
Change. commit() takes the same condition: `... && one_token_self_commit() && !g_qfuse()`. Without STRATA_QFUSE nothing changes.
Result (same l620 run). QFUSE on with this change against QFUSE off, same build: logits md5 and the 96 greedy tokens byte-identical
on all 4 prompts (IDENTITY PASS); the build without the change differs on all 4. The default path (QFUSE unset) of the merged head
against the f-027-int2 head (2a2264c21): first-window logits and greedy tokens byte-identical on the same 4 prompts.
Not covered. Serve answers and draft counters (3948-qfuse.sh step 2, SERVE=1), sm_86, and the reporter's 40-prompt set.
The report asks for the first data mismatch between the GGUF row bytes and what the direct reader delivered, and for an early
check of the gathered values; neither exists. Its trace stops at "the embedding has 96 non-finite values" without saying whether
the reader gave a wrong row or the table holds one.
With STRATA_PLE_CHECK=1 (a tester switch, off by default; one static flag read per gather call when off) every direct-mode decode (read_row, collect,
gather_batch, also from the prefetched slots) checks the row's 160 values. For a value that is not finite or beyond 1e5 (an IQ4_NL
row cannot decode beyond 8.3e6; real PLE values are orders smaller) it re-reads that row from the file with a buffered read and prints
one line to stderr (at most 8 per table): the row, its file offset and page offset, the value, and either "the delivered bytes ARE the
file's bytes" or "THE DELIVERED BYTES DIFFER FROM THE FILE from byte N" with the first 8 bytes of each. That separates a reader
fault (the bug) from a damaged table. gates/3602-ple-direct.sh sets it in its direct arms.
Checked: ple_reader_test --selftest on z440 with the switch on prints the "ARE the file's bytes" line for the synthetic table's random
rows (which decode to huge values); with it off the output and the bytes are unchanged (ple_reader_selftest passes).
Not checked: the "DIFFER" branch has no natural trigger (that is the point); it was read, not run.
Problem. #3948: with STRATA_QFUSE=1 the verified text changes and MTP acceptance drops (RTX 3080, sm_86). The report suspected the
group counter or the nat_xq_ offsets of fused_gr.cu; the audit and the new two-group test (previous commit of this stream)
rule those out. What the new test found instead, on an RTX 5070 (sm_120, CUDA 13.0, MSVC), over 10,500 random windows of two groups: in
about 1 run of 400 one q8_1 code of the fused image differs by exactly one from native_quantize_q8_1's, in every mode (group-relative,
window-indexed, direct, graph replay). The scale/sum pair and the other codes agree.
Cause. native_mmvq.cu, which holds native_quantize_q8_1 (the quantiser every consumer of a q8_1 image was written for), is
compiled with --use_fast_math (CMakeLists.txt, the native_* list), so its `amax / 127` and `x / d` are the fast full-range division
(div.full.f32, up to 2 ulp off). fused_gr.cu (the hyper-connection read's tail, gr_up_q32_kernel) and verify_kernels.cu (the GDN
output norm, gdn_q8_1_store) are not in that list and divide with IEEE rounding. roundf(x / d) then lands on the other side of a .5
at the rare element whose quotient is next to a boundary. One changed activation code per ~10^6 values is about one per few
hundred tokens at 80 hyper-connection reads of 2560 values per token: enough to move a near-tie in the verify, and the
draft (which quantises with the separate kernel) disagrees with the verifier a little more often, which lowers acceptance. That
matches the report (acceptance 0.755 against 0.79, different text, and an unchanged text where no draft is made).
Change. include/strata/kernels/q8_1_finite.hpp: q81_div(a, b) is `a / b`, except in a translation unit that defines
STRATA_Q81_FASTMATH_DIV (CUDA device code only, never HIP), where it is the same PTX division the fast-math file gets (div.full.f32).
q8_1_quant and the scale use it. fused_gr.cu and verify_kernels.cu define the macro; nothing else does, so every other kernel,
the separate quantiser and every HIP build compile to what they did. STRATA_QFUSE is off by default on CUDA (it is on by default on
gfx1151, which is HIP and unchanged), so the default path is bit-identical.
Result (RTX 5070, sm_120): gr_parity's two-group window test: 10,500 runs with 1,500 random inputs for each T, all modes, zero
differences (before: dozens); the whole of gr_parity, gdn_parity and gdn_parity with STRATA_GDN_SPLIT=1 pass.
Tests. gr_parity: the two-group test prints the first differing bytes (block, field), and STRATA_GRP_REPS=N / STRATA_GRP_KEEP_GOING=1
repeat it longer and count every failure. gdn_parity: the q8_1 destination check uses its own stream (native_quantize_q8_1 refuses the
null stream, which crashed the test on Windows), unbuffered stdout and STRATA_GDN_PARITY_TRACE=1.
Not covered. sm_86 (the report's card) has not run the new test yet (z440 was reserved for other jobs); the cause is not
architecture specific. The model run (gates/3948-qfuse.sh) is still owed: QFUSE on against off must now be byte-identical, and the
draft counters equal.
Problem. #3844: --spec 5 measured +6.1% decode on an RTX 5060 Ti, but a saved preset that carries it was left out entirely
("outside the safe values [3, 4]"), and an imported one was refused. The tuner's BOUNDS are the values it tries; they are
not a finding that a value is harmful. The never list (--spec 6 and wider: 4% to 16% slower, --expert-cache 0, ...) is.
Change (Terminal / Control; touches strata_control/tune_knobs.py, strata_control/presets.py, strata_cli/tune_commands.py and
serve/server.py).
- tune_knobs.check_setting_parts(args, env, ctx) returns (never, bounds). check_setting is the two lists together, in the
order it had, so the tuner still skips --spec 5 and its tests are unchanged.
- presets.overlay applies a preset whose only problems are outside BOUNDS and returns info["warnings"], one short line each
("--spec 5: not tried by the tuner, kept because it is your setting"). A never-list setting still means "not applied".
presets.create and import_preset refuse only never-list settings.
- serve/server.py prints each warning as "[strata] warning: ..." when the preset is applied; strata_cli tune_commands
_effective adds "warnings" to its payload.
Tests. tests/tuning/test_presets.py: a preset with --spec 5 is saved, applied (the flag becomes 5, the input config is not
modified), warns once, and an in-range preset has no warnings; the never list still refuses --spec 6; import of --spec 5 works.
tests/tuning/test_autotune.py: --spec 5 stays in the tuner's refused list. python -m pytest tests/tuning: 274 passed;
tests/terminal -k "tun or preset or overlay": passed.
Not covered. The app's Details text does not show the warning yet (it reads the payload key above).
Problem. #2408 / PR #2409: restoring a large parked conversation on a layer split ended the server with a CUDA out of memory
on a later card. The PR checks the free VRAM of every card against --vram-reserve-mib before a restore. Its own review
points out that a loaded card is always below that reserve, so the check would skip nearly every restore on a split (and
on one GPU). It is not ported.
What a restore allocates. conversation_snapshot_restore writes into the session's existing buffers (pools(st), the GDN
state, the indexer tails); refresh_residency only launches kv_stream_reset or copies into the existing ring. The elastic K/V
growth (kvg_ensure) allocates, but only on one GPU. The one thing a restore can create on a card is the calling thread's
private copy stream for it (conversation_copy.hpp private_stream), which it takes from the pool of finished threads when
there is one (#3350). A failed restore is fatal by contract (the session may be half written). So the need is one stream
and one event, and only for a card whose stream does not exist yet (a snapshot loaded from disk, or a restore on a thread
that never copied on that card).
Change. conv_cache.hpp: kRestoreFirstUseMib (64) and restore_fits(free, stream_ready, n, need): a restore is refused only
for a card that has no copy stream yet and less than 64 MiB free. conversation_copy_stream_ready(device) says whether a
card has one. generate.cpp: before the restore, each card of the run is checked; a snapshot that would not fit goes back to
the cache and the prompt is read again, with a log line. A card that has its stream is never refused, however full.
No sample (a failed cudaMemGetInfo) means the restore goes ahead as it did.
Tests. conv_cache_test: the predicate (cards with a stream are admitted at any free size; the edge at the need; no
sample / no cards / no need). conversation_copy_streams_test: copy_stream_ready before and after the first copy, in the pool
after a thread ends, and the measured cost of a card's first copy stream and event (printed; the test requires under
16 MiB, so 64 MiB leaves a wide margin). conversation_snapshot_test: after the first restore, six more restores of two
images leave the free VRAM where it was.
Not covered. The pure predicate and the generate.cpp code build on z440 (CUDA 13.0); the GPU tests are listed in the stream
notes. No model run: gates/2409-restore-guard.sh (a held-VRAM start on a layer split) shows the skip line and an
unchanged answer on the second turn.
Problem. On a layer split the later cards are left with very little VRAM after loading (#3702: 68 MiB on the last card; #3638:
268 and 666 MiB, then every batched prompt read of more than ~400 tokens ends non-finite after hours; #3797: the weight arena
fails with 320 MiB free and the message does not say who holds the rest). The window-graph eviction (12b0bce95) and the copy
stream pool (#3350) are in already; this is the rest of the code audit of what is allocated after load.
Audit. The table of every lazy allocation after load (what, when, size, how a failure is handled) is in
phaseA-tests/notes-027-gh-eng-d.md. Bugs found and fixed here, each small:
- verify.cpp capture / capture_batch and mtp.cpp finish_capture: the cudaGraphUpload and its sync were only printed. The
upload is a warm-up, but a failure (no VRAM) stayed as the thread's pending error and the next kernel wrapper's
cudaGetLastError ended the engine. The error is now cleared and the first launch uploads the graph.
- verify.cpp capture_commit and mtp.cpp finish_capture: the commit graph called cudaGraphInstantiate directly (no
eviction of older layouts, unlike the window graphs) and reported the status of the end-of-capture call, which is success,
so the message read "no error". capture_commit now goes through instantiate_evicting and reports the instantiate's status.
- instantiate_evicting instantiates into a local and sets the member only on success (what a failed instantiate leaves in
its output is not specified, and a non-null member means "captured").
- prefill.cpp carve: after a failed take of the shared region the sub-allocators had no base and each cudaMalloc'd its
buffer (hundreds of MiB in all) on the card that had just run out. carve now returns at once; the caller reports the
failure as before.
- generate.cpp: a failed set_stage_helper / set_peer is reported and the run goes on, but its CUDA error stayed pending;
cleared.
Found and not changed: relend() re-plans only the first participant (#3638 comment). Read: it runs only from the VRAM
command, which is off on a layer split, so it cannot touch a later stage's loan. Other candidates (ignored cudaEventCreate for
the pipelined fences, unchecked expert copies, cudaFuncSetAttribute flags set before success) are in the notes.
Changes for the report of #3797 and the reserve (#3702).
- include/strata/program/gpu_report.hpp: on a failed weight-arena allocation (first card and each stage) the engine now
also prints, for every card of the run, the free and total memory and the largest other programs that hold memory on it
(NVML, loaded at run time; where the driver does not list them it says so). Text only; the failure is the same.
tests/program/gpu_report_test.cpp pins the text.
- include/strata/program/stage_reserve.hpp: the arithmetic of stage_room (what the split keeps free on a card) moved out of
generate.cpp unchanged; tests/program/stage_reserve_test.cpp compares it with the old expression over a grid. The file
also holds a PROPOSAL for a larger window term on a later card (96 MiB + 6 MiB per further window size + 48 MiB, 174 MiB
for --spec 4) which nothing calls: the default placement is not changed. gates/3702-later-reserve.sh measures it.
- The low-VRAM warning of a later card now adds how many MiB the plan meant to keep free there and how far off that was.
Tests: gpu_report_test, stage_reserve_test (host only, built on z440).
Not covered: no model was run. What the ~700 MiB the plan does not count on the last card is (window graph sets, kernel code
loads) needs the run in gates/3702-later-reserve.sh.
Problem. #3948 reports that STRATA_QFUSE=1 changes the verified text and lowers MTP acceptance on an RTX 3080 (sm_86); the
control (QFUSE off against off) was noisy too. The reporter read the code and suspects the per-group counter in
fused_gr.cu gr_q8_tail (not reset when a graph replays with another T) or the nat_xq_ offsets of a window of two groups.
Audit (no model run). The counter is reset by the block that finishes the group (atomicExch before it reads the 32 values),
so a launch leaves it at zero whatever T is; two blocks own a 32-column group in every kernel that calls gr_q8_tail, the
32-column kernel (gr_up_q32_kernel) quantises on its own without a counter; the fence pattern is the usual
threadFenceReduction one. The two destinations are consistent with their consumers in verify.cpp: the attention half's
xq_ is indexed from the group's first token (the projections read it from row 0), the expert half's nat_xq_ by the window's
token (the experts read it from tb). The GDN side (verify_kernels.cu gdn_q8_1_store, rows from t_out_begin) matches its
consumer too. I found no defect by reading. One thing the tests below will show if it matters: nvcc ignores the
`#pragma clang fp contract(off)` in gdn_q8_1_store, so on CUDA the ds.sum field could differ by one ulp from the separate
quantiser (the q8 bytes would not).
Tests (synthetic data, no model).
- gr_parity: the engine's G = 2 call pattern - the window split at (T + 1) / 2, two fused_gr_read_multi calls in a row on
one stream with the same counters, group-relative destinations (xq_) and window-indexed ones (nat_xq_), direct and as a
graph replay; fresh random inputs for every one of 200 repetitions for each T = 2..8; after every run both groups' q8_1
bytes equal native_quantize_q8_1 of `mixed`, nothing outside the written ranges changed, and all counters are back at zero.
- gdn_parity: gdn_step_norm_multi with a q8_1 destination for every T and t_out_begin (verify and commit calls): the bytes
equal native_quantize_q8_1 of the output rows, the outputs and the state are the ones of the call without it, and
nothing is written past the rows. A second ctest, gdn_parity_split, runs all of gdn_parity with STRATA_GDN_SPLIT=1 (the
4-blocks-a-head step, which has its own output-norm kernel).
Not covered. The reporter's model run (QFUSE on/off, MTP, 40 prompts): gates/3948-qfuse.sh. The tests were built on z440
(CUDA 13.0, sm_86); their GPU run is listed in the stream notes.
Problem. #3602 reports a deterministic all-logits NaN with --ple-io direct on IQ3_S (Windows, two GGUF shards) where
--ple-io mmap passes; a diagnostic build found non-finite values already in the gathered PLE embedding of one token. The
bytes the model got were not the file's (the largest finite value, 7.9e6, is what a random fp16 scale gives an IQ4_NL
row). This commit audits the reader and builds the test the report asks for; it does not claim to have found that
fault.
Audit (no model run). src/ngram/ple_reader.cpp, src/platform/direct_file.cpp and the consumers (PleTable gather_batch /
prefetch_rows / collect, prefill's double-buffered ple_gather, verify's windows): rows are copied and cached under the
reader's mutex after a complete read, short reads are checked per row, straddling rows get a two-page read, the offset
arithmetic is 64-bit, the set-associative cache and the slot buffers are only touched under the lock, and prefill's two
host buffers are ordered by the ple_copied events. I found no way to hand out a row before its read completed or from a
stale buffer.
One real bug, in the error path. After a failed read the I/O worker parks for good (worker_loop), but issue() went on
queueing jobs, so the next collect() waited for them for ever. issue() now returns a ticket with no jobs once the reader
has failed, and collect() returns the error at once (and drops the ticket). Same change in the SYCL copy
(sycl/src/ngram/ple_reader.cpp).
Tests.
- src/ngram/ple_reader_stress_test.cpp (ctest ple_reader_stress): 7 table shapes (header 0/192/4000/4051, rows of 1, 90,
110, 160 and 4096 bytes, with and without bytes after the table) x 15 reader configurations (io thread or caller,
in-flight 1/2/3/64, row cache of 0, 16 and 100000 rows, injected completion delays, keep-alive), four client threads
with one to three tickets outstanding collected in any order; every delivered row is compared byte for byte with a
buffered read of the file; collected buffers are compared again later (a late write would show); rows on every page
phase, the last rows, out-of-range rows; the reader after a failed read (the file is cut short under it); with --big a
5 GiB sparse table with rows around 2^31 and 2^32 and at its end.
- table_is_consistent_under_concurrent_use in src/ngram/ple_reader_test.cpp: one PleTable used by a prompt thread
(big gather_batch), a decode window (prefetch_rows, gather_batch) and single gathers at once, against the mapped reader.
Results (z440, Linux, g++ 13, reader + direct_file + test built alone): plain build with --big OK; ASan+UBSan OK. The
error-path test against the unchanged reader hangs (the test exits 3 after 20 s); with the change it passes. ThreadSanitizer
flagged a data race that was a mistake in the test (the delay changed from a second thread in the caller-thread mode,
which has no lock); fixed in the test.
Not covered. The NaN itself: it needs a run on the reporter's configuration (gate script gates/3602-ple-direct.sh). The
Windows run of the new test is listed in the stream notes.
Gyro-L (agentionai's 50 GiB file) stores its experts as a pair trellis: 128
values per block, a fp16 scale and 64 steps of K bits (K = 5, 6, 7), two
weights per step taken from the first pair of the TQ2_T codebook entry the
step's state selects. The engine refused the file because it did not know the
three type ids. This adds them next to the existing trellis types 144-146
without changing those.
What changed:
- gguf_reader.hpp / tools/gguf_reader.py: names and block sizes (42 / 50 / 58 bytes).
- tq_cpu.*, tq_states.hpp, tq_avx2.cpp: tq_supported / tq_kbits / tq_block_bytes
for the new ids (the bytes per block are 2 + 8K for the pair layout, 2 + 4K
for the others), a scalar pair kernel (the reference) and an AVX2 pair kernel
that is bit for bit equal to it; AVX-512 uses the AVX2 pair kernel.
- iq_kernels.cu: TqInfo, the lane decode (tqp_states8, a byte window), the
dot product, the decode-once Split, the fp16 dequantizer and the format lists
for 147-149. They are reached only by a pack that holds these types.
- moe_mmq.cu, strata_ggml_tq_overlay.cmake, third_party/ggml-tq/tq.cuh,
ggml-common-tq.h: the MMQ prompt path for the three types (overlay
enumerators, traits, tile loader branch, instances).
- native_expert_parity: names tqp5-7 and skips its synthetic check, with a
message, on a ggml that has no trellis traits.
- docs/GYRO.md: the "not supported" sentence now says how Gyro-L is read.
Changes to the contributor's patch:
- tq_cpu.cpp called the AVX2 pair kernel without the x86 guard the rest of the
file has, which breaks the Metal build (it compiles tq_cpu.cpp without
tq_avx2.cpp): the pair rows go through one helper that falls back to the scalar kernel.
- moe_mmq.cu named GGML_TYPE_TQP5-7 under STRATA_MMQ_TQ, but the public
agentionai llama.cpp has the older trellis types and not these: a new
STRATA_MMQ_TQP (set for the overlay and for a tree that has the tqp instances)
guards them, so that tree still builds. The HIP overlay lists get the three
instances too, and ggml_tq_overlay_test checks them (and sets the CMake
policy its IN_LIST needs).
- native_expert_parity: the skip test compared a type-name pointer to null, but
ggml_type_name of an id past the table reads past it (the test crashed on
an upstream ggml); it compares against GGML_TYPE_COUNT.
- docs/GYRO.md keeps the plain description and leaves out the contributor's
measured speeds (we have not repeated them).
Tests added: src/kernels/cpu/tqp_cpu_test.cpp (no ggml, no model: the scalar
kernel against a bit-by-bit reference from the format definition, every ISA
equal to scalar for 1-8 tokens and for gate/up rows, the type predicates for ids
0-299 against the old formulas), the pair types in tq_pascal_bench (fp16
dequantizer bit for bit against the definition, decode kernels within 3e-2,
decode-once == per-column), tools/test_gguf_tqp.py, and native_expert_parity
tqp5/6/7 registered when the ggml tree has the tqp instances.
Checked (no model): tqp_cpu_test passes on x86 (AVX2 on z440; AVX2 and
AVX-512 on the owner PC with MSVC); tq_pascal_bench on an RTX 5070 (sm_120,
MSVC build): 0 failures, the TQP dequantizer matches the definition in all
bits, and the checksum lines of TQ2_T/TQK6/TQK7 are byte-identical to the
build without this change. Builds: CUDA sm_86 (gcc), sm_60 (CUDA 12.9),
sm_120 (MSVC), SYCL, and the CPU files for aarch64.
Not covered: the contributor's own parity runs need a ggml tree that carries the
types (not public here), and the 50 GiB model run (load check, coherent
answer, 32K prompt, MTP, Gyro-S/M identity) is still to be done. Metal, SYCL and
HIP have no kernels for the types; HIP was read, not built.
Problem (KeyanHu-git): the gated proposal needs the choice between the MTP window and a
lookup window to be fixed by the history, not by the draws of the rounds it will verify.
Freezing the policy for the whole request (the first commit) does that, but also freezes
the cost table, so a repeat-heavy request kept the window sizes it inherited. In the
author's diagnostic workload the frozen policy chose mean lookup windows of 5.87 and 3.38
on two repeat requests; with this change 4.31 and 4.05 (from the PR text; not
reproduced here).
Change: the request's copy of the DraftPolicy (source_policy) gets three calls, used only
when gated_proposal is on:
- observe_committed(): the round's accept and probe counts, applied at once. A lookup
window counts as a probe of its size (tried_), so a size whose cost is only guessed is
tried kProbes times, not every round.
- defer_cost(end, t, ms): the round's time is held. A time covers the window and the next
MTP draft, so it depends on every Philox counter up to
max(p + T - 1, p + a + 1 + S_mtp).
- apply_costs(committed): applies, in order, the held times whose end is behind the
committed position, with the same update observe() makes (including the STRATA_SPEC_COST
clamp), in the context bucket the round ran in; delayed_n_ keeps a held time from
counting as a second probe.
choose() counts the probes as cost_n_ - delayed_n_ + tried_.
Port note: the DraftPolicy on the current tree keeps its costs per context bucket
(STRATA_SPEC_COST), so tried_, delayed_n_ and the held times carry the bucket. A gated
request does not run with STRATA_SPEC_COST (it falls back to the older rule), so its
copy stays in bucket 0.
Default path: unchanged. choose() reads two new terms; on any path that never calls the new
functions tried_ and delayed_n_ are +0.0 and cost_n_ - 0.0 + 0.0 == cost_n_ exactly.
Checked by driving the old DraftPolicy (namespace-renamed copy of f-027-gh-eng) and the
new one with the same random call sequences (observe, observe_chain, observe_probs,
set_context, seed_costs, set_robust, choose, chain, keep_draft; 400 sequences of 3000
steps): 25,983,538 bitwise comparisons of picks, costs and rates, 0 mismatches; one
observe_committed call makes the same harness report 60,528 mismatches. generate.cpp:
nothing calls the new functions when gated_proposal is false.
Tests: draft_policy_test (the PR's added assertions, on the bucketed class) and
spec_admission_test pass on z440 and on the owner PC.
Not covered: the sampling argument for the delay boundary (that a held time cannot steer
a window through a counter that is drawn again) is the author's; a unit test cannot show
it, only the distribution run (gate script 3416-spec-prob-dist.sh, arm NEWLK) can.
Problem (KeyanHu-git): the rejection rule of the previous commit judges each row against
the target's probabilities p. For a native (IQ) pack the CPU experts round a token's rows
differently depending on how many tokens share its expert in the window (the grouped
kernels against ggml's one-token dot, #152). So p of one row moved with the window width,
and the same sampled text could depend on the draft length. On random blocks a grouped
Default call differs from the token alone in 700 to 12000 of about 18000 output floats
for the IQ gate/up and IQ4_NL down formats.
Change: NativeArithmetic { Default, WindowInvariant }, fixed for one pool batch.
native_gu_rows / native_down_rows become thin wrappers of one template with the arithmetic
as a parameter. The Default instance keeps every old branch (the new conditions are
compile-time true there). WindowInvariant gives each token the bits of the one-token call
on this build and CPU: kernels that already match it stay grouped (NVFP4, Trellis and
Hadamard rows, KQ256 when ggml runs its AVX2/FMA dot); the IQ rows that differ use
iq512_reference_gu_rows (types 18, 21, 22: the same decode, ggml AVX2's accumulation
order, one weight decode per group) or iq4nl_reference_rows (IQ4_NL down), and anything
else falls to the one-token dot per token. ExpertPool::run_split_multi_native takes the
arithmetic and expert_pool_dispatch_multi passes it; generate.cpp uses the invariant
pool only when the MTP drafts probabilistically (mtp.prob(), i.e. only with
STRATA_SPEC_PROB=1). Prefill, greedy decode and every other decode run Default.
The pool picks the entry point with one compare and a direct call per chunk.
Default path (STRATA_SPEC_PROB unset): unchanged, shown by running the old and the new
native_gu_rows / native_down_rows (Default) on the same random blocks: 61 format, width
and amplitude lines (IQ2_XXS/XS/S, IQ3_XXS/S, IQ4_XS, NVFP4, Q4_K, Trellis with and
without folded Hadamard, IQ4_NL / Q2_0 / Q5_1 / Q8_0 down; widths 1 to 8; row sub-ranges),
output hashes identical in 10 settings of the environment switches on z440 (AVX2) and
in 13 on the owner PC (Zen4, AVX-512, MSVC), including STRATA_NO_IQ512, STRATA_IQ_GATHER
and STRATA_IQ_MT_MIN=1. g++ codegen of the AVX-512 row kernels is not byte-identical for
5 of 87 functions (inlining and register allocation of the same source); the AVX-512
kernels were run only under MSVC.
Tests: native_arithmetic_parity, eight ctest configurations (default, AVX, KQ256,
IQ_MT_MIN=1, no512, ggml only, NVFP4 AVX2, NVFP4 ggml), registered in CMakeLists.txt and
sycl/CMakeLists.txt. It compares the invariant entry points at every group width 1..8
with the one-token calls, bitwise, for the IQ, NVFP4, Q4_K/Q5_1/Q8_0, Trellis and
Hadamard formats. All eight pass on z440 and on the owner PC; the existing CPU expert
tests (iq_avx2_parity, iq_avx1_parity, iq512_gather_parity, pool_tasks_test,
nvfp4_cpu_parity, router_dot_parity, q2_bitplane_parity) pass.
Not covered: an ARM CPU (not built here), an AVX-512 machine with fast gathers (Sapphire
Rapids), the layer-split call site, and any model run.
Problem (KeyanHu-git, refs #2028, #2591): with STRATA_SPEC_PROB=1 the default
--spec-min-p gate tests the token drawn from the draft head's distribution q, but the
verifier still judged the draft against the unfiltered q and treated a failed gate as a
plain sample from p. That shifts the output distribution: with p = (0.5, 0.5),
q = (0.875, 0.125) and threshold 0.5 it emits (0.5625, 0.4375).
Change: the gate is part of the proposal. Keep q's mass on the tokens that pass the gate
without renormalising (mu); a draw that fails the gate becomes NULL. An offered token D
is accepted with min(1, p(D)/mu(D)); a rejection and NULL draw from the normalised
positive part of p - mu; a position with no attempted proposal is a plain sample from p.
Serve and the CLI build the same packet (core/gated_proposal.hpp) and give it to the
verifier explicitly (Verifier::set_spec_proposal, sample_tokens_proposal). The explicit
call has no plain-p fallback: an invalid packet row is reported, not sampled around.
STRATA_SPEC_PROB_GATE=top keeps its normalised-q rule.
Default path (STRATA_SPEC_PROB unset): unchanged. mtp.prob() is false, so gated_proposal
is false, the new draft-result argument is null, GatedDraftLimit does nothing,
Verifier::run takes the old branch (set_spec_q clears the measure flag), and the
sampler kernel's old instance is the old code with the new branches compiled out.
Read hunk by hunk; no hunk runs a new branch or a new allocation per token. The
default-identity model run is written as a gate script and not run yet.
Port notes against the PR as submitted:
- Unsupported combinations (parallel or multi-device decode, --lookup-chain,
--adapt-swaps, PCIe share, --spec-min-p not finite, STRATA_SPEC_COST, DFlash2,
STRATA_SPEC_T_CYCLE, --spec-corrupt, Metal) used to be an error that ended the
request or the server. They now warn once and keep the older rule, so setting the
variable never stops a run. --adapt-swaps 96 and the PCIe share are on by default, so
the exact rule needs --lookup-chain 0 --adapt-swaps 0 --pcie-frac 0.
- The lookup source of a probabilistic window is chosen by a copy of the policy fixed at
the request start, without asking whether the MTP's first guess agrees. Learning from
committed rounds inside a request is the policy commit of this series.
- The CLI loop on the current tree has the out-of-memory window retry, DFlash2 and
STRATA_SPEC_T_CYCLE; the gated path does not use the retry (a narrower window would
change the room the packet was built for). Serve keeps merged first windows
(STRATA_ADMIT_MERGE_FIRST) at their length.
- docs/DETAILS.md: two short bullets instead of the PR's long paragraph (that paragraph
no longer exists on this tree). bench/results and the PNGs are not ported.
Tests: gated_proposal_test and spec_admission_test (default, pick, top, exact-top-only)
pass on z440 (CPU, g++ 13) and on the owner PC (MSVC). On the owner PC's RTX 5070
(sm_120, MSVC build, synthetic data, no model) spec_proposal_parity, spec_verify_parity and
sampler_parity (default, one-block and old-kernel configurations) pass. Also built:
CUDA sm_86 on z440 and boston, SYCL on P1006 (engine and host tests). The Metal path is
not built here.
Not covered: a model-level distribution test (chi-square of the first tokens against
plain sampling) and the identity run with the variable unset. Both are written as gate
scripts (3416-spec-prob-dist.sh, 3416-default-identity.sh) and not run.
Prefill on one card, decode on another, session handed over through RAM; exact token for token (state hash match at 4K
and 16K on 2x A4000 and 2x A6000). Nothing changes unless split_serve / --split-serve is set. On two alike cards it is
slower than the layer split on every workload measured (docs say so); it can only pay on dissimilar cards. Idea and
manual setup by MarcoTool (#4000).
Measured on 2x RTX A6000 (IQ3_S): slower than the layer split on every workload;
exactness checks pass (Q2_0 on 2x A4000, IQ3_S on 2x A6000).
Idea and manual setup by @MarcoTool in #4000.
No copy, code-constant, JSON-spelling, retrieval or unanswerable defects (retrieval 12/12, JSON 3/3, unanswerable 9/9, extraction equal to off); two long-winded answers, which the runs with the prior off also show two or three of per run. Story +15%, code +37% against off.
f-027-gh-eng gained STRATA_EXPERT_NIBBLE (4-bit expert slots, the sm_60 experimental
build): only run_impl gathers such a layer for MMQ. The layer-major walk would read
the slot as GGUF blocks, so lm_prepare now declines when any layer is converted and
the prompt takes the chunk-major path. Compiled out of every build without
STRATA_EXPERT_NIBBLE_BUILT.
The engine did not build on aarch64: the CPU expert path assumed the x86 AVX
kernels, the CPUID probes and the x86 spin/fence intrinsics. After this change
the strata target compiles on aarch64 (CUDA host code, native GGUF packs).
On x86 nothing is selected differently.
What changed:
- CMake: STRATA_CPU_X86 is decided from CMAKE_SYSTEM_PROCESSOR (an unset
processor stays the x86 build). On aarch64 src/kernels/cpu/portable.cpp
stands in for the AVX kernel sources; native experts run on ggml-cpu's NEON
dot products. -ffp-contract=off on aarch64. STRATA_ISA_FLOOR is ignored with
a warning there.
- include/strata/platform/cpu_relax.hpp: strata_cpu_pause / strata_store_fence
map to _mm_pause / _mm_sfence on x86 and to yield / dmb oshst on aarch64.
verify.cpp, pool.cpp, session.cpp, mtp.cpp and graph.cpp use them.
- expert_layout.cpp: the CPUID/XCR0 probes compile on x86 only and report no
AVX on other CPUs.
- generate.cpp: --cpu-oracle-q8-0 is rejected on a non-x86 CPU.
- setup.py: nvidia-smi's "[N/A]" memory on GB10 (unified memory) takes the
pool size from /proc/meminfo on ARM64 only; the CPU probe no longer needs
the x86 "flags" line; setup warns that the ARM64 engine is compiled on the
PC and which packs work.
Changes to the contributor's patch, found while moving it onto the 0.2.7 line:
- portable.cpp had no iq512x_rows / iq512x_supported / iq512_gather_on (added
to the x86 kernels after the PR was written): the aarch64 link would fail.
- The Metal build (Apple silicon) shares these files. It keeps its barrier
(dmb ishst), its plain router loop (its bf16_rows_dot_multi is a stub), its
Metal start-up block in generate.cpp and its "unknown" CPU name.
- The CPU side of the GPU doorbell: strata_load_fence() after the spins that
wait for a flag the GPU raised (dmb ishld on ARM, nothing on x86 and on
Apple silicon), so the data written before the flag is not read early.
- A card whose memory nvidia-smi does not report stays unlisted on x86 (the
new fallback is for ARM64 only).
- tests/terminal/test_setup_arm64.py.
Checked: with the x86 CUDA build (sm_86, gcc 13 -O3, z440) the objects of
verify, session, mtp, graph, expert_source, pool, expert_layout and generate
are byte-identical before and after (the x86 wrappers are macros for the same
intrinsics). Every C++ file of the strata target compiles for aarch64 with
clang, and the symbols of the CPU kernel library link up (syntax and symbols
only: no ARM machine, no ARM run). The Metal-shared files compile for
aarch64-macos with STRATA_GPU_METAL.
Not covered: no ARM hardware here. The tests that include x86 headers
(iq_avx2_parity and the other AVX parity tools) and the DeepSeek target do
not build on aarch64; the strata target does not need them. Memory ordering
on ARM is reasoned, not run. The Q2_0 and Gyro packs need the x86 kernels.
DETAILS.md gets two sections: the decode switches (GRAPH_PRIO, MMVQ_IL_PF) with the
two sm_80-only defaults named, and the prompt switches (PF_FUSED_NT / BATCH / F16 /
WS, GDN_TC, PLE_COMPACT / PLE_COPY_STREAM, EMBED_VRAM, DEQUANT_SIDE) with what is
bitwise and what is rounding-level. MMVQ_IL_TABLE.md gets the sm_80 row of the table
list and a paragraph on the hoisted-load tables (kIlPf_80 / kIlPf58_80, Q6_K only) and
on measuring a card with mmvq_il_parity --bench --emit-table --pf 0,3,4.
Rewritten from the PR's text: every figure is labelled as the reporter's (one CMP
170HX; we have no sm_80 card and measured none of them), and the prompt-switch
section is new (the PR documented only the decode switches and the conversation
cache).
Original docs by linxuhao (PR #3779, commit 6d6f0f39).
Problem: the interleaved small-batch matvec (native_mmvq_il) has one weight block
in flight per warp: its four `if (kbx < bpr)` sections each sit behind their own
branch, so on a GA100 (ncu: 432 GB/s on the qkv Q6_K matrix) the loads of a group
are not issued together.
Change, opt-in (every variant is bitwise native_mmvq): native_mmvq_il_pf_kernel
issues every weight load of a group (4 virtual warps x R rows) before any of the
group's dots, with the addresses clamped to the row's last block for a tail, and
with the ring variants the next group's loads before this group's dots. The
per-lane accumulation order is the plain kernel's. P2's forms replace
conversions by exact ones where the value is a small integer: (float)p becomes
p + 1.5 * 2^23 minus 1.5 * 2^23 for |p| < 2^22, and the saturating byte subtract
of Q6_K's codes (0..63, so nothing saturates) becomes two instructions. Q4_K,
Q5_K, IQ4_XS and Q6_K have the variants; Q3_K (the Q2_0 pack's type) keeps the
plain kernel.
- STRATA_MMVQ_IL_PF=1: the measured table (kIlPf_80 / kIlPf58_80: Q6_K only, read
only when the card is sm_80; on every other card 1 does nothing);
=2..6: one variant for every type and shape (A/B only, any CUDA card).
- Unset or 0: the plain kernels through the same launch code (the variant argument
is 0, and the added code is one cached environment read and a switch).
- CUDA only; HIP has no interleaved path.
Tests: mmvq_il_parity --pf a,b,.. bit-checks each listed variant against
native_mmvq for every type and shape, T = 1..8 and rows 0/1/2/4, and with --bench
times them interleaved (PFBENCH lines) and prints the joint (variant, rows) table
in the source form of kIlPf_<cc>. New ctest entry mmvq_il_parity_pf runs it with
--pf 0,2,3,4,5,6 on whatever card is present.
Reporter's numbers (linxuhao, CMP 170HX): decode +1.4 % (two reference / on / reference /
on sets) over the plain kernels; with STRATA_GRAPH_PRIO decode +1.3 % in their
per-feature pair (4K prompts +3.9 %, 16K +1.2 %, 52K +0.2 % in the same pair, which
the reporter reads as noise). Q4_K, Q5_K and IQ4_XS variants tie the plain kernels
there. Theirs, not measured here.
Original change by linxuhao (PR #3779, commit 116992ab).
Harness only; no engine code. The bench part of mmvq_il_parity times the
interleaved matvec cold (24 copies of the weights rotated past the L2), hot (one
copy), hot with other weights' reads in between (--gap MB, what a window's GEMVs do
between a prefetch and its consumer), and with the timed weights in a persisting L2
set-aside (--persist); --ab runs the five arms per case in one process, because
the clocks ramp over the first ~100 ms of a run. A 150 ms warm-up of launches runs
before the first timed case, and MMVQ_IL_TMIN bounds T from below.
The parity part is unchanged: without --bench the same checks run in the same
order (every rows choice against native_mmvq, bitwise). The bench code is CUDA only
(the test is built with STRATA_ENABLE_CUDA); the persisting-L2 calls are left out
under STRATA_HIP_GFX906.
Original change by linxuhao (PR #3779, commit 1f4b8139).
Problem: the decode window graph runs every kernel node at the launch stream's
priority, so the one-block indexer kernels of the side branch wait behind the
queued blocks of the big q GEMV, and the shared expert competes with the routed
experts.
Change, opt-in (bitwise by construction: no kernel changes, the same launches on
the same inputs; only the dispatch order moves): with STRATA_GRAPH_PRIO=1 the
side mixer's branch stream (GDN gates, QSA indexer chain) is created at the
device's highest priority, cs_ and side2_ in the middle, the shared expert's
stream at the lowest, and the window graph is instantiated with
cudaGraphInstantiateFlagUseNodePriority. If a driver did not copy the stream
priority into the nodes, the capture-time notes (the branch forks and the
priority streams' tails) set the attribute per node. One log line per captured
window prints the node priority histogram. A card whose priority range is 0..0
changes nothing.
CUDA only: the members and the three methods sit under !STRATA_USE_HIP &&
!STRATA_HIP_GFX906 && !STRATA_GPU_METAL; graph_prio_on() is false there.
Merged by hand with #3702 (the window-graph eviction): instantiate_evicting now
takes both keep_t and the flags, and the retry after freeing other window
sizes keeps the flags (the merged code would otherwise have dropped the node
priorities on that retry).
With the variable unset: every stream is created as before and the capture-time
notes return at once.
Tests: none of its own (a priority is a scheduling hint with no value to compare).
The window graphs are exercised by the engine; see the checklist for the
reporter's greedy-answer identity check.
Reporter's numbers (linxuhao, CMP 170HX, IQ3_S, --spec 3, 3K context, interleaved in
one session, each arm with STRATA_QIL=1): reference 125.3 / 126.6, GRAPH_PRIO
127.8 / 129.0, reference 125.0 / 126.8 tok/s, +1.8 %. Greedy 11 of 11 identical. In
the later per-feature pair together with STRATA_MMVQ_IL_PF: decode +1.3 %. Theirs, not
measured here.
Original change by linxuhao (PR #3779, commit 68c487c3).
One section in DETAILS.md for the switches the four conversation-cache commits
add: what each copies where, its VRAM and host memory cost, that they stay off
on HIP, gfx906 and Metal, and the measured times.
Rewritten from the PR's section: the times are labelled as the reporter's
(linxuhao, one CMP 170HX on PCIe Gen2 x4; we have no sm_80 card and measured
none of them), the independent RTX 3090 data of bitaccesscomau is added, and the
pool row describes STRATA_CKPT_POOL_READY=N instead of the PR's change to the
meaning of STRATA_CKPT_POOL=N (STRATA_CKPT_POOL=1 is unchanged).
Original section by linxuhao (PR #3779, commit a04f4399).
Problem: restoring a parked conversation copies its K/V from the host over the
link: 27-107 ms over a PCIe Gen2 x4 link.
Change, opt-in and only with STRATA_PARK_ASYNC=1 (CUDA only): a park also keeps
the whole K/V of its image in a VRAM slot of its own (64 MiB granules, up to N MiB
in all, only while 1.5 GB of VRAM stays free beside it); the background copy to
the host reads from the slot, and a restore of that image copies device to device
from the slot (about 1 ms) without waiting for the background copy. Each K/V
layer carries its own join-and-release token, so any order of destruction is safe
(the layers outlive their image when they are retained for the next park). Under
VRAM pressure the elastic K/V's growth releases every slot and the stage, then
retries once (both growth sites in generate.cpp).
With the variable unset the slot code is not reached: the budget is 0, a restore
finds dev_tag 0 and takes the host path, and the growth sites see
conversation_park_vram_release_all() return 0 and behave as before.
Tests: none of its own (it needs a card with the VRAM for the slot and a parked
conversation); see the checklist for the reporters' A/B/A conversation check with
STRATA_PARK_VRAM_MIB set.
Reporter's numbers (linxuhao, CMP 170HX): restore of a 3K conversation 28-47 ->
7-12 ms; outputs identical. The first restore of a run is slower (41.6 / 56.9 ms)
than later ones (6.5-12.6 ms) and they have not found why. Theirs, not measured
here.
Original change by linxuhao (PR #3779, commit bafca410).
Problem (reporter, CMP 170HX, PCIe Gen2 x4): a conversation switch's park still
paid 38-52 ms for the live state, because the STRATA_CKPT_POOL pool was empty at
every park (a switch takes two checkpoints), so the 118 MB GDN vector was
allocated fresh and page-faulted; and a checkpoint's 37 small synchronous
device-to-host copies queue behind the 4 MiB DMA pieces of STRATA_CKPT_ASYNC.
Change:
- STRATA_CKPT_POOL_READY=N (opt-in, 1..16, with STRATA_CKPT_POOL on): the pool
keeps N ready checkpoints, pre-sized to the session's state sizes with their
pages in, topped up on a background thread after any take (not with a layer
split).
- STRATA_CKPT_GATHER=1 (CUDA, opt-in): the small running-state vectors (ple, and
each layer's indexer tail, dead row and block position) are gathered device to
device into one buffer and fetched with one copy into a pinned buffer.
- The park stage is allocated rounded up to 64 MiB (a growing conversation does not
cudaFree/cudaMalloc it every time) and a VRAM-resident K/V skips the pointer
attribute query.
- The background helpers (pool top-up, the checkpoint DMA, the park DMA) leave the
host's pinned core (aux_cpus.hpp); no output depends on where a thread runs.
- A parked image whose live GDN state is still being copied joins before its
buffers are freed (SavedConversation::live_pending).
- STRATA_PARK_TRACE prints the joins, the live-state vectors' readiness and the
K/V sub-timers.
Differs from the PR as submitted: the PR made STRATA_CKPT_POOL=1 allocate two
ready checkpoints at startup (+236 MB of host RSS) and read STRATA_CKPT_POOL=N as
a count. That changes what an existing switch does, so here STRATA_CKPT_POOL=1
keeps its meaning (the pool holds dropped checkpoints only, nothing is allocated
ahead of time) and the ready pool has its own variable, STRATA_CKPT_POOL_READY.
A user of the PR's STRATA_CKPT_POOL=2 sets STRATA_CKPT_POOL=1
STRATA_CKPT_POOL_READY=2. The PR's one-checkpoint refill after a park with
STRATA_PARK_ASYNC is replaced by the top-up, as in the PR, and so is only active
with the new variable. Merged by hand into #2526's park_source.
With none of the new variables set: the stage size is rounded up only when the
park stage is on, the extra code is trace output behind STRATA_PARK_TRACE and
the join calls.
Reporter's numbers (linxuhao, CMP 170HX): park at a 3K context 64-75 -> 21-40 ms
(all park switches together, A/B/A conversation check identical, greedy 11 of 11
identical). Theirs, not measured here.
Original change by linxuhao (PR #3779, commit 56088e71).
Problem: parking the outgoing conversation copies its K/V device to host and
deep-copies its checkpoints before the next prompt can start: 213 ms at a 3K
context, 334 ms at 13K on a PCIe Gen2 x4 card (the reporter's CMP 170HX).
Change, opt-in and CUDA only (off on HIP, gfx906 and Metal):
- STRATA_PARK_ASYNC=1: conversation_kv_save copies the device K/V device to
device into a VRAM stage (only while 1.5 GB stay free beside it; otherwise the
save copies as before), and a background thread moves the stage into the host
buffers in 4 MiB pieces through two pinned buffers. conversation_park_join
waits for it: every reader of a parked image's K/V, the cache before it frees an
entry, and the image itself (SavedConversation::pending) before its buffers go.
ConversationBuffer::resize_for_overwrite leaves the added bytes unfilled for
that copy; the buffer's segments use an allocator that value-initializes only
when a fill value is given, and resize(n, value) is unchanged for every other
caller.
- With it, the checkpoints this request drops right after the park (all of them
when another state comes in, else the ones off this prompt's path) move into
the parked image instead of being copied, and a background allocation refills
STRATA_CKPT_POOL for the next save. Only the main session's checkpoints move,
not an idle batch slot's (the park_source of #2526 serves both).
- conversation_snapshot_save takes dropped checkpoints' storage (the existing
STRATA_CKPT_POOL) for its copies; with the pool off, nothing is taken.
- Unset: park_drops is null, no storage is taken, the K/V copies are the
synchronous ones; the added work is the K/V size sum for the stage request and
a few join calls.
Differs from the PR as submitted: the PR's join in conversation_file.cpp
(put_payload) is not ported; the SAVE/RESTORE handler of generate.cpp joins the
park first (see the previous commit, which carries the PR's 33a8146a fix).
Hand-merged into the park_source / ParkSource refactor of #2526.
Not covered: if the background copy fails after resize_for_overwrite, the
parked image keeps unfilled bytes; the failure is logged. A restore of such an
image would read them. No test of the park path with a failing copy.
Tests: the PARK_ASYNC path has no unit test in this commit (it needs a card with 1.5 GB free for the
stage); the reporter's A/B/A conversation check covers it, see the checklist.
Reporter's numbers (linxuhao, CMP 170HX): park 213 -> 64-75 ms (3K), 334 -> 68 ms
(13K); greedy answers identical on their 11 prompts and the A/B/A conversation
check. Theirs, not measured here.
Original change by linxuhao (PR #3779, commit 3e0c141c).
Problem: on a slow host link a turn checkpoint copies the whole GDN state to
the host on the request path (118 MB at 1.64 GB/s on a PCIe Gen2 x4 card, 72 ms;
an independent RTX 3090 on PCIe 3.0 x16 showed 73 ms to 62 ms with the switch).
Restoring a parked conversation copies it back.
Change, both opt-in and CUDA only (off on HIP, gfx906 and Metal):
- STRATA_CKPT_ASYNC=1: the GDN state is copied device to device into a VRAM stage
(one buffer the size of the state, allocated at the first checkpoint save, plus
a pinned host buffer of the same size; if either allocation fails the save falls
back to the synchronous copy), and a background thread moves it in 4 MiB pieces
into the checkpoint's host vector. Every reader joins first and only for the copy
it needs (restore, snapshot save and restore, the parked copies, the serve loop at
each request start, the SAVE/RESTORE session handler).
- STRATA_CKPT_VRAM_SLOTS=N (with ASYNC): the last N saved GDN states stay in VRAM,
tag-checked, only allocated while 1.5 GB stays free; a restore of a checkpoint
whose tag a slot holds copies device to device.
- STRATA_TRACE lines carry a steady-clock timestamp; STRATA_PARK_TRACE prints the
timing of the save steps.
- Unset: the code adds a mutex-guarded no-op join per checkpoint save, restore and
snapshot, a tag field that stays 0 and a few clock reads; the copies are the
synchronous ones as before.
Differs from the PR as submitted: the PR put a join into conversation_file.cpp
(put_checkpoint), which broke the conversation_file_test and
session_save_reclaim_test builds without CUDA and matched only the copied vector,
not the one the copy in flight writes, so a SAVE within one checkpoint copy after a
turn could write an incomplete GDN state (found by bitaccesscomau; fixed in the
PR's last commit 33a8146a). This commit carries that fix: conversation_file.cpp
stays as it is and the join runs first in the SAVE/RESTORE handler of
generate.cpp. The commit 33a8146a is therefore not a separate commit here.
One more join, found by running the new ctest entries: conversation_snapshot_sources (the SAVE
path of a session file) saves the running state with the background copy and the caller writes
it to the file at once, after the handler's join has already run. It now joins the copy of
that checkpoint before it returns; without it a session file written with
STRATA_CKPT_ASYNC=1 could hold an incomplete GDN state (conversation_snapshot_test failed
at "both files read" before the join).
Tests: conversation_snapshot_test now joins after each save and, with
STRATA_TEST_BIG_GDN=1 (a GDN state over 1 MiB, so the async path is taken), runs as
the new ctest entries conversation_snapshot_test_ckpt_async and
conversation_snapshot_test_ckpt_vram_slots: every checkpoint, capture and restore
must give the same bytes as the plain run.
Not covered: no test destroys a checkpoint while its copy is in flight (readers
join; the cache's eviction paths do not); the reporter saw 3 of about 35 engine
starts stop answering after a park with evictions, on the paths with the
switches off, and could not reproduce it in 15 reruns. The VRAM stage has no
headroom check of its own (an RTX 3090 with 147 MiB minimum free showed no failure
in 42 of 42 saves; the allocation failure falls back silently).
Reporter's numbers (linxuhao, CMP 170HX): turn checkpoint 71.4 -> 0.9 ms,
restore 114-126 -> 28-47 ms. Independent data (bitaccesscomau, RTX 3090 24 GB,
PCIe 3.0 x16, release defaults, 4 pairs): checkpoint save 73.2 -> 62.4 ms, tiny-prompt
time to first token median ratio 1.038, outputs byte-identical. Theirs, not measured
here.
Original change by linxuhao (PR #3779, commits 0d512925 and 33a8146a).
Problem: the prompt path's DeltaNet recurrence is a large part of the prompt time
of a GA100 card.
Change, opt-in and rounding-level (the output is not the FP32 kernels' bits):
src/prefill/gdn_tc.cuh holds the chunked WY-form recurrence with every product
on the tensor cores (f16x3 splits with power-of-two scaling, one block per value
head) and its prep kernel; kernels.cu dispatches it before the STRATA_GDN_CHUNKED
branch for prompt chunks of 128+ tokens when STRATA_GDN_TC is set.
Where it runs and where it does not:
- gdn_rec_tc returns "not supported" and the existing kernels run unchanged on a
card below sm_80, with fewer than 48 SMs, or with less than 122 KB of
shared memory a block (the kernel needs 124,944 B). That excludes sm_86, sm_89
and sm_120 (99 KB opt-in): the code is compiled into those builds and never
launched there. It runs on sm_80 and sm_90. A failed scratch allocation
(60.3 MB a device) also falls back.
- Device code for cards below sm_80 compiles to trapping stubs behind
STRATA_GDN_TC_SM80; the whole header sits under !__HIPCC__.
- Unset, the dispatch costs one cached flag test per call.
Tests: gdn_rec_parity gains the tc variant (T = 1..4099, 8192, 32768, six stress
inputs against gdn_rec_kh_kernel, gate 1e-4; "not on this card" where it cannot
run), --selftest-frag (each tensor-core helper on one warp against a double
reference) and --bench-tc. The new ctest entry gdn_rec_parity_selftest_frag runs
the helper self-test, which any sm_80+ card can run (77 below sm_80). The full
tc variant can only run on an sm_80 or sm_90 card.
Reporter's numbers (linxuhao, CMP 170HX): recurrence 105 -> 34 ms at 3.3K tokens
and 1,705 -> 547 ms at 52K; prompts +3.9 % (4K), +6.0 % (16K), +6.7 % (52K) over
the same base; within 2.3e-6 of the FP32 kernel in gdn_rec_parity; teacher-forced
KL against the bitwise base mean 0.0157, median 0.0011, max 0.098, top-1 100 %;
1 of 5 conversation checks identical and decode -2.6 % in their pair (they
think draft acceptance moves with the rounding; not checked). Theirs, not
measured here.
Original change by linxuhao (PR #3779, commit f1cabc99).
Problem: the NT = 256 kernels still decode every 32 values to int8 and scale
the int32 products in FP32.
Change, both opt-in and both rounding-level (the output is not bit-identical to the
int8 kernels; they are the PR's "rounding-level" switches):
- STRATA_PF_FUSED_F16=1 (implies the NT = 256 set): the decoded weights times
their block scales and the activations times theirs as F16, m16n8k16 F16
products into FP32 accumulators. The k order inside a 16-value group is
permuted the same way for both operands.
- STRATA_PF_FUSED_WS=1: native_ws_kernel, producer warps decode into an F16 ring
in shared memory and consumer warps run the same F16 products (named barriers),
so it has the F16 path's rounding. One block per multiprocessor; about 101 KB of
shared memory at its defaults, so it is set up on sm_80 / sm_90 cards and
says "not available" and keeps the 512-thread kernels on an sm_86 / sm_89 /
sm_120 card (99 KB). Compile-time knobs STRATA_WS_PROD / _SLOTS / _NB / _KS / _HTAB.
- Both take a WideBatch like the NT = 256 kernels, so STRATA_PF_BATCH above 128
reaches them. CUDA only (under !__HIPCC__); the Turing pass compiles the F16
instances against __trap() stubs and never runs them (dev_info stops before the
NT = 256 setup on Turing).
- With neither variable set the kernels, their launches and their shared-memory
sizes are as before (the F16 template parameter defaults to false).
Differs from the PR as submitted: when STRATA_PF_FUSED_Q2N=1 is on as well
(#3809, after the PR was written) the WS setup also sets up the Q2_0 gate/up
instance.
Tests: prefill_fused_iq_test_f16 and prefill_fused_iq_test_ws (ctest) run the
fused-experts test (double-precision reference and MMQ) with the switch set. The
test now skips (77) when a requested NT / F16 / WS kernel set is not set up on the
device, so a card that cannot run WS does not report a pass for the 512-thread
kernels.
Limits of the test for these two arms only: with STRATA_PF_FUSED_F16=1 or _WS=1 the
test uses the limits it already has for STRATA_PF_F16 (F16 products): fused against the
int8-rounded model rel RMS 5e-3 (2e-3 otherwise) and against MMQ 5e-3 (1e-3 otherwise).
Measured with F16 on an RTX A4000 (sm_86, z440), over the 11 pairs: against the int8-rounded
model rel RMS 2.85e-3 to 2.90e-3 and worst row 5.9e-3 to 8.1e-3; against MMQ the same;
against the FP32 reference 1.438e-2 to 1.480e-2 with MMQ at 1.435e-2 to 1.478e-2 (the
fused path stays within 1.003x of MMQ's rel RMS and 1.03x of its worst row). The rounding is the F16 rounding of the
weights times their scales and of the activations times theirs (an 11-bit mantissa, about
2^-11 relative per operand), which the int8-rounded model does not have; it is not a
different summation order. The same arm on an RTX PRO 5000 (sm_120): up to 2.90e-3 (rel RMS), worst row up to 8.1e-3.
The default path and the NT=256 arm keep the old limits and their printed errors are
unchanged: with no switch set the test prints identical output (52 lines) from the base
tree and this tree, on an RTX A4000 (sm_86) and on an RTX PRO 5000 (sm_120).
Reporter's numbers (linxuhao, CMP 170HX): one MoE layer of 3,256 tokens 8.27 ms
(NT = 256) to 7.90 ms (F16) and 7.65 ms (WS); prompts +2..+6 % over the
256-thread kernels, 1 of 11 greedy answers changes. Teacher-forced KL against the
bitwise base on their 20-prompt set: F16 / WS mean 0.0062, median 0.00006, max
0.076, top-1 100 %. Theirs, not measured here.
Original change by linxuhao (PR #3779, commit 579cbfbb).
Problem: a layer split with --batch pipelines the slots in groups, and the pipelined groups run no MTP drafts.
--batch-mtp with serial windows leaves all but one card waiting. A multi-stream deployment wants both.
What changed: STRATA_PIPED_MTP=N (with --batch-mtp) keeps the pipeline and drafts in it. A group's window carries
each of its slots' rows: the slot's token and up to N MTP proposals (at most --spec - 1; the first always, each
further one while the draft layer was at least --spec-min-p sure of the one before). The last stage judges each
slot's proposals on their own. The number of rows to keep is known only then, and by then the earlier stages run
other groups' windows, which reuse the scratch the commit replays from. So a small save graph copies those rows
right behind each window (Verifier::batch_launch_rows), and every stage queues the commit of the kept rows after
the verdict (Verifier::batch_commit_rows). The slot drafter catches up on its own rows and drafts the next
proposals as an asynchronous chain (MtpDrafter::chain_launch). Unset, nothing changes.
Changes on top of the PR as sent (review):
- The verifier window size is worked out where the verifiers are built, from the slots that fit and the final
groups, not earlier from --batch as asked. Unset it is the old expression.
- STRATA_PIPED_MTP is decided again once the slots that fit are known. It says it is off, and why, when fewer than
2 groups are left, when STRATA_BATCH_SELF_COMMIT is on, or when a group's window would hold more than
kVerifyMaxT rows. The start line names groups and slots correctly.
- batch_groups::resolve keeps its old mtp_dropped rule; only mtp_piped is new, and it also applies with
--batch-groups given (the PR's fix 1). The "--batch-groups 1" test case now expects "not dropped".
- --remote-expert-opt alone no longer blocks --batch-mtp on a layer split: the object it guards is only made with
a helper expert cache (the PR's fix 2).
- docs/BATCHING.md rewritten with the numbers measured here.
Tests: batch_groups_test (the PR's cases, four "unset is as before" cases); builds on CUDA sm_86 (Linux), SYCL
(the sycl tree has its own verify.cpp and generate.cpp and does not include this code) and MSVC. Model runs:
see docs/BATCHING.md and the gate scripts 4121-*.sh.
Not covered: AMD (HIP) was not built or run; the draft layer on the first stage; helper expert caches.
2.0 showed copy and code defects on the RTX 3060 (IQ3_XXS); 1.5 showed none in 19 tasks and keeps the speed
(story +15%, code +37% vs off). The 61-task check decides 1.5 vs a 1.0 fallback.
+6.9% Qwen IQ3_S decode on Strix Halo Windows, prompts unchanged; token ids identical in 4 starts x 5 prompts.
Linux gfx1151 and other cards unchanged; STRATA_DEFAULTS=026 restores it. Docs label the DeepSeek Strix numbers by build.
The Windows guide says what 0.2.7 does and how to turn the switch back on, backed by five fixed greedy prompts
(895 ids) run four times with the switch on, off, on, off: identical ids. STRIX_HALO.md and the 0.2.7 release notes
get a line. The DeepSeek 4K prompt figures of 2026-10-25 (78.7, 87.2) stay and are labelled with their build; a table
of the 0.2.7 candidate changes one thing at a time (chunk size, DS4_DEFAULTS=026).
Interleaved A/B on a Strix Halo Windows PC (Qwen IQ3_S, engine 0.1.51): chat decode 43.6 tok/s with the fork,
46.6 without (+6.9%), prompts unchanged. Linux gfx1151 and every other architecture keep their table.
STRATA_SH_STREAM=1 turns it back on, STRATA_DEFAULTS=026 keeps v0.2.6's table.
arch_default_env(arch, defaults026, windows) is a pure rule; arch_defaults_test checks both platforms.
Ported PRs (authors kept) and our fixes, plus the owner's part-2 items: the restored TROUBLESHOOTING entries with a
link check, #3895 (no shared-expert fork calibration on Windows HIP with --kv-resident; a warning on a hand-set
STRATA_SH_STREAM=1), #3930 (docs), #3421 (Docker stops when models are present and MODEL_PATH is unset), #3570
(launch --show-key reads setup's config), #3311 (engine paragraph only) and #3029 (Strix Halo Windows guide,
re-measured on 0.2.7).
Structure from the guide in PR #3029 by HaiYunChuang (what the PC reaches, memory and the BIOS, the configuration, what is on by default, traps),
rewritten for the 0.2.7 install and with every number measured again: the Terminal and the gfx1151 engines built from f-027-int2 (3e468f5de), a
local signed test feed, strata launch as a standard test user on the NucBox EVO-X2 (Adrenalin 26.10.1), tools/strata_bench.py over a tunnel.
Qwen IQ3_S: 44 tok/s chat, prompts 755 / 821 / 792 tok/s at 4K / 16K / 32K. DeepSeek 2.00 bpw: 12 to 14 tok/s chat, prompts 214 to 342 tok/s.
An interleaved A/B of STRATA_SH_STREAM on Windows gfx1151 (three pairs): off is 6.9% faster in chat decode, prompts unchanged.
What launch chose and logged is quoted instead of a hand-made config. The first guide's tuning rows (engines 0.1.41 to 0.1.47) are not repeated.
Co-authored-by: HaiYunChuang <[email protected]>
Part 1 fixes and 12 ported PRs (authors kept), and the owner's part-2 items: the Monitor readings strip with
local currency (#3628), approval notification + opt-in minimise to tray (#3477), opt-in done sound (#3586),
app port setting with a busy warning (#3402) and interface size up to 200% (#3970). Host tests: only the two
known base failures on Linux; connector-setup and the new tests pass on Windows.
Problem: Gemm::native dequantizes each dense weight to FP16 in front of its
GEMM on the compute stream, so the dequant never overlaps the kernels queued
ahead of it.
Change, opt-in (CUDA only, always off on HIP): with STRATA_DEQUANT_SIDE=1 the
prompt path takes a second GEMM scratch (64 MiB more of its region, counted in
bytes_needed and carried through relayout) and Gemm::set_dequant_side hands it
to the Gemm. native() then dequantizes the next weight on a non-blocking side
stream into the scratch the compute stream read last but one. Four events order
every write and read of each scratch (free -> dequant -> ready -> GEMM -> free),
so each product sees the same bytes as in the serial form. The row-sliced path
(a weight larger than the scratch) stays serial. Metal answers false and keeps
the serial form.
Differs from the PR as submitted: if the side stream, its events or the
setup fail, the PR failed the prompt-path init. Here one line is printed and the
dense weights are dequantized in front of their GEMMs as before. Unset, nothing
changes (gemm_dequant_side() is one cached flag; the scratch, stream and events
are not created).
Test: new gemm_dequant_side_parity (ctest, GPU, synthetic, 77 without a device):
eight Q4_K / Q5_K / Q6_K weights of the engine's shapes (one larger than the
scratch, so it takes the sliced path) run three rounds through Gemm::native
serially and with the side stream, with and without a spin kernel queued ahead of
each GEMM; every output must be identical bit for bit.
Not covered: dense weights read by the side stream are the resident ones; a weight
that arrived by an upload on another stream would need its own event (the prompt
path has none that do).
Reporter's numbers (linxuhao, CMP 170HX): measured with STRATA_PLE_COPY_STREAM and
STRATA_EMBED_VRAM, the four PLE / embedding / dequant switches together give
+2.2..+2.6 % prompts. Theirs, not measured here.
Original change by linxuhao (PR #3779, commit 741fc6ab).
DS4_DEFAULTS=026 behind the DeepSeek speed-ups Off toggle, the 12 GB NVIDIA route-prior row, the Advanced-only
Qwen '0.2.6 behaviour' row (STRATA_DEFAULTS=026), and the app-ui Quit test fix. Terminal suites pass; host suite
failures on Linux also fail on the base.
Owner decision for 0.2.7: the tables of PR #3453 (the cells where an RTX 4070 and an RTX 4080 community run gave the same rows count, and
the one-column Q3_K row) ship on for sm_89 cards, untested on Ada by us. The output is bitwise the same with them off (a table only
changes how many rows a warp takes); no end-to-end gain was seen on a 4090 or 4080 in the report, so the reporter tests 0.2.7 and reports.
The rule is a pure function in sm120_series.hpp: mmvq_il_89_default(hip, cc, v026) is true for cc 89 only, false in HIP builds and under
STRATA_DEFAULTS=026 (defaults_026.hpp, as 38c9080db did for the other defaults); mmvq_il_89_resolve() adds the variable: "1" and "0" win,
unset or other text is the rule. native_mmvq.cu reads it once. arch_defaults_test checks cc 89 on, every other cc off, HIP off, 026 off,
"0" off, "1" wins over 026, the empty text. mmvq_il_parity expects the on-tables by default and the old ones with =0 or 026.
Docs: MMVQ_IL_TABLE.md, DETAILS.md and the 0.2.7 release notes (a line that says it is not measured on an RTX 40 card, with the off switch).
French catalogs for the app (3,528 strings + fix/tune/tools catalogs) and the served web page (247 keys),
README.pl.md from the current README (#3355, Jakub Luwierski), Polski in every README language row, and the
README.md language row's double-encoded names fixed. Tests: 21 host files 180 pass on p1006, Terminal and web
i18n checks on l620.
serve/web/locales/fr.json (247 keys, plurals one/other, same terms as the app catalog) and the French line in serve/web/i18n.js. tools/web_i18n.py and serve/test_web_i18n.py (node and python) pass.
Ports the Polish translation from #3355 and the Polski entry in the language row of every README. The pull request translated the older long README, so README.pl.md is rewritten from the current README.md (logo, centred language row, DeepSeek speed tables including Strix Halo, Gyro, coffee button). README.md's own language row had garbled names (double-encoded text); it is written out correctly.
Co-authored-by: Jakub Luwierski <[email protected]>
fr.json, fix-fr.json, tune-fr.json and tools-fr.json (3,654 strings, vous, plain short words), the Francais entry in the language picker, every locale test and gap check now covers fr, a french-catalog test (spacing before ? ! : ;, apostrophes, short labels, no leftover English), scripts/fr_sync.py to list and merge strings added later, and scripts/language-shots.mjs for screenshots of the app in one language.
Wide/narrow CSS (#3176), page translations with a language choice, opt-in KaTeX math (#2972, #2974),
sandboxed svg/html preview (#1628), Load/Unload in the header (#3550), and one chat-history design
(#1787, #2615, #3823, #4004). Tests: browser suite on the owner PC, node + pytest serve suites on boston
(457 passed, 2 skipped).
Problem: on a GP100 die the verify window's gate/up kernels for the IQ2_S,
IQ3_XXS and IQ3_S expert formats are not held by memory bandwidth but by each
block's serial chain (unpack the codebook indices, the emulated dp4a, the
16-lane reduction); the same kernel with every dot product removed still takes
about two thirds of its time.
Change: with STRATA_EXPERT_NIBBLE=1 the engine, after the expert cache is
filled and every expert is resident, transcodes the gate/up blocks of every
fully resident IQ2_S / IQ3_XXS / IQ3_S layer (n_embd 2560, n_ff 640) into slots
of 4-bit codes, 16 bytes per 32 weights, with the integer scales and fp16 d kept
as they are; the down rows stay GGUF. The decode gate/up reads those slots with
kernels that run the production integer chain and float tail in the production
order (R4S / R4P / R4SV / R4PX by format and group size); the prompt path's MMQ
gather rebuilds the GGUF blocks from a slot on the fly. At load every item is
round-tripped and every slot's rebuilt GGUF blocks are compared with the staged
blob before the cache is sealed (ExpertCache:🦭 a later fill_slot stops the
engine). New: include/strata/core/expert_nibble.hpp (slot contract, switches,
per-layer table), nib_gu_proto.cuh (STRATA_NIB_GU_PROTO=1 variant), the
transcoder, hash and rebuild kernels and the 4-bit gate/up kernels in
iq_kernels.cu, generate.cpp (the fill, the VRAM accounting, STRATA_EXPERT_NIBBLE_
RECHECK), iq_nibble_parity and native_expert_bench --nibble.
Review changes against the PR:
- An option the switch cannot work with (--peer-device, --kv-grow, --vram-elastic,
remote caches, STRATA_MMQ_DIRECT, ..., a card that is not cc 6.0, a build
without the kernels) now prints a warning and the start goes on with the usual
GGUF slots, where the PR refused the start. What is known only later (an expert
out of VRAM, a loan, the prompt chunk) still refuses, since slots are
converted by then.
- The checks this switch adds to expert_source.cpp and prefill.cpp are compiled
out of every build that has no 4-bit kernels, so the default builds add nothing
to the hit split or the prompt loop.
- iq_nibble_parity gets --synthetic-only (every gate/up format x both down
types on random experts, no model file), skips with 77 on a card that is not a
GP100, and the synthetic and --header runs are ctests.
- The PR's docs section is rewritten for these changes and its measurements are
labeled as the author's.
Default path (switch unset): the table is empty, nibble::slot_bytes(l) is
expert_layout().blob_bytes(l) and nibble_max_slot_bytes() is max_blob, every
sizing the PR touched returns what it returned before, and the new branches in
generate.cpp (guard_raw before the adaptive tier's copies, the profile fill's
test), expert_cache.cpp (one bool in the fill calls and admit) and verify.cpp
(a capture-time layout choice, compiled out without the kernels) are not taken.
Proof on the kernels: in the sm_60 build (CUDA 12.9) the SASS of all 446
kernels iq_kernels.cu had before is identical with and without this change,
as is that of the 603 in native_mmvq.cu, 29 in qsa_select.cu and 10 in
moe_mmq.cu; 21 kernels are new.
Tests: iq_nibble_parity --header passes on z440 and boston (the host half of the
contract; also a ctest). The GPU parts (--synthetic-only, a model shard) refuse a
card that is not a GP100, so they were only compiled here (sm_60, CUDA 12.9,
P1004: iq_nibble_parity and native_expert_bench build; the whole tree builds).
sm_86 build without the SM60 option (z440): exit 0, the 4-bit paths compiled out.
Not covered: no GP100 run. The model gate is
phaseA-tests/gh027/gates/3455-expert-nibble.sh: iq_nibble_parity --synthetic-only
and on a model's layers on a P100, then identical replies with the switch on and
off (a layer is converted only when every expert of every layer is resident, so
the rig needs the pack to fit whole), then the decode A/B. The contributor's
numbers (four GP100 dies, Swift 1.5 IQ3_S, 47 of 48 layers converted, +12.7 GiB
VRAM, decode 47.3 -> 51.9 tok/s, +9.7%; prompt reading 2.3-3.0% slower; replies
of 15 requests equal) were not re-measured, and the enabled path was written
against an older base than this branch's.
Problem: on a GP100 die (compute capability 6.0: no dp4a, no tensor cores) the
prompt pass's routed experts run as MMQ launches per 16-expert group: a
gate/up, a swiglu, a quantize and a down launch per group, each group's experts
gathered and padded to the group's largest row count, every MMQ tile decoding
its weight blocks again. With every expert resident the per-group overhead, not
the weight bytes, is a large share of a prompt read on this card.
Change: with STRATA_MMQ_WS=1, a layer whose routed experts are all resident on
this GPU runs quantize_scatter, then gate/up as ONE weight-stationary launch
over all its experts (read where they lie in the expert cache), ONE
swiglu_quant over all its rows, then the down as ONE launch. New
src/prefill/moe_ws*.cu(h): a CTA holds 128 (or 256) weight rows of one expert
and runs them against up to 64 of that expert's activation rows, a thread per
weight row, so each weight block is decoded once per pass; gate/up for
IQ2_XXS, IQ2_XS, IQ2_S, IQ3_XXS and IQ3_S, down for Q2_0 and IQ4_NL. Every
product keeps MMQ's int8 values, scales and per-32-value epilogue in its K
order, so GU, H and the down rows are MMQ's bits. ws::faster carries the
author's measured ws / MMQ time ratios per format and picks a layer only where
both products win at its rows per routed expert. prefill.cpp decides per layer
after quantize_scatter (all experts resident, this GPU alone, no peer or CPU
share, not the fused layout, not STRATA_MMQ_DIRECT, no Hadamard, no NVFP4 tail,
no FP4 activations, aligned, not STRATA_DBG_NAN); every other layer keeps its
path. STRATA_WS_FAST=1 (second opt-in, same bits) selects a down kernel for the
IQ4_NL 32-token tier with fewer instructions per pass.
Gating: the kernels and STRATA_PREFILL_WS exist only in the CUDA build with
STRATA_EXPERIMENTAL_SM60 (not HIP, not Metal; prefill.cpp links stubs that
answer mode 0 elsewhere). At run time ws::mode() is 1 only with STRATA_MMQ_WS=1
and every visible GPU at compute capability 6.0. Unset, nothing is allocated
(the offset and chunk buffers, under 1 MiB per die, exist only when the mode is
not 0), the arena layout and the mapped table are the ones before, and the
per-layer test short-circuits on the cached mode.
Not ported: moe_ws_down_bench (needs a model shard) and its CMake target.
Added: ws::set_mode (tests only) and tests/cuda/prefill_ws_parity_test.cpp, a
synthetic layer-level check: random experts of all 5 gate/up x 2 down formats,
random routing, the per-group MMQ (mmq::Context per expert) against ws::gate_up
+ swiglu_quant + ws::down, GU, Hq and Dm compared byte for byte, the down also
alone on the reference H and with STRATA_WS_FAST on. It gives its verdict on a
compute capability 6.0 card only (the kernels follow MMQ's sm_60 SASS); other
cards run it for information with --any.
Tests: prefill_ws_parity_test on boston (2x RTX A6000, sm_86, CUDA 13.1, gpu lock,
SM60 build, --any because it is not a GP100): 40 cases (all 5 gate/up x 2 down
formats, four routings each), 0 declined; the largest GU and down-alone error
against that card's MMQ is 5.2e-07 of the largest value (float rounding: MMQ's
sm_86 epilogue is not the sm_60 one the kernels follow), the full chain's Dm
differs up to 1e-3 where int8 rounding of H flips on those rounding
differences, and STRATA_WS_FAST output equals the default kernel's in all
cases, 0 differing bytes. The byte-for-byte verdict against MMQ is the P100's.
sm_60 compile of the whole tree (CUDA 12.9, P1004): exit 0; sm_86 build without
the SM60 option (z440): exit 0. SASS of the 10 kernels moe_mmq.cu had before
this change is identical in the sm_60 build.
Not covered: no GP100 run here. The contributor's numbers (prompt reads 31-37%
shorter on four GP100 dies with every expert resident: 240 tokens 2,156 -> 1,364
ms, a fresh 26K read 41,231 -> 28,555 ms; one run per arm) were not re-measured.
The model gate is phaseA-tests/gh027/gates/3752-mmq-ws.sh.
Problem: the token-embedding table of a native pack lives in mapped pinned host
memory, so every prompt's scattered row gather crosses the PCIe link. On a Gen2
x4 link that costs more than the table's size in VRAM is worth.
Change, opt-in: with STRATA_EMBED_VRAM=1 the table is copied into the current
device's VRAM when the table plus 512 MiB stay free (322 MiB for an IQ4_XS
table). The gather kernels read the same bytes through dev_ as before. If the
VRAM is not there or the copy fails, the table stays in mapped host memory as
before and one line says so. The start-up line now says where the table is
("in VRAM" also for the two existing VRAM fallbacks, which it called mapped
host memory).
With the variable unset nothing changes (one getenv at load).
Not covered: a layer split across several devices places the table on the
device that is current when it loads, as the existing VRAM fallback does. No
unit test (it moves the same bytes to a different memory; the gather kernel is
unchanged).
Reporter's numbers (linxuhao, CMP 170HX): measured together with
STRATA_PLE_COPY_STREAM and STRATA_DEQUANT_SIDE on a 51.8K-token prompt, the
embedding step went from 191-201 to 70 ms. Theirs, not measured here.
Original change by linxuhao (PR #3779, commit 952dd784).
Problem: with STRATA_PLE_COMPACT the raw-row upload sat on the compute stream
before layer 1, which idles for the DMA (4.7 MB per 3K-token chunk, 2.9 ms on a
Gen2 x4 link).
Change, opt-in, only with STRATA_PLE_COMPACT=1: the raw rows go up on a
non-blocking stream of their own once the previous chunk's dequant has read the
device buffer (ple_raw_free event); the compute stream waits for the upload
(ple_copied) only at the dequant. Same bytes, same kernel. The stream is drained
when a run ends so no upload is left in flight from a buffer the next run
refills.
Differs from the PR as submitted: if the stream or the event cannot be created,
the PR failed Prefill::init. Here it prints one line and the rows go up on the
compute stream as with STRATA_PLE_COMPACT alone. With the variable unset
nothing changes (the stream and the events stay null).
Reporter's numbers (linxuhao, CMP 170HX): PLE time of a 51.8K-token prompt 95-103
to 47 ms, measured together with STRATA_EMBED_VRAM and STRATA_DEQUANT_SIDE
(+2.2..+2.6 % prompts for the four PLE/embedding/dequant switches together).
Theirs, not measured here.
Original change by linxuhao (PR #3779, commit 0ba47e61).
Problem: the prompt path reads 16 PLE rows a token, dequantizes them on the
host and uploads 10,240 floats a token. On a slow host link (PCIe Gen2 x4) that
upload is on the critical path of the first layer.
Change, opt-in with STRATA_PLE_COMPACT=1: for an IQ4_NL table held in memory
(--ple-io ram / mmap) the prompt path copies the raw 90-byte rows (1,440 bytes
a token) into the same pinned buffer, uploads them and runs
ple_iq4nl_dequant on the device (d times the code's value, one multiply, the
split-half nibble order of iq4nl_dequant_row). Any other table (Direct I/O,
pageable fallback, another format) and a failed allocation take the host path
as before. With the variable unset nothing changes: the new code is behind one
cached flag, and the extra members of the Prefill state stay null.
Two things differ from the PR as submitted:
- PleTable::gather_raw_iq4nl writes a row past the table as scale 0 and code 8
in every nibble. The PR wrote zero bytes, which the device decodes to
-0.0f (0 times code 0's value -127), while the host path writes +0.0f for such
a row.
- New test ple_iq4nl_dequant_test (ctest, synthetic, no model, 77 without a
device): the device decode against iq4nl_dequant_row, bit for bit, over every
fp16 scale pattern with random codes, random rows, an odd row count and the
out-of-table row.
Reporter's numbers (linxuhao, CMP 170HX, PCIe Gen2 x4): prompts +1.0..+1.7 %
(4K, 16K) with the switch on. Theirs, not measured here. On a Gen4 x16 card
the gain will be smaller.
Metal gets a stub (the Metal prompt runs through its window and never calls
it).
Original change by linxuhao (PR #3779, commit da818261).
Problem: the prompt path stores H (swiglu of the gate/up rows) as floats and
then quantizes it to the q8_1 rows the down product reads: two kernels and an
H round trip per MMQ group.
Change: swiglu_quant_kernel, a block per row (128 threads, 4 values each,
blockIdx.y the row's 512-value half), computes swiglu_kernel's expression for
each value and quantize_mmq_q8_1<D4>'s result for each 32-value block, so hq
holds the same bytes and H is never stored. swiglu_quant_ok(type) says where
it applies (a q8_1 D4 down type, no FP4 activations). Built only in the CUDA
build with STRATA_EXPERIMENTAL_SM60; HIP and the other builds get
swiglu_quant_ok() == false and a no-op. Nothing calls it in this commit; the
weight-stationary path of the next commit does. prefill.cpp gets the matching
stubs for builds without the MMQ sources.
Tests: prefill_swiglu_quant_test (new, registered under STRATA_BUILD_TESTS in
the SM60 build, skips with 77 without a device) compares the fused kernel with
swiglu + quantize byte for byte for Q2_0, IQ4_NL and Q8_0 down types, split and
interleaved gate/up rows, a zero row and a row with one huge value.
Run on z440 (RTX A4000, sm_86, CUDA 13.0, SM60 build, gpu lock): all 18 cases print
"0 of N bytes differ" (1, 37 and 4113 rows) and the test ends "prefill_swiglu_quant_test:
ok". sm_60 compile of the tree (CUDA 12.9, P1004): exit 0.
This is a partial port of another contributor's prompt-path fusion change;
only the kernel is taken, as in PR 3752.
Problem: on a GA100 card (70 SMs) the 512-thread fused prompt-expert kernels
leave the card idle at every barrier and at every grid tail.
Change, all opt-in:
- STRATA_PF_FUSED_NT=256 sets up a second set of native_kernel instances with
256 threads and two blocks a multiprocessor (WW 2 only), with two-pass weight
loads (the loads of the next k step are issued before the current step's
decode; load_unit_l gives the same words as load_unit). Every output is the
same thread's sum in the same order.
- STRATA_PF_BATCH=N launches up to N experts at once (WideBatch holds 512 blob
pointers, 4 KB of kernel parameters: CUDA 12.1+, Volta and newer; HIP, gfx906
and the Pascal build keep kMaxBatch). Above 128 it needs the NT = 256 kernels.
Below 128 it only changes the launch grouping.
- With neither variable set the 512-thread kernels, their launches and their
load sequence are unchanged: tile_rows(ww, 512) equals the old 256 / ww, the
two-pass code is behind a constant-false TWO_PASS, and the one added test in
experts_native is occ_nt > 0, which only STRATA_PF_FUSED_NT=256 sets.
- If the 256-thread instances cannot be set up (shared memory, occupancy) the
engine prints one line and keeps the 512-thread kernels.
Cost: the NT = 256 instances are new device code in every CUDA build (the
reporter measured 365,632 bytes for sm_80).
Tests: new ctest entry prefill_fused_iq_test_nt256 runs prefill_fused_iq_test
(the fused kernels against the double-precision reference and the MMQ path)
with STRATA_PF_FUSED_NT=256; it exits 77 below sm_80.
Reporter's numbers (linxuhao, CMP 170HX): one MoE layer of 3,256 tokens
8.27 ms with NT = 256; prompts +7.6 % (4K), +4.9 % (16K), +4.3 % (52K) for
NT=256 with BATCH=512 on the PR's own build. Greedy answers identical on their
11 prompts. These are theirs.
Original change by linxuhao (PR #3779, commit 056d36a9). Merged onto the Q2_0
fused path of #3809 (STRATA_PF_FUSED_Q2N=1) without textual conflicts; the PR
predates it, so one line is mine: the NT = 256 setup also sets up the Q2_0
gate/up instance when Q2N is on (without it that instance would have launched
without its shared-memory attribute).
moe_shared_init and moe_scatter_add came with the layer-major walk and had no test.
lm_kernels_parity runs them on synthetic data of the model's width (N = 2560, no
model, a few MiB) against a host reference:
- moe_shared_init: acc = shared * sigmoid(sg), within 1e-5 relative (the device
sigmoid uses __expf);
- moe_scatter_add: the per-token fma chain in row order, bitwise equal to the host
chain and bitwise equal between two launches (no atomics), over groups of one
row, one sub-chunk, several sub-chunks, one row per token and eight per token.
Registered in CMakeLists.txt next to gr_upmix_parity (CUDA builds, not gfx906),
SKIP_RETURN_CODE 77 when no device is visible.
Not covered: the walk itself needs a model (see the 3976 gate scripts).
Problem: an sm_80 card took the shared interleaved-MMVQ row tables, which
were measured on other cards (#3252 reports A100 slowdowns).
Change: kIlRows_80 (2-4 columns), kIlRows58_80 (5-8 columns) and kIl1Rows_80
(one column), looked up under the key 80 only. They were read off
mmvq_il_parity --bench --emit-table with the same 3 % rule as the other tables.
STRATA_MMVQ_IL_SM80=0 maps an sm_80 card to the key 81, which has no table, so it
gets the shared table again (and no one-column table, as before).
A table only picks the rows a warp (0, 1, 2 or 4); 0 means native_mmvq's kernels.
mmvq_il_parity already checks every rows choice against native_mmvq bitwise for
every type and shape, T = 1..8, so any entry of the new tables is covered.
Reporter's numbers (linxuhao, one CMP 170HX): decode +2.7..+3.2 % over the shared
table, Qwen3.8-Flash-Next IQ3_S. The measured card is not an A100; the rows may
want re-measuring there.
Original change by linxuhao (PR #3779, commit eaedca84); the commit message here
is rewritten. The sm_89 table of #3453 is a separate change in the same list.
Problem: sm_80 (GA100: A100, A30, CMP 170HX) ran the per-group dequant kernels
that sm_61 and sm_86 had already left for the wide ones, because
dequant_wide_on() had no sm_80 entry.
Change: dequant_wide_on() answers yes for compute capability 8.0 as well. The
test is the raw (major, minor) of the device, so it is false on every other
card (sm_60, 70, 75, 86 keep their answer, sm_89, sm_90 and sm_120 stay off as
before). STRATA_DEQUANT_WIDE=0 still turns the wide kernels off.
Reporter's numbers (linxuhao, one CMP 170HX, GA100 sm_80 with 70 SMs): the GPU
time of a 3,259-token prompt went from 857 to 792 ms, prompts +6.3..+7.2 % at
3.3K tokens and +3.9 % at 52K. We have no sm_80 card; these are theirs.
Bit identity: the wide kernels are the ones sm_61 and sm_86 already run by
default. dequant_wide_parity compares them with the per-group kernels byte for
byte for FP16, BF16, FP32, the padded stride and a row slice. The kernels have
no sm_80-only code, so the result does not depend on the card; what only an sm_80
card can show is that ptxas for sm_80 keeps the same arithmetic (see the
checklist in phaseA-tests/gh027/gates/3779-sm80-checklist.md).
Original change by linxuhao (PR #3779, commit 48dc535e); the commit message
here is rewritten.
Problem: a request declares tools; the model writes a complete call for a name that is not among them (bash for shell, a dropped
prefix, an MCP name with every "__" written as "_"). Since #3075 such a call stays text, which is the rule we keep, but the client
saw an ordinary answer with finish_reason stop and nothing else, so the agent turn ended with raw XML and the model never learned
of its mistake. Reporters on OpenCode, Pithagoras, OpenClaw, LibreChat, Open WebUI and Pi hit it.
What changed (declared-only stays: nothing undeclared is ever returned as a call, finish_reason is unchanged):
- serve/frontend.py OutputParser: a declared name that contains "__" is also read when written with every "__" as "_"
(mcp_playwright_browser_navigate for mcp__playwright__browser_navigate); the call goes to the client under the declared name.
Same rules as the other aliases: only when unambiguous, and a tool declared under that spelling wins. A reordered or renamed
tool (server1_label_values for label_values_mcp_server1, bash for shell) is not guessed.
- serve/server.py: when a block went back as text for an undeclared name, the reply carries
"strata_rejected_tool_calls": [{"name": ..., "reason": "undeclared"}] (the names the parser already collected, up to five): at the
top level of a Chat Completions reply and in its last streamed chunk, and in the message_delta event and the message of
/v1/messages. It is absent when nothing was rejected, so every other reply is unchanged. The text itself is untouched.
- The server window line says the response carries the names; the count goes to totals["tool_calls_unread"] and to
/metrics/prometheus as strata:tool_calls_unread_total.
- docs/DETAILS.md: the alias, the marker, the counter.
Tests: serve/test_tool_name_alias.py (collapsed spelling at several chunk widths, a declared spelling and a plain tool unchanged,
a tool declared in the collapsed spelling wins, two tools with one collapsed spelling are not guessed, reordered/renamed names stay
text and are named), serve/test_server.py ForcedToolChoice (the field in Chat Completions whole and streamed and in /v1/messages,
absent for an ordinary answer and a declared call, the collapsed name is a call; the counter). The whole serve/ suite: 1055 tests,
two failures that also fail on the base tree (test_fatal_recovery on this PC, RestartWindow.test_not_alive_until_ready under load).
Not covered: returning an undeclared call as a real tool call (the owner keeps declared-only); a block for a DECLARED name that came
back as text because it was malformed (counted, no name to list); the /v1/messages field with a real client; a stderr or log-file copy
of the warning.
Problem: a decode window's QSA select runs one CTA per query for the top-k
(block_topk_wide_kernel past ~135K cells) and the warp-shuffle block scorer
(block_scores_multi_kernel), in each QSA layer. Both grow linearly with the
context: on one GP100 die at 300K cells a top-k call takes 0.39 ms with 1-5 SMs
busy and a scores call 0.22 ms (two queries). At short contexts the same work is
small and an extra kernel is pure overhead.
Change: qsa_select.cu (inside !__HIPCC__ && STRATA_EXPERIMENTAL_SM60) gets
block_topk_multi_* (six kernels over several CTAs per query, one per 2,048
blocks of capacity, at most 32, exchanging through caller scratch: no clusters,
no spin waits, capturable) and block_scores_dx_kernel (the multi scorer's
per-head sums built in registers as the same 32-leaf xor-butterfly tree, no
__shfl_xor_sync, the tail block stays block_scores_tail_kernel's). New public
qsa_topk_multi_scratch_ints, qsa_block_topk_long and qsa_block_scores_long
forward to qsa_block_topk / qsa_block_scores outside the build gate or their
conditions. verify.cpp keeps a second set of window graphs (exec_[sel_long_][T])
for windows that reach STRATA_SELECT_LONG_AT cells (default 65536); the choice
is made before a capture and handed to the next stage of an overlapped layer
split, capture_all captures both, pl_launch picks by the window's position.
The ids and the score bits are the replaced kernels'. STRATA_TOPK_MULTI=0 and
STRATA_SCORES_DX=0 turn either half off inside the long graphs.
Review changes against the PR: in a build without the multi-CTA kernels
STRATA_SELECT_LONG=1 now says so and is ignored (the PR would have kept two
sets of graphs for nothing there); Metal gets stubs for the three new calls
(the SYCL tree has its own verify.cpp and qsa_select copy).
Default path: unset, select_long_on() is a cached false, topk_scratch_ is not
allocated, sel_long_ stays 0, and every window uses exec_[0][T], the graph it
used before; the index change is the only difference in verify.cpp.
Merge note for f-027-gh-eng (#3702 evict_window_graph, not in this branch's
base): its exec_ / exec_nr_ / launched_ accesses need the same [sel_long_]
first index.
Tests: src/kernels/qsa_select_long_parity.cpp (new; registered with
STRATA_SCORES_DX=1 STRATA_TOPK_MULTI=1 in the SM60 test block, and a second
case with STRATA_TOPK_MULTI_N=3; skips with 77 without the kernels) compares
qsa_block_scores_long with qsa_block_scores (nq 1-8, n_bid 8 .. 131072, tail
blocks of 0-3 cells, every block 0..n_bid memcmp) and qsa_block_topk_long with
qsa_block_topk (nq 1-16, contexts 1,000 .. 524,288, random scores, all-equal
scores and 7 distinct values for ties across CTAs, selected ids of every query;
the multi scratch is junk-filled and must have been written). z440 (RTX A4000,
sm_86, CUDA 13.0, gpu lock, build with -DSTRATA_EXPERIMENTAL_SM60=ON): "172
cases, 0 differ -> PASS" with one CTA per 2,048 blocks and with N=3. sm_60
compile of the whole tree (CUDA 12.9, P1004): exit 0.
Not covered: no GP100 run here (the kernels' last word is the P100, and the
speed claim). The contributor's numbers (four GP100 dies, every expert resident,
--pipeline-windows 3: 400 tokens at 115K context 51.0 -> 52.5 tok/s, +2.9%;
prompts below 64K unchanged; replies and state hashes equal on 15 requests) were
not re-measured. The model gate is phaseA-tests/gh027/gates/3820-select-long.sh.
Everything the Chat view shows (header, tabs, message box, settings, toasts,
tool blocks, the sidebar, the history and Load/Unload texts) now goes through the
translation catalog: serve/web/locales/{de,es,ja,ko,pt,ru,zh}.json beside en.json
(written from the English text in the code by tools/web_i18n.py). Plurals follow
the language (Russian has few and many, Japanese and Korean only one form).
serve/test_web_i18n.py fails when a catalog lacks a key, has a stale one or
changes a {placeholder}. A ninth language is one file plus one line in
serve/web/i18n.js. The Monitor, Images and About views are still English (the
language setting says so); they are the next step.
Browser tests: a language pick translates and is kept, /?lang= is for one visit,
an unknown code stays English, all seven catalogs load without an error.
docs/DETAILS.md: a section "The web page" (every switch, what the preview can and
cannot reach, the languages, the tests), the Sampling drawer's new name, and the
credits (KaTeX and the contributors of the PRs ported into the page).
Five overlapping proposals, one set of controls (Chat settings > Chats, all off
by default; off, the page is the page from before: one chat, New chat clears it
with an Undo, Send is off while an answer is written, no new request field).
- Saved chats list (#1787): the chats kept in this browser (IndexedDB, one record
per chat, names of attachments only), a list with day groups, search over
titles and text, rename, delete with Undo, collapse; a private window or full
store keeps the list in memory and says so once. Nothing goes to the server.
The list has the New chat button, so the composer's one is hidden while it is on.
- Edit and rewind messages (#2615 + #3823): under each message, Regenerate (a new
answer to the same question), Edit answer (its text only: thinking and tool
calls go, no request) and Rewind to here. Rewind asks, as #3823 proposes: "Edit
and send again here" hides this message and the ones after it, puts the
question and its attachments in the box and keeps one Undo bar ("2 messages
hidden"); "Branch from here" starts a new saved chat with the messages before
it and leaves this one alone. A branch keeps a title that says where it came
from ("Branch of ...", also at the very first message) and links back. One
control does what #2615 had as Edit on a question and as Remove from here.
- Send while answering (#4004, for the web page): "Wait" (default, as before),
"Send after" (the message waits and goes when the answer is done) or
"Interrupt" (the answer stops where it is, what is written stays and is marked
interrupted, the message goes at once; a running tool finishes first). One
message waits; Take back returns it to the box; Stop keeps it for you to send.
The pure logic is serve/web/history.js, the store serve/web/chatstore.js, the
page's part serve/web/chats-ui.js. Tests: serve/test_web_history.js (Node, with
the IndexedDB stub and cases of #1787), serve/test_web_history.py, and 17 browser
tests (tools/webtest/browser_test.mjs). Screenshots: tools/webtest/shots.mjs.
Also: the Sampling drawer is now Chat settings (its button says so), a wide
window lets a message use the column (#3176 widened only the column).
Ported by hand and merged: #1787 by Scotty Hudson, #2615 by Joyer Huang (uppet),
issues #3823 and #4004 by xvcezrqpbxtpzblrt (their proposals, for the desktop app,
carried to the web page).
Co-authored-by: Scotty Hudson <[email protected]>
Co-authored-by: Joyer Huang <[email protected]>
Co-authored-by: xvcezrqpbxtpzblrt <[email protected]>
The layer-major walk is a copy of run_impl's attention half with the experts moved
out. It does not carry four things run_impl does, and when one of them is asked for
the walk now declines (the prompt is read chunk-major, and the one-time "layer-major
off for this prompt" line names the reason):
- the DFlash2 drafter's prompt taps (tap_layers_): the walk would leave the taps
stale and the drafter would read old rows;
- STRATA_KVA (approximate late layers): the walk would silently read exactly;
- a stop_at_chunk callback (the pipelined batch reader that yields at chunk
boundaries): the walk reads every token, and stopped_at() must stay -1;
- a pending force_chunk (a --prefill-pipe plan): it is left set for run_impl.
The walk also resets stopped_at_ like run_impl does at the start of a run.
With STRATA_PREFILL_LAYER_MAJOR unset none of this is reached.
Not covered: no model run in this commit.
Only NVIDIA CUDA has run the layer-major walk. On a HIP build (wave32 or the
gfx906 compat layer) lm_prepare now declines with a reason, so a prompt takes the
chunk-major path and the one-time "layer-major off for this prompt" line names it.
The code still compiles everywhere the shared prefill.cpp does; nothing else
changes. With STRATA_PREFILL_LAYER_MAJOR unset nothing is reached.
Not covered: HIP and SYCL were not built or run (SYCL has its own prefill.cpp).
run_impl never clears a later stage's hand_in_, so after one short prompt (read
chunk-major) every later prompt was declined with "a layer split's chunk hand-off
is in use". lm_prepare_chain now declines only when a hand-off is really in flight
(a stage's next_run_ is valid) and otherwise clears the stale hand-in pointers.
A prompt longer than the common super-chunk is split into the same number of
super-chunks of near-equal size (32K at a 28K fit: 2 x 16K, not 28K + 4K). That
keeps a stage's residual rows in VRAM and lets a split's stages overlap on every
super-chunk. STRATA_PREFILL_LM_EVEN=0 gives the old largest-first split. It only
acts when STRATA_PREFILL_LAYER_MAJOR is set.
The split was declined whenever the model has PLE layers. The PLE block (layer 1)
reads the token history, which only the first stage has, and a later stage never
reaches it. So the split is declined only when the first stage does not hold both
layer 0 and layer 1.
Problem: layer-major prefill declined every layer split, so a split model kept
re-reading each layer's streamed experts once per 8192-token chunk.
Changed:
- lm_prepare_chain sizes every stage with the same super-chunk (the smallest any
stage fits, up to four rounds until they agree) and gives them back together;
- a stage runs only its layers [LB, LE); a later stage's rows arrive in its pinned
R_host;
- a stage before the last writes its final rows into the next stage's R_host on
the copy stream, then runs the next stage on a thread (std::async) while it reads
its own next super-chunk. There is one hand-off buffer: a stage's final pass waits
until the next stage has finished the previous super-chunk;
- layer_major_mid() asks the last stage, the one that reports the rows;
- run() drains the pipeline after the walk and clears the mid flag on every stage.
Not covered: only two stages were tried by the reporter (RTX 3060 + RTX 5060 Ti);
the peer-GPU tier and the stage helper have no layer-major form and decline.
run_layer_major called mmq::iota on ids_identity, which is null when no layer is on
MMQ; it now skips that. The QSA shapes come from core::qsa_shapes(g) as run_impl
does, instead of being filled in by hand.
lm_prepare required ids_identity and tok_dev even when no layer is on MMQ. Only a
model with an MMQ layer needs the identity table, so the check now asks for it only
then. Without this a model whose layers all take the FP16 path was declined.
Problem: lm_prepare declined every model whose experts are not all on MMQ. All 48
layers of UD-Q4_K_XL take run_impl's FP16 path, so layer-major never ran on them.
Changed: a layer outside the MMQ plan dequantizes each routed expert to FP16 once
per super-chunk and runs two FP16 GEMMs per entry on that entry's FP16 input rows
(a new per-super-chunk buffer mixed_h, the bits run_impl's gather reads), then
adds the group's rows per token with moe_scatter_add as the MMQ path does. The
buffers each path needs are checked separately, and a missing one gives a decline
reason that names it. The summary line counts the groups on the FP16 path.
Not covered: no model run in this commit.
Problem: the layer-major walk (#1670) was written for an older base. It read the
hyper-connection matrices as bf16 only, accepted expert formats its MMQ pass does
not carry, and called the QSA indexer with the older argument lists.
Changed (Norhiro's port of vizdatom's walk, kept as one step):
- the HC projections are dequantized per half when STRATA_HC_Q8 holds them as Q8_0
with no bf16 copy, and released after the GEMMs that read them are enqueued;
- lm_prepare declines (the prompt then runs chunk-major) for a layer that needs the
WMMA tiles, an APR (Hadamard-folded) layer, and an NVFP4 layer;
- STRATA_PREFILL_DIRECT is not used by the walk.
Not covered: no model run in this commit.
A super-chunk of S tokens is read layer by layer: each 512-token sub-chunk runs the attention half,
the FFN half hyper-connection read, the router and the shared expert in order (GDN / QSA / PLE state
advance as in the default path); then the layer experts run once for the whole super-chunk - each
routed expert staged once, its rows of every sub-chunk in one MMQ group, the results added per token.
Expert traffic drops by about S / 512 (a 12.9K prompt: ~509 GB -> ~79 GB on ryzen).
- STRATA_PREFILL_LM_TOKENS: the super-chunk target (default 8192), shrunk to fit free VRAM minus
STRATA_PREFILL_LM_MARGIN_MIB (default 96).
- STRATA_PREFILL_LM_R=vram|host|auto: the residual rows of the super-chunk in VRAM, or in pinned RAM
with the copies overlapped on the copy stream (auto: VRAM only when the whole target fits there).
- STRATA_LM_PREFETCH=0: no prefetch (default on: every streamed expert of a layer goes to the stager
when its attention halves start; the ones the routing skips are skipped if not started).
- Not for a layer split, a peer GPU, image prompts, the fused layout or a prompt under two chunks.
- on_chunk still reports every sub-chunk in order; layer_major_mid() tells serve not to checkpoint
before a super-chunk end (the state is already at that end).
Port to 0.1.43.1: the copied run_impl body follows the new base. Every blob the stager or the copy stream reads later
comes from blob_stable() (blob() may hand out a recycled STRATA_IO_PREFETCH buffer); the token ids go up through tok_up,
the QSA step records and the PLE rows' pageable fallback through up_ring (#1712, #1996); the PLE gather threads pin to
--aux-cpus; bf16x2 takes m.f16_io; the K8V4 append is upstream's, which writes the host copy and the staging pool by
halves. The indexer passes the INT8 pooled keys as the original does, in the QSA port's order (pooled_i8 and
pooled_scale after block_pos, the int8 qsa_block_scores(_tc) overloads); STRATA_IDX_INT8 stays opt-in and unsupported on
the prompt path, as in the original. stage_one calls the stager's pass() before wait(): with batched claims a routed job
after about a ring's depth of skipped prefetch jobs waited forever. lm_prepare reads the fused layout from m.fused_bufs
and declines while STRATA_DBG_FEAT, STRATA_DBG_ROUTE_OVERRIDE or STRATA_PREFILL_DUMP_R_ALL is set. serve's near-tail
checkpoint (--prompt-cache-tail) is not taken mid-super-chunk, so under layer-major it is taken only when the last
super-chunk is a single sub-chunk.
From vizdatom's #1670, original commit .
moe_shared_init writes the shared expert part of moe_combine before the routed experts are known;
moe_scatter_add adds one expert group rows into a per-token accumulator, per token in a fixed order
(deterministic, no atomics).
From vizdatom's #1670, original commit f0173382d0646907917edf99e59617c1412d898e.
Chunks below stream_all_min() stage each layer's unpinned experts on the stager's few threads, and with a mapped
file tier each of those is a chain of 128 KiB page faults, a shallow queue for the drive. With STRATA_PREFILL_WARM=1
(and a source that warms: the GGUF in place, or experts.bin with STRATA_BIN_WARM=1) a background thread hands the
source the layer's whole list at once before the stager starts, and the next layer's list from the previous chunk
(less what is on the GPU or pinned) while this layer computes. It only warms pages: the stager copies the same
bytes. At most four lists queued; a thread that falls behind drops the oldest. Off by default: the A/B arm.
Port to 0.1.43.1: the warmer thread pins itself with aux_cpus::pin_current_thread() first, like the other helper threads since --aux-cpus (it would otherwise inherit the host thread's CPU).
From vizdatom's #1670, original commit cabe33c4f585d4c9dd0af1fc9f014d00e2920042.
issued is a watermark, so a job nobody waits for could be passed while its thread was still writing
the pinned buffer the next job on that buffer was filling - a corrupt expert blob. A job now also
waits until the job kRing before it has marked itself ready. skip(j) lets a job the routing turned
out not to need be marked ready without reading it (the layer-major prefetch). The default walk
consumes every job in order, so neither changes it.
Port to 0.1.43.1: the stager now claims batched runs and sleeps in issued.wait, so the ready[i - kRing] wait
sits in the run's per-job wait loop after the issued wait, and a skipped job is left out of the run's gather
(no memcpy, prefill_copy or copy_blobs entry); its ready flag and `active` stay with the run's single store and
decrement. New pass(j): skipped jobs are never issued, so the layer-major walk moves the watermark to the
routed job it is about to wait for (store plus notify_all, which wakes sleeping threads); without it a routed
job after about a ring's depth of skipped ones waited forever. The default walks never call it.
From vizdatom's #1670, original commit 149b899548fbc98977219e6fd4ee7866384e61f6.
One button in the header of every tab: Load while the model is unloaded, Unload
otherwise. It calls POST /load and POST /unload (the server already had both,
for its own page only). Unload asks first and never interrupts a turn: the
button is off while an answer is being written, an image is made or a request is
queued. The pill and the Monitor say Unloaded / Loading / Unloading, a message or
an image on an unloaded model shows Loading, and a poll that started before the
POST cannot flip the button back (the POST's state is held until a later poll
agrees). On a phone the pill's text is the first thing the header row gives up
(it measures the row, not a fixed width), so Load/Unload and the theme button
stay on the screen from 320 px.
Browser tests (tools/webtest/browser_test.mjs): unload with its question, load
back, declining, not while answering, a message that loads, the header at 320 to
640 px.
Ported by hand from PR #3550 by Sakurai Momoka (sakuumomo); the texts now go
through the page's translation catalog.
Co-authored-by: Sakurai Momoka <[email protected]>
A complete ```svg or ```html block in an answer gets a Preview button. The
block's own code is drawn in an iframe (srcdoc) that is cut off from the page:
sandbox="" (an opaque origin: no access to the page's DOM, storage, cookies or
server, no scripts, forms, popups, downloads or navigation of the page), a CSP
in the frame's document (default-src 'none': no network, fonts, frames or media;
styles inline, pictures as data: only; no <base>, no form target), no referrer
and no permissions. A well-formed svg is shown as an <img> with a data: URL: an
image runs no script, loads nothing and animates in every browser. The frame
takes the page's light or dark colours and is built again when the theme
changes. A page (html) can also be run WITH scripts, one block at a time, after
a warning that is shown every time: sandbox="allow-scripts" only (still an
opaque origin), the CSP allows the block's inline script and no address; closing
the preview ends it.
What the sandbox flags alone do not stop is a link or <meta refresh> inside the
model's html loading a web page into the frame. The browser test shows it does
(tools/webtest/browser_test.mjs, with and without the header). So the page is
served with Content-Security-Policy: frame-src 'none' (nothing else: a srcdoc
frame inherits the page's policy, and the script preview needs its inline
script), and a <meta http-equiv=refresh> in a previewed page is dropped.
tools/webtest holds the mock server (the page's own files, scripted answers,
no model) and the browser tests; serve/test_web_preview.js checks the documents,
flags and frames, serve/test_web_preview.py the header and that the page never
listens for a frame's postMessage.
Ported by hand from PR #1628 by thoives.
Co-authored-by: Thomas Ivesdal-Tronstad <[email protected]>
A PC whose CPU lacks the AVX-512 set that setup checks for gets the native Q2_0 pack. On a CUDA card of sm_80 or newer its
Q2_0 / Q2_0 prompt experts run through MMQ, although the fused int8 kernel for that format (#2145) is in the binary and set
up only on a Turing card (b6443f3ae).
STRATA_PF_FUSED_Q2N=1 sets that kernel up on any sm_80+ card and takes it for the Q2_0 / Q2_0 layers, under its own variable
so that STRATA_PF_FUSED=1 keeps its bits. STRATA_PF_FUSED=0 and STRATA_PF_FUSED_NATIVE=0 still turn it off. A failed setup
leaves the occupancy as it was and the layers on MMQ.
- moe_fused_iq.cu: DevInfo.q2n; dev_info() sets up Q2_0 gate/up when the switch is set; gu_covered() and native_supported()
take the pair only when q2n is set (a Q2_0 gate with another down type is refused, as on Turing).
- tests/cuda/prefill_fused_iq_test.cpp: --unset-pf-fused leaves STRATA_PF_FUSED unset (the engine's default) and skips below
sm_80; CMakeLists.txt: the ctest entry prefill_fused_iq_test_q2n runs the Q2_0 / Q2_0 pair with the switch alone (exit 77 =
skipped without an sm_80+ card; CUDA branch only).
- docs/DETAILS.md: one section in the author's measured numbers (RTX 5080) and the A4000 caveat below.
Unset (the default), nothing changes: dev_info() reads the variable once and sets nothing up, gu_covered() and
native_supported() return what they returned. Read line by line against base: the Turing path (t75) and the HIP path are
untouched. Set, the prompt arithmetic changes (first-token KL against FP16 0.0110, MMQ 0.0112, the author's figure) and the
fused path's smaller buffers let --prefill auto pick a larger chunk, which is most of the author's 16K gain; the generated
text can change.
Not verified: an RTX A4000 (sm_86). The kernel was taken on every CUDA card in #2145 and then kept to Turing after an A4000
no longer matched the old output and read 4K prompts 4.5% slower; the PR does not repeat that run. Merge note: #3277 and #3779
change dev_info() and native_supported() in the same file.
Tests: prefill_fused_iq_test_q2n and prefill_fused_iq_test pass on an RTX A4000 (sm_86, synthetic data); on the Q2_0 / Q2_0 layer the
fused kernel is 1.76x / 1.30x / 1.24x the MMQ path at 2048 / 4864 / 8192 tokens with the same relative RMS against FP32. Builds on CUDA
sm_86 (Linux) and sm_120 (MSVC). Not covered: no model run; the engine-level A4000 check is the gate.
Chat settings > Render math (off by default). When it is on, \( \), \[ \], $x$
and $$ $$ in an answer are drawn as formulas by KaTeX 0.19.0, which is vendored
in serve/web/vendor/katex (script, style, 20 woff2 fonts, MIT licence and a
PROVENANCE.json with the npm integrity and the sha256 of every file, so the page
needs no network). Nothing of KaTeX loads until an answer holds a formula.
The shared core (math.js, from #2972) keeps the text of every message exactly as
the model wrote it: formulas are swapped for placeholders before the Markdown
runs and drawn over them afterwards, so history, copying, the Markdown export
and the requests never see the drawing. Code, link targets, prices ($5 and $10)
and a formula still being streamed stay text. A formula that fails, is over
16 KiB or times out stays as its source text.
Safety: KaTeX runs with trust off (no \href, \url, \includegraphics, \htmlClass:
they are drawn as their own name), maxExpand 1000, maxSize 20 and a fresh macro
table per formula; it builds DOM nodes itself, so no HTML string from a model
reaches the page. A formula already drawn is placed again at once while an
answer streams, so it does not flicker back to source.
serve/test_web_math.js runs the real vendored KaTeX against hostile formulas;
serve/test_web_math.py checks the vendored files against PROVENANCE.json and
that nothing outside this server is addressed. serve/web/vendor is -text in
.gitattributes so the hashes hold on a Windows checkout.
Ported by hand from PRs #2972 (core) and #2974 (KaTeX) by CC-David-CC.
Co-authored-by: David <[email protected]>
The served page had no way to translate its text: every string was English in
the HTML and the script. Now the English text stays in the code next to a key
(tr("key", "text", vars), data-i18n="key" in the HTML), a language is a catalog
serve/web/locales/<lang>.json, and a catalog that is missing or has no entry for
a key leaves the English text. The language is the user's choice (About >
Settings), kept in this browser; nothing is chosen from the browser's language,
because a page translated in part would mix two languages. /?lang=de shows one
for a visit.
The server hands out /web/locales/<lang>.json and /web/vendor/katex/** from
serve/webfiles.py: whitelisted names and types, no backslash, colon, percent or
dot-dot, and the resolved path must stay inside the folder (also against a link
out of it). tools/web_i18n.py reads the English text out of the code, writes
en.json and checks every catalog for the same keys and {placeholders}; the
check runs once the catalogs exist.
Only the part of PR #3311 by gopinath87607 that is about this engine: a periodic mark is saved at a chunk boundary of the batched prefill and the read carries on
(src/program/generate.cpp, periodic_checkpoint), so the interval is not a piece-size dial; what cuts a read into runs is the re-read, the root, a message
boundary and the start of the last turn; --prompt-cache-every 0 skips the periodic and the near-tail marks. The wall-time numbers and the comparison with the Maya
checkout belong to another engine and are left out. New section in docs/PROMPT_CACHE_TAIL.md, linked from the option's row in docs/DETAILS.md.
--show-key only knew a running launch or a preset. A model set up by setup.py (START-HERE.bat, ./setup.sh) keeps its generated key in strata-<model>.json, so the command
said there was no key and people read the JSON by hand. With no launch and no --preset, it now looks at the run configs of the installed models: one model with a key
prints that key (stdout is the key alone; a hint line with the model and the file goes to stderr), several models ask for --model NAME (any case, the size or the full
id), a name that matches nothing says so. --preset never falls back to a setup config, and no key is printed while the choice is open. Tests:
tests/terminal/test_launch_show_key_setup.py.
A volume or bind mount from the 0.1.x image keeps its models in /data/models. The entry point moves a folder that is part of the volume, but a folder the person
mounted there is left alone, and a failed move leaves the files too; with no MODEL_PATH the start then did not see them and downloaded the model again
(tens of GB) without a word. Now, when .gguf files are still in /data/models and MODEL_PATH is not set, the container stops with code 2 and names the folder,
says what to do (MODEL_PATH to the folder with the shards, or the files into /data/Strata-data/models) and that nothing was downloaded.
ALLOW_DOWNLOAD=1 goes on anyway with the same text as a warning. docs/DOCKER.md lists the setting. The other points of the issue (the key, the removed
variables, the status loop) are the owner's design questions and are not changed here. Tested by running the guard function with sh (tests/terminal/test_docker_files.py);
the whole container was not run (no docker on this PC).
Prefill on one card, decode on another, the session handed over in RAM through the engine's
own session file (SAVE/RESTORE). The engine's new request form GEN 0 prefill=1 reads a prompt
as a request does, up to where the first verify window starts, and stops with the session at
n - 1 tokens, which is the place a continuing request resumes from. Opt-in; without
split_serve nothing changes.
Idea and manual setup by @MarcoTool in #4000.
Problem: the strata-dsv41 packages were built for sm_86 and sm_120 (+ PTX 120). An RTX 4090 user was stopped at setup on 0.2.4.1; since
b750d757f an sm_86 package is accepted on an sm_89 card (cubin same-major rule), but the engine had no sm_89 image of its own, and the docs said
nothing about RTX 40 or about building the engine for a card the packages do not cover. The reporter built strata-dsv41 for 86-real;89-real
themselves and ran the 2.00 set at 40.2 tok/s on an RTX 4090 48 GB (theirs, --dspark on).
What changed:
- packaging/build_dsv41_cuda.sh and build_dsv41_windows.bat: the default architectures are now 86-real;89-real;120, so the DeepSeek CUDA 13
packages list sm_86, sm_89, sm_120, compute_120 (Windows and Linux alike). The CUDA 12.9 line in the script header says ARCH="86-real;89".
- CMakeLists.txt: a build of strata-dsv41 with no -DCMAKE_CUDA_ARCHITECTURES defaults to 86-real;89-real;120 (the Qwen engine's default
stays 120, and the warning now names the architectures it picked).
- strata_control/engine_component.py arch_native(): true when a package holds machine code for the card itself. resolve() and
feed.select_engine() use it as the last tie-breaker, so with otherwise equal packages the one with the sm_89 image is taken on an RTX 40;
an sm_86-only package still fits (no refusal).
- setup.py: the "engine not available" stop now points at the build recipe (docs/DEEPSEEK41.md); the developer line names the
architecture to pass. The card list already said RTX 30, 40 and 50.
- docs/DEEPSEEK41.md, docs/MODELS.md: RTX 40 coverage, the reporter's measurement quoted as theirs, no claim of our own (we have no Ada card),
and a section "Building the engine for your card".
Tests: tests/terminal/test_engine_manifest.py (arch_native; the build scripts and the CMake default name sm_89 and both OSes agree; the
package list they give is sm_86 sm_89 sm_120 compute_120; an sm_89 card is accepted and prefers the package with sm_89). tools/test_setup_dsv41.py
unchanged and passing.
Not covered: no run on an Ada card (gate script in phaseA-tests/gh027/gates/3935-ada-dsv41.sh).
One bullet from PR #3930 by NorwinYu: on two identical RTX 2080 Ti over a dual-socket Xeon E5-2696 v4 a plain split moved decode from 35.2 to 51.0 tok/s and an
11.6K prompt from 672 to 908 tok/s, because the CPU pool per window fell from 9.52 to 0.79 ms (STRATA_DECODE_TIMING=1). The page's own 5080 + 3090
figures, where the second card is on par, came from a fast CPU; the bullet says so and that it is one measurement. No code changes: setup already splits
across two cards. Shipped without a hardware run of ours; the reply asks owners of two older cards on slow CPUs to compare and report.
The chat, About and Monitor columns grow with the window (the chat and the
composer to 1100 px, About to 1100 px, Monitor to 1920 px), the metric cards
wrap with auto-fit, and under 640 / 560 / 480 px the controls get bigger
touch targets, the composer bar wraps, and long words, code and tables stay
inside the viewport.
Ported by hand from PR #3176 by Matthias Nott (mnott); the CSS is his.
Two paragraphs from PR #3895 by Vop3p, put under the table row they explain: the fork pays only when it has a CPU-pool wait to hide (4,859 slots on one
RX 6900 XT, GPU-reach wait 20.7 of 39.1 ms; with every expert resident it is 43.4 against 62.9 tok/s on an R9700), and the Windows RX 6900 XT fault with
the fork plus --kv-resident, its bisection and its limits (one PC, kernel not isolated). The text now says what the code does about it and asks RX 6000 owners on Windows to test.
On one Windows RX 6900 XT STRATA_SH_STREAM=1 together with --kv-resident failed the engine on its first request (verify: unspecified launch failure).
setup --calibrate (tools/calibrate.py) no longer measures the fork on Windows when KV streaming is on (the fork is off by default on AMD, so
nothing changes unless the person sets it), strata tune's never-pick list holds the combination (strata_control/tune_knobs.py), and a config that
sets STRATA_SH_STREAM=1 with --kv-resident on Windows HIP gets one warning line when the engine starts (serve/server.py sh_stream_warning). Nothing is refused or
removed: the variable still reaches the engine. Shipped without a hardware run, from one report; the reply to #3895 asks RX 6000 owners on Windows to test
and report. Tests: tests/terminal/test_sh_stream_windows.py.
The 0.2.5 rewrite dropped about 25 entries: the Mac quarantine steps, strata doctor, sending logs, the first-start freeze, the graphics memory the app keeps free,
Smart App Control, UPDATE.bat and git pull (#1276), the build tools and the gcc/glibc compile problems, the engine checksum, the engine stopped unexpectedly,
slower than the tables, the empty-reply and Turing switches, --expert-cache and --resident-budget-gib, the page-locked OOM, the NVIDIA 0x141 reset, CUDA 13.2 on
RTX 50, the AMD section (KFD, the build list, 0xC0000005, the driver reset, pinned allocations, verify timeouts) and the body of the verify-timeout entry, which
the rewrite had left empty. Each is back with the 0.2.7 facts: the app's reserve defaults (native-resources.mjs), the Drafting setting, --vision on|off|gpu|cpu
in the Terminal and --vision no in a source checkout, Intel Arc's CPU reader, and the Terminal route next to the source-checkout one.
Every pointer resolves: setup.py's driver-reset pointer, the server's Smart App Control message, DETAILS.md's STRATA_KEEP_EMPTY_TURNS / STRATA_TOPK_STREAM line,
OLDER_GPUS.md and MAC.md. tests/terminal/test_troubleshooting_page.py checks the entries, those pointers and the page's own links and anchors.
Problem: below sm_70 the prompt reads each QSA layer's attention through the
decode kernels (qsa_prompt_attn_batch returns false without tensor cores): 32
queries a call, every chunk's partial results (817 KB a query) written to
scratch and merged by a second kernel.
Change: src/kernels/cuda/qsa_pattn.cu keeps the contributor's default variant
only, ps_score_x_kernel (q in registers, the chunk's K staged with 16-byte
loads, two cells at a time through a select-free reduce-scatter) and
ps_value_d4_kernel (four dims per thread, merged into the output as the merge
kernel does). Every float operation of attn_chunk_kernel_pre75 and
attn_merge_kernel is kept in its order, so the output is the decode kernels'
bit for bit. The PR's other variants (per-chunk, fused persistent, bench split
codes), pattn_bench and the variant argument are not ported; pattn_bench needs
a STRATA_QSA_DUMP file from a model run.
Gating: compiled only in the CUDA build with STRATA_EXPERIMENTAL_SM60 (HIP and
Metal link a stub that answers false; prefill.cpp keeps the call out of the
Metal build). At run time only with STRATA_PATTN_SM60=1, on compute capability
6.x, int8 K/V, head_dim 256, 12 query heads per KV head, at most 33 chunks.
Unset, the only cost on the default path is one cached int test per QSA layer,
and only when qsa_prompt_attn_batch has already declined. Any refusal or a
failed launch returns false and the usual 32-query loop runs (one warning, no
exit; the PR called exit(1) on a launch failure).
Tests: src/kernels/qsa_pattn_parity.cpp (new, registered in CMakeLists.txt in
the experimental sm_60 test block, skips with 77 when the kernels are absent)
compares the kernels with the Daily's chunk kernel plus attn_merge_kernel on
random int8 K/V: a shuffled page table with ~3% unresident pages and a run of
64 unresident pages (whole chunks masked, some queries empty), three selection
shapes, caps 64 / 130 / 2051 / 2112, widths 0..cap, n_q up to 40 with a
scratch that holds only 2 decode-layout queries (several batches), junk in the
scratch and output. On a Pascal card it also requires the checked entry to take
the shapes and qsa_decode_attn_batch to match. The kernel entry that skips the
compute-capability test (qsa_pattn_batch_any_arch) exists for this test only.
Results: z440 (RTX A4000, sm_86, CUDA 13.0, build with
-DSTRATA_EXPERIMENTAL_SM60=ON, gpu lock): "qsa_pattn_parity: 468 cases, 0
differ, 29724792 nonzero output values in all -> PASS". sm_60 compile of the
whole tree (CUDA 12.9, P1004): exit 0.
Not covered: no P100 run here (the Pascal-only checks of the test and the
speed claims wait for the gate, phaseA-tests/gh027/gates/3943-pattn.sh). The
contributor's numbers (43.3 -> 30.8 ms a 1,536-query chunk per QSA layer at
26K context, 2-3% less time on 8K-150K token prompts on four dies) were not
re-measured; compute capability 6.1 is covered by the same switch and was not
measured by anyone.
The Speed group gets a row for Qwen on a discrete CUDA card of 13 GiB or less: "Prefers the parts of the model already on the graphics card, which makes answers
much faster on 12 GB cards. Turn off if answers look worse than before." On (recommended) by default; Off writes STRATA_ROUTE_PRIOR=0, On removes it. The
DeepSeek speed-ups row, when Off, writes DS4_DEFAULTS=026 and the six DS4_*=0; On removes all of them, and a file that holds only the six (an earlier test
build) shows as Off. The Qwen "0.2.6 behaviour" switch is Terminal-only and is not in the app. New strings fx82 and fx83 in the eight fix catalogs.
- DeepSeek speed-ups Off now writes DS4_DEFAULTS=026 (the engine's master switch back to the 0.2.6 defaults) plus the six DS4_*=0 for older engines; On removes
all seven. A file with only the six (an earlier test build) still reads as Off and On removes them (`legacy` entries).
- New switch "Faster answers on 12 GB NVIDIA cards" (the residency-biased routing default): on by default, Off writes STRATA_ROUTE_PRIOR=0. Applies to Qwen
Flash-Next (no arch) on a CUDA build with a discrete NVIDIA card of 13 GiB or less (not a chip that shares system memory), the engine's own rule.
- New Terminal-only switch "Qwen: 0.2.6 behaviour" (STRATA_DEFAULTS=026, off by default), a row of the Advanced table for every Qwen model, for
troubleshooting; the app never shows it. DeepSeek has DS4_DEFAULTS instead.
tests/terminal/test_speed_switches.py covers the new defaults, applicability, the seven DeepSeek entries and the legacy six, the Advanced rows and the
comparison with the app's table (the Terminal-only switch is left out of it).
/v1/chat/completions on the CUDA and HIP engine accepted logprobs and
returned an empty choices[0].logprobs.content. Now "logprobs": true with
"top_logprobs" 0 to 20 (and the legacy "logprobs": N) returns one entry
per token with its alternatives, streamed or not, and "allowed_token_ids"
returns a distribution renormalised over exactly the ids the caller named.
Engine (src/program/generate.cpp): the GEN line takes logprobs=K and
allowed=a,b,c. After each "T id" the engine writes one "LP <logprob>
[id:logprob ...]" line from the verify window's head logits before
sampling, penalties and temperature. The SYCL engine already wrote LP
lines for logprobs=K; it now reads allowed= as well, with the same line
format. A request without logprobs reaches none of this: both call sites
are behind "req_logprobs >= 0", a local int of the request, so the default
path prints what it printed before and does no extra work per token.
Changes from the PR: the call sites are guarded instead of calling the
function for every token, the function is static, the token index is
checked against the row, and the Chinese comment is in English.
Server (serve/server.py): LP lines are read before the token is yielded
and belong to the request that asked (tagged with its cancel Event), the
entries spell the reply, json_schema replies return no logprobs rather
than a mismatching set, an allowed id past the vocabulary is a 400, and
logprobs on a --parallel server is answered for max_tokens 1 only (a
slot's BT lines carry no LP line). serve/structured.py: _extract_json is
now extract_json (the logprobs alignment uses it); the two uses added by
the #2893 retry follow the rename.
Left out: the PR's bench and result files (none in the diff). #3332 (/v1/completions) and #4056
(logprobs on it) build on this change; this one needs nothing from them.
Tests: serve.test_logprobs and serve.test_logprobs_parallel (30 tests with
the PR's), serve.test_structured, serve.test_responses: 94 tests OK.
Not covered: no model ran here. The PR author reports 28 of 28 on an RTX
5080 with MTP and on a GTX 1080 Ti; the layer-split pipelined decode path
(the first call site) was not run by anyone, and the LP values were not
compared with a CPU softmax of dumped logits.
Co-authored-by: Leonardo Murillo <[email protected]>
Review of the first port: it registered kIlRows_89 / kIl1Rows_89 for every
compute capability 8.9 card by default. Bit-identity of an interleaved-kernel
choice to native_mmvq is a property of the compiled code, and the release
engine builds its own sm_89 cubin that no card of ours can run; the issue also
saw no end-to-end decode change on a 4090 or a 4080, and the 4070 and 4080
runs agree in 53 of the 75 cells of the 2-4 column table (33 of them non-zero). So the default
change has no measured benefit and an unproven bitwise claim on the card it
changes.
Change: sm_89 answers the shared table, with no one-column row, exactly as
before, unless STRATA_MMVQ_IL_89=1 is set (read once, like the neighbouring
switches). With it set the cells from the two community runs apply. Nothing
else in the lookup changes; sm_86, sm_120 and the others are untouched.
Why it is bitwise: native_mmvq_il_kernel keeps a separate accumulator set per
row (acc[v][i][c]) and sums each row over the same chunk order and the same
warp_sum, so the rows-a-warp count R only changes how many rows a warp walks,
not the order of any row's float additions; the plain-versus-interleaved
equality is what mmvq_il_parity measures (memcmp against native_mmvq for rows
1/2/4, T 1-8, odd row counts). That test passes on sm_86; it has not run on an
Ada card.
Tests: mmvq_il_parity now checks the unset behaviour (sm_89 equals the shared
table, Q3_K and the one-column row answer 0) or, with STRATA_MMVQ_IL_89=1, the
4070/4080 agreement. A second ctest, mmvq_il_parity_89table, runs the table
check with the switch set and no kernel case. On z440 (2x RTX A4000, sm_86,
CUDA 13.0, gpu lock): mmvq_il_parity and mmvq_il_parity_89table both pass, and
the binary prints "mmvq_il_parity: OK (every case, T 1-8, rows 1/2/4 and the
table, bitwise)" with the switch unset and set.
Model gate owed on an sm_89 card: phaseA-tests/gh027/gates/3453-sm89-il.sh
(mmvq_il_parity on the card, then identity of the answers with the switch on,
then a speed A/B that decides whether the default flips).
Co-authored-by: rodi4eg <[email protected]>
With --batch-groups above 1 on a layer split the pipelined batch loop never starts a round of the adaptive VRAM tier (the
serial batch windows start one every --adapt-every windows). While the slots are busy the tier only moves when they go back
to the solo path, so a fresh engine keeps its start-up profile for minutes.
STRATA_PIPE_ADAPT_EVERY=N (opt-in; unset, 0 or not a number is off): after N group windows pump starts no new group on the
first stage, lets the groups in flight finish, then runs one round as batch_step does between serial windows, adapt() and
apply_pending(true), so nothing in flight reads a cache slot while experts swap and the new experts have landed before the
next launch. Not while a prompt is read beside the slots (in_prompt_read, its loan of cache slots), not with --adapt-async,
not while the last round's copies are in flight, not with the tier off. The rule is two pure functions in
include/strata/program/batch_groups.hpp (pipe_adapt_every, pipe_adapt_due) with 12 new cases in batch_groups_test. With the
switch set the engine says so once at start and the "strata batch (pipelined ...)" line adds the rounds, the experts
swapped in and the time spent. docs/BATCHING.md documents the switch with the author's table (2x RTX 5090, six streams,
309 to 410 rows/s over eight bursts after a fresh start, two engines per setting).
Reviewed: unset costs one counter increment per group window and one false check where a group starts, and prints no new
line. A due round cannot wait forever: no new group starts, the groups in flight finish, then the round runs. A prompt read
sets in_prompt_read before the reader thread starts (including STRATA_PIPED_CONCURRENT) and clears it after the reader
joined; the drain after the read runs the round if it is due.
Tests: batch_groups_test (the new cases). Not covered and still the author's to show (the PR is a draft): a soak with
prompts read beside the slots, a Linux run, another GPU pair, more than two stages, and exactness settings with the switch
set. The model gate is written down.
STRATA_ENGINE_BACKEND=hip-rdna2 worked for strata update (updater.detect_hardware), but strata update --backend hip-rdna2 offered no package and
strata launch stopped with "STRATA_ENGINE_BACKEND='hip-rdna2' is not one of cpu, cuda, hip, metal, sycl". engine_component.split_backend reads
hip-<variant> as the hip family and its package; detect_backend (the start path), detect_hardware and update --backend all use it. Any other name is
checked as before. Tests in tests/terminal/test_gfx_selection.py.
STRATA_RAM_ADAPT (opt-in, --serve, one GPU, needs --adapt-every) lets the resident RAM copy follow the conversation:
experts the windows keep routing to take the RAM places of unused experts of their layer. Each move re-read the incoming
expert from the file and a layer moved one pair per window, although the incoming expert was usually read for the window
before and its blob still sits in that read's stage buffer.
What changed:
- FileExpertSource::replace_resident_staged(layer, in, out): the same move as replace_resident, with the bytes copied out
of the finished stage buffer of `in` (no file read). The copy runs under stage_mu_, so claim_stage cannot give the buffer
to another blob meanwhile; it needs a non-busy buffer whose key is `in`, and returns false (nothing changed) otherwise,
and the pair then goes the old way. The Foresight hold/unhold around the copy is the one install_resident uses.
- generate.cpp: a layer pairs up to 4 of its most routed outside experts with up to 4 of its least routed ones in the copy
(outgoing count below 0.5, incoming 2 or more), largest difference first. A pair whose incoming expert is staged moves by
copy; `n` now caps only the pairs that need a file read. The log line counts the moves and how many came out of a stage
buffer (ram_staged()).
- docs/DETAILS.md: the RAM_ADAPT bullet in the low-RAM list in plain words and one row of the measured table (the author's
numbers: RTX 3060, IQ2_XS and IQ3_XXS, 120 ShareGPT requests, 3 interleaved runs each).
Unset (the default) nothing runs: the block is guarded by ram_adapt_n > 0 as before. The weights moved are byte copies of the
same blob; which experts sit in RAM changes, as it already does with RAM_ADAPT, so greedy token ids may differ from run to
run once it is on. Changed from the PR: the DETAILS.md hunk was rewritten against the current, shorter bullet.
Tests: file_expert_source_test (new staged cases; the positive case needs unbuffered load and is skipped without it).
Not covered: no model run. Model gate written down: RTX 3060 ShareGPT replay =8 against =1.
The Build section said nothing about it. gemm.cu (create_hipblaslt_state, own_tuning_dir) takes STRATA_HIPBLASLT_TUNING, else the file
<engine folder>/tuning/<arch>-hipblaslt-<version>.txt, else plain hipBLAS. The report measured 826 against 1,043 tok/s at 8K prompts on a
gfx1151 (one run per arm, stated as such) and the log line that tells whether a table was taken. The extra log line the issue suggests for the
no-table case is a small engine change and is not part of this commit.
The first #3218 commit made setup.py admit a Maxwell (sm_52) card only
when a tester set STRATA_EXPERIMENTAL_SM52=1, and then routed it to the
CUDA 12 engine. That engine holds no sm_52 code, and users never set
environment variables to get a fix.
sm60_card() is back to Pascal and Volta (6.0 to 7.0). A PC whose only
NVIDIA cards are Maxwell now gets a warning (maxwell_note) that says no
ready-made engine holds code for it, gives the source-build recipe
(-DSTRATA_EXPERIMENTAL_SM60=ON -DCMAKE_CUDA_ARCHITECTURES=52, CUDA 12.x)
and the one reporter's result, and points to docs/OLDER_GPUS.md. The card
table line says Maxwell runs only from a source build. A Maxwell card
beside a newer card prints nothing new. Naming the card (--gpu, --cuda
12) or any env var does not admit it. The CMake floor from the first
commit is unchanged (opt-in build only).
Tests: tools.test_setup_older_gpus 23 tests OK (a Maxwell-only PC through
setup.main() with and without the old switches installs nothing and
prints the recipe; a Maxwell card beside an RTX prints no note).
Not covered: no engine was built for sm_52 by this commit; the P1004
compile is recorded in the notes.
The report written by BENCH.bat and bench.sh had fields missing or wrong: nvidia-smi's 'CUDA UMD Version:' (current drivers) was not read, the value
was the newest CUDA the driver supports but was labelled like the engine's toolkit, storage always said 'not measured', and Windows 11 showed as
Windows 10. Now the report says 'CUDA (driver supports up to): X; engine built for CUDA Y' (BUILD.json beside the engine), names the storage device
(Get-PhysicalDisk, system_profiler, lsblk -e 7 to skip snap loop devices), adds a 'GPU detail (after the run, near idle)' line with power limit and draw,
PCIe generation and width, pstate and temperature, and recognises Windows 11. The standard report (strata bench run, strata_cli/bench_facts.py) already
read the UMD version, storage and power limit; this brings the legacy tool to the same. Ported from PR #3250 by JoShMiQueL with its tests.
Two 0.1.44.1 changes no page told. TROUBLESHOOTING.md gets the start error 'the engine process ended: ...' (serve/server.py start_exit_note: the signal or
Windows exit code, and for an illegal instruction the AVX, AVX2, FMA and F16C flags the CPU lacks). SOURCE_INSTALL.md says a setup run again
pre-selects the installed size, marked '(installed)', and --yes takes it (setup.py size_tag, installed_size). Written by hand from PR #3243 by
Buddychek, whose hunks no longer fit the rewritten pages; each sentence checked against the two functions.
Review fix on the #4005 port. STRATA_PREFILL_COPY_WAIT_DEDUPE=1 skipped the copy stream's wait on a slot's release
event when the previous entry had waited on the same event. Two cases break that:
- At idx == ring the previous entry (ring - 1) is still in the first round, so its wait was on the event of an earlier
walk. The same slot id can be this walk's first flush, which re-records the event after that wait; skipping the wait for
slot 0 then lets the next copy overwrite a slot that the flush's gather has not finished reading.
- With a ring of one MMQ group (16, the smallest ring) a flush of the whole ring re-records the event right after the
previous slot's wait.
The rule is now copy_wait_dedupe_ok() in include/strata/prefill/copy_events.hpp: idx > ring, ring > group, same event.
copy_events_test models two to three walks with random group sizes and random issuer/consumer interleaving and checks that
a refill always waits on a record at or after the flush that released its slot. The same model with the PR's original rule
(idx >= ring) finds the stale-event case, so the test can fail. The default path is unchanged (dedupe off).
Tests: copy_events_test passes (z440, g++ -std=c++20). Not covered: a GPU run.
Review fix on the #2526 port. At an admission from a slot the port copies every checkpoint of the slot that is not
longer than the copied-back length into the main session's chain. Such a checkpoint is valid there only when its tokens
are a prefix of what was copied back (its K/V cells are then the ones the main session holds) and it has no pictures
(the main session's pictures are cleared at a give-back). The PR did not check either; a checkpoint left over from a
branch the slot no longer holds would sit in the main chain with cells that are not its own.
Both trees (CUDA/HIP and SYCL) now test the prefix and the empty picture list per checkpoint. The checkpoint the
request continues from is copied as before. Nothing changes when every checkpoint is a prefix, which is the normal case.
Tests: compiles on CUDA and SYCL (see the build lines of the branch). Not covered: no model run; the gate is written down.
docs/INTEL.md and docs/INTEL_ARC.md described an SSD expert tier that strata-sycl.sh's default (STRATA_VERIFY_NO_HOST=1) does not have: every
expert must be in VRAM or in the pinned RAM mirror, and the engine refuses to start otherwise (sycl/src/program/generate.cpp, 'REFUSED: ... are
neither in VRAM nor mirrored'). A native (IQ) pack refuses --spec 0 and 1 (sycl/src/program/generate.cpp and src/program/generate.cpp). Ported by hand
from PR #2153 by nekomario28 onto the rewritten pages, each sentence checked against those two messages. The hunk for docs/TROUBLESHOOTING.md is left
out: the entry it changed is no longer on that page.
Adds a minimal handoff configuration (--prompt-cache 1, --prompt-cache-every 0, --prompt-cache-root 0, --conversation-cache-mib 0), what each
option does, that it costs RAM, that it does not make cold prompts prefill in parallel, and how to look for the handoff in the engine log, /metrics
and /v1/status. Ported from PR #3334 by oscarlius (continuing #3035). The four options exist in src/program/generate.cpp and the o.prompt_cache > 0
gate matches the text; the log evidence was not reproduced here (no model run).
metal/src/mtl_rt.mm names MTLGPUFamilyMetal4 (macOS 26 SDK) and MTLCompileOptions.mathMode (macOS 15 SDK) behind @available, a run-time check;
the names must still exist in the SDK the engine is compiled against, so the metal-link job failed on the macOS 14 SDK on every pull request that
touched src/, include/ or metal/. The job now runs on macos-26 (the runner-images repository lists a macos-26 image). Ported from PR #3870
by stuchapin909. A GitHub Actions run cannot be made here: the author ran this workflow on his fork; the first run after the merge proves it.
#3824 changes the same line with macos-15 and an Xcode 26.0.1 path pin, and is not taken.
make_blob quantized its random test rows with ggml-cpu's type traits from_float, which the i-quants do not have: an IQ3_S pair called a null
pointer (an access violation on Windows, SIGSEGV on Linux) before the first timing. It now uses ggml_quantize_chunk, with a flat importance
matrix where the type needs one. A Q2_0 gate/up or down is timed as pool.cpp runs it (activations as an ActQ, Strata's own Q2_0 rows, SwiGLU
in between) instead of with ggml-cpu's Q2_0 dot, and the pair's header line says so. Other formats are timed as before. Ported from PR #3871 by
Licenser.
Built on this PC (MSVC, CPU only) and run: iq3_s/q2_0 0.531, iq2_xs/q2_0 0.336 and q2_0/q2_0 0.214 ms per expert per thread (--mb 512, 1 thread).
Checkpoint.raw mapped every shard for the life of the process. On Windows each faulted page of the ~64 GB of expert bytes is charged to the
process working set, so the pack step looked like a 64 GB RAM demand and was killed by free-RAM watchers on a PC with a model loaded (nothing
leaked: it all came back at once). raw() now reads the tensor's byte range into a fresh buffer through one cached file handle per shard; the
pages go to the system file cache, which counts as available. The author measured the same bytes, 1.2x faster, flat working set, and a
byte-identical pack. raw() is no longer thread-safe (this file uses no threads). Ported from PR #3378 by jerry78424 (the fix for his issue #3377)
with its two tests; I ran the tests (13 pass); the full 102 GB pack was not run.
A GGUF header with an unknown metadata type or an impossible length raised KeyError or OverflowError inside the reader, and setup, which expects
a ValueError, printed a traceback instead of its diagnostic naming the broken shard. Strings, scalars, arrays and tensor dimensions now share
one bounded read; counts are checked against the bytes left in the file before anything is constructed; type ids are checked before their values
are decoded; nested arrays are reported as unsupported; general.alignment must be a positive multiple of 8. The file name is added once to the
message. Ported from PR #2288 by KeyanHu-git with its tests (tools/test_gguf_reader_malformed.py).
Checked against the old reader on the 18 GGUF v3 vocabulary files of the pinned llama.cpp: same metadata, tensors and data start.
When one assistant message made several tool calls and the client returned the results in another order, the call ids were dropped before the
prompt was rendered, so the model saw anonymous results in the wrong order. The ids are now kept while the request is normalized; a run of
results is reordered to the call order only when the call ids and result ids are non-empty unique strings with matching sets. Anything else
(missing, duplicate, unknown or incomplete ids) keeps the received order, and the ids never reach the template. Ported from PR #3395 by
j-luwierski with its seven tests (serve/test_chat_tool_results.py); the hunk was merged by hand beside require_messages.
Problem: an Ada card took the shared MMVQ-IL table, which has no Q3_K row, so
Q3_K shapes ran the plain kernel and the other types used a table measured on
sm_86 and sm_120.
Change: kIlRows_89 (2-4 columns) and kIl1Rows_89 (one column) are registered
for compute capability 8.9. The 2-4 column table keeps only the cells where the
RTX 4070 run (rodi4eg) and the RTX 4080 run (kzhv) gave the same rows count and
sets the rest to 0, the plain kernel; the 4070 and 4080 disagree in 13 of 15
cells of the unmerged tables. The one-column row is Q3_K only, which both cards
gave. The issue's 5-8 column table is not taken: the 4080's changed in 10-14 of
20 entries between two runs of the same card, so there is no kIlRows58_89. The
bench output files of the PR are not ported.
Every choice in a table is bitwise native_mmvq's output (kernel row split only),
so answers do not change; this is speed only. No end-to-end decode change was
seen on a 4090 or a 4080 in the issue, so the gain is the microbenchmark's.
This does change which kernel an sm_89 card runs by default, so the model gate
(identity on sm_89) is written down in phaseA-tests/gh027/gates/3453-sm89-il.sh.
Tests: mmvq_il_parity (registered in CMakeLists.txt) now checks sm_89's table
against the two source runs (check_tables) instead of expecting the shared
table, and still checks bitwise parity of every case and rows choice on the GPU
it runs on (sm_86 on z440). docs/MMVQ_IL_TABLE.md updated.
Not covered: no sm_89 card here, so the table speed is the contributors'
numbers, not ours.
Co-authored-by: kzhv <[email protected]>
When the elastic image encoder failed to restart, its exception left request preparation and the client got a closed connection. It now goes
through the EngineDied path: HTTP 503 with the reason, a later picture tries again, text requests and a later good start are unaffected. An idle
encoder stop is logged even when the cache shrink was refused, and a layer split gets a line saying its encoder VRAM has to be kept free with
--vram-reserve-mib. Ported from PR #2435 by i-Vladimir-i with its tests (chat, messages and responses, streamed and not, plus a real encoder
that exits with code 71 on Linux) and the docs/VRAM_ELASTIC.md text.
On a layer split the later card is left with a few dozen MiB after loading, and each single-request window size is
captured the first time it is used (68, 64, 58, 52 and then 10 MiB free in the report); when one capture's
cudaGraphInstantiate ran out of memory, instantiate_evicting could only free graphs of batch slot layouts, a
single-request run has none, and "verify: instantiate: out of memory" ended the engine.
Now, after the batch layouts, a single-request window's capture frees the graph of the largest other window size (never the
one being captured or a prelaunched one; pick_window_victim) and instantiates again, and says so once: the freed size is
captured again when it is next used. capture_all keeps every size at once and is excluded. Only a failed instantiate does
anything new; with enough VRAM nothing changes.
window_evict_test (CPU, header-only) covers the choice. The real gate needs a model on a 2-GPU split with a thin later card
(listed in the triage file); not changed here: the later-stage reserve of the split search (kWindowMib), which is a default
placement change.
Under WDDM every cudaEventRecord and cross-stream wait is a submission of its own (the author measured ~80 / ~45 us each
on an RX 6900 XT under Windows 11), and the streamed walk paid one record and one wait per expert although a group
gathers after one wait on its last copy. Two tester switches, threaded issuer only:
- STRATA_PREFILL_COPY_EVENT_EVERY=N (2..64): the copied event is recorded every N-th entry, before the issuer waits and on
the last; the compute stream waits on the first recorded entry at or after the one it needs (the copy stream is in order).
- STRATA_PREFILL_COPY_WAIT_DEDUPE=1: the copy stream skips its wait on a slot's release event when the previous slot was
released by the same event.
The same copies into the same slots: the same bytes.
Default path (both unset): read line by line. The issuer records copied[slot] and waits on used[used_of[slot]] exactly as
before (the counters are bookkeeping, cover() is a no-op). Added to the PR so this holds by construction: wait_copied takes
the slot the old code used and, without the sparse switch, waits on copied[slot] at all three consumer sites (the PR derived
the slot from the entry number). The record/wait selection is now three pure functions in
include/strata/prefill/copy_events.hpp, with a host test (copy_events_test, registered in CMakeLists.txt) that models the
issuer and consumer with a full ring and checks that every wait names a live recorded entry at or after the one needed.
Tests: copy_events_test passes (z440, g++ -std=c++20). Not yet: a CUDA build of prefill.cpp in this tree (queued), no GPU
run. Not covered: the Windows HIP / WDDM speed claim and NVIDIA (the gate is written down). Needs the author's numbers
repeated before any default talk.
--host :: failed before the model loaded because Server always opened an IPv4 socket. Server now selects AF_INET6 when the bind address contains a
colon; the ready line, the 'from other devices' lines and the browser address bracket an IPv6 address. The bind is still the person's choice and
no default changes. Ported from PR #3212 by purrsec with its tests (serve/test_security.py, including a real :: bind). setup.py's --host help and
docs/DETAILS.md say :: is all IPv6 interfaces; the README hunks of the PR (they name strata launch --host, not checked) are left out.
A reporter built Strata for two Quadro M6000 (sm_52) by lowering two
floors and saw about 22 tokens a second with Qwen3.8-Flash-Next IQ1.
CMakeLists.txt: the experimental build (-DSTRATA_EXPERIMENTAL_SM60=ON)
now accepts sm_52 and prints a warning for it (Maxwell, one reporter, no
native half precision, CUDA 12.x). Without the flag the build still
stops at sm_75 as before, and the architectures a build uses are not
changed.
setup.py: sm60_card() admits a 5.2 card only when the tester sets
STRATA_EXPERIMENTAL_SM52=1, because the ready-made CUDA 12 engine holds
no sm_52 code and a Maxwell card must not be routed to it by default.
With the variable set, setup warns about the Maxwell card instead of
refusing. Default: a Maxwell card is "not supported" exactly as before.
docs/OLDER_GPUS.md: a Maxwell row and a short section, with the build
command and a request for an answer-text check.
Tests: tools/test_setup_older_gpus.py (20 tests, new Maxwell cases for
default, opt-in and a newer card present). Not covered: no model run, and
the answer text on a Maxwell card is unchecked.
When HEAD and the one-byte range probe both gave no length, download() took an early EOF as a finished file even though the repository pin (or
ModelScope's published size) names the size. The published size now stands in for the length, so the existing resume, retry and parallel code
publishes the file only after it has all its bytes; a size that disagrees with the pin is still refused before the transfer and the SHA-256
check is unchanged. Ported from PR #2287 by KeyanHu-git with its tests (tools/test_setup_download_length.py).
install_own_path treated a file as present only when <file>.done existed, and that mark is written by setup's own download or by verify_sha256,
which a file from another tool (LM Studio) never reached: setup said it was missing and asked for the full download size in free disk. With
--gguf-dir an existing file and its mmproj go to verify_sha256(own=True), which checks size and SHA-256, leaves a mismatching file alone and
writes the mark. A file that is not in the folder is still reported missing. Ported from PR #3606 by ai-bits with its two tests.
Under WDDM cudaMemGetInfo answers for the calling process: memory another process holds on the card (the image encoder,
a browser, another engine) is not in its free figure. The reporter measured, on an RTX 5090, 30,994 MiB free from CUDA
where NVML read 24,348 with 6 GiB held elsewhere, so --expert-cache auto sized the cache too large and shrank it by
guesses.
device_free_bytes() now takes min(CUDA's figure, NVML's figure for the whole card) when STRATA_VRAM_FREE=device is set.
nvml.dll is loaded from the system directory at first use (LoadLibraryExW with LOAD_LIBRARY_SEARCH_SYSTEM32), the card
is matched by its PCI bus id, and a missing nvml.dll, a missing export or a failed nvmlInit leaves CUDA's figure with
one line on stderr. The serve check of "VRAM free with everything loaded" uses the same function. The value can only get
smaller, so the switch can only shrink the expert cache.
Unset (the default), nothing changes: the NVML code is compiled only for _WIN32 and not for HIP, Metal or the no-VMM
build, and the env variable is read once. The serve check's min() is the same value elsewhere (on a unified-memory
system device_free_bytes() is never below cudaMemGetInfo's figure). Added to the PR: the NVML block is also left out
of a SYCL build (SYCL_LANGUAGE_VERSION), whose CUDA names are a compatibility layer. docs/DETAILS.md: one row in the
measured-and-left-off table instead of the PR's long paragraph.
Tests: builds on Linux CUDA (the Windows-only block is skipped), SYCL (P1006) and on Windows MSVC (owner PC build of
the final tree). Not covered: no run with a model; the Windows gate is written down.
resolve_rope accepted nan, inf, zero and values below 1 and handed them to the engine. It now stops with a message that names --rope-scale and the
valid range, before any configuration is written; valid factors and the automatic scaling are unchanged. Ported from PR #3393 by j-luwierski
(follow-up to #2950) with its tests; the --help text gets the range.
tools/check_cuda_spills.py reads the resource usage the compiler wrote into the kernel libraries (cuobjdump
--dump-resource-usage: STACK and LOCAL per entry function and per sm_) and fails on any Strata kernel outside its ALLOWED
table, which carries the measured allowance of the kernels that hold such an array on purpose; ggml's vendored kernels
(no namespace) are skipped and counted. ctest cuda_spill_guard runs it on eight libraries, skips (77) without cuobjdump or
a demangler, CUDA builds only (HIP has no cuobjdump); check_cuda_spills_unit runs tools/test_check_cuda_spills.py
(stand-in demanglers are sh scripts, so not on Windows).
Changes against the PR: (1) a demangler that exits 0 but prints a blank name used to make the kernel's key empty, which
reads as vendored, so a spilling Strata kernel was skipped; that is now "no verdict" (77) with a test. (2) The unit test is
registered. (3) The ALLOWED table was re-measured on f-027-int: sm_120 with CUDA 13.0 on Windows (MSVC) and sm_86 with the
CUDA 13.0 objects built on Linux (cuobjdump 12.9); allowances are only raised (new DFlash2 gemm8_kernel and
rope_table_kernel, native_gu_swiglu_kernel on sm_86, dequant_kernel / expdec_gu_kernel / prompt_attn_i8v3_kernel /
native_gu_swiglu_kernel on sm_120), none lowered or dropped. Result: 0 violations on both architectures; a test shows the
same object fails once one allowance is lowered.
Ported from PR #2683 (the tracking issue is #2682).
A json_object or json_schema answer that holds LaTeX such as \alpha is
rejected (502) because \a, \g and the like are not JSON escapes. New
opt-in: "repair_latex_escapes": true in response_format (and in
text.format on /v1/responses). Only after the strict parse failed, the
server doubles each backslash that does not start a valid JSON escape and
parses once more with the same duplicate-key and NaN rules, then checks
the schema as before, and prints one log line. Without the option, and
for any answer that parses on the first try, nothing changes.
Escapes that are valid JSON (\b \f \n \r \t \/ \" \ and a well-formed
\uXXXX) are left alone, because they cannot be told from LaTeX: a \beta
next to a \alpha comes back with a backspace. Truncated output, arrays,
NaN and duplicate keys still fail.
Tests: three new tests in serve/test_structured.py (default still 502,
opt-in 200 and stream, failures stay failures, helper cases, Responses
format passthrough). serve/test_structured.py 13 tests OK.
With STRATA_DECODE_TIMING=1 the serial line listed `verify + commit/emit + draft` after `ms/window =`, while the adaptive
wait before verify and the join after the draft were timed but printed after the sum, so the named parts did not add up
to ms/window. The serial line now reads `adapt wait + verify + commit/emit + draft + adapt join + other`, where `other` is
the decode time minus the named parts (not a timed step). The pipelined decode keeps its old line, byte for byte, and
nothing is printed with the timing off or with no windows. Diagnostic only: no decode path changes.
docs/DETAILS.md: a short paragraph on the new line (the hunk was rebased by hand onto the current file).
Tests: compiles on CUDA (the format string was counted against its 22 arguments); no parser in the repo reads
`ms/window = verify`. Not covered: no run of the line itself (it needs a decode).
The quick speed check has its own GPU probe, separate from strata_cli/bench_facts.py, and tried nvidia-smi then rocm-smi only. On a Radeon with
amd-smi alone it wrote the server's vendor stub and 'GPU driver: not read'. The same amd-smi probe as #3519 is added in the shape this file
uses, and a bare vendor name from the server no longer replaces a card the probe named. Ported from PR #3630 by l33tm4st3r.
The three new tests are POSIX-only (symlinks, /sys); run on z440 (Linux, Python 3.12): OK. tools/test_strata_bench.py on Windows: OK.
A batch slot holds the reply it decoded. The main session was parked with the conversation's prompt only (the
admission's copy), so when the next admission or a yielding read overwrote the slot, the reply lived nowhere: the
conversation's next turn found its prompt-read state and read the whole reply again (the author measured 255 of a
256-token reply re-read on four 200K chats over two slots).
What changed:
- park_current is split into park_source (one budget / RAM-admission / snapshot path) and two callers: park_current
(the main session, output line and byte accounting as before) and the new park_slot (an idle slot's ids, checkpoints,
sessions and draft K/V). The ParkSource passes what a park captures; the main session's retained K/V (#1503
take_reuse_for) is taken only for the main session, never for a slot.
- park_slot runs before copy_to_slot at a slot admission and before a BYIELD read overwrites a slot. It does nothing
when the live main session or a parked conversation already holds that state (ConversationCache::holds /
conversation_covers, new, with unit tests), for a slot with pictures, or without --batch-mtp when the drafter is on.
- When a slot is admitted into the main session its checkpoints come along (the shorter ones too), so the chain keeps its
branch points and the parked copies a turn back are superseded (#342) instead of taking an entry each.
- Same refactor in sycl/src/program/generate.cpp.
Changed from the PR: parking is an optimization like a checkpoint (#3350). The PR ended the request (ERR) or the batch
(batch_fatal) when park_slot's snapshot failed; here a failed snapshot, a failed device sync and a host allocation
that runs out are logged, the slot is overwritten as before and the request goes on. The slot's extra checkpoints that
come along at an admission are copied under a bad_alloc guard (the one the request continues from is copied as before).
Tests: conversation_cache_test (new holds() cases and the four-conversation replay at policy level) and
conversation_split_failure_test. Not covered: no model run (the batch path itself needs one); the gate is written down.
The Responses validator still refused anything below 1 after the chat
path made max_tokens of -1 and 0 mean no cap (documented in
docs/DETAILS.md). It now accepts -1 and 0 as no cap; values below -1,
floats, booleans and strings stay a 400. Nothing else changes: the
server already passes max_output_tokens or 0 to prepare(), which gives
the answer the room left in the context.
Tests: OverHttp.test_max_output_tokens_zero_and_minus_one_mean_no_cap
(stream and non-stream, status completed) and
test_max_output_tokens_bad_values_are_400. The first fails without the
change. serve/test_responses.py 43 tests OK.
Problem: with STRATA_Q8_SM60=1 the main model's Q8_0 matrices run on the GP100
Q8_0 kernel, but the MTP drafter's own Q8_0 draft head and projections kept the
generic kernels.
Change: with STRATA_MTP_Q8_SM60=1 (on top of STRATA_Q8_SM60=1) the drafter packs
its Q8_0 projections as packed copies, as the main model does (the prompt path's
draft_kv still reads the GGUF layout, so a copy is kept), and packs the draft
head's subset in place when every column count fits the kernel's shared-memory
planes and the data is 16-byte aligned (otherwise a packed copy). The in-place
pack writes the GGUF blocks back if the device write fails. Release unregisters
the entry without a cudaFree when the data was packed in place. New
native_q8_0_sm60_pack_quiet and native_q8_0_sm60_pack_in_place; HIP returns
false; metal/src/engine/stubs_0145.cpp gets matching stubs that answer
"nothing packed".
Replies do not depend on the drafts (they are verified); only the draft rounding
changes, because the Pascal kernel sums in another order.
Default path: mtp_q8_sm60() is false unless the env is set, so the packing calls
are skipped. The destructor now walks the tensor list once at teardown, not in a
hot loop. The SYCL tree has its own sycl/src/core/mtp.cpp, so it is not touched.
Tests: compile-only on sm_60 (CUDA 12.9) on P1004. Model gate (identical replies
with the flag on and off, decode speed on a P100) is written down in
phaseA-tests/gh027/gates/3751-mtp-q8-sm60.sh.
Not covered: no GPU run here; the contributor measured +4.4% on four GP100 dies
with all experts resident, one run per arm, not on a single streamed P100.
#1337 made a failed engine start in a later calibration step a candidate
that loses. An engine that started and then exited during its own rate
measurement still failed the whole run. measure_candidate() now treats
both the same: the candidate is dropped and recorded in
report["failed_starts"], the earlier results stay. It is used for the CPU
worker step, the adaptive expert tier step and the AMD shared-expert
fork. An operator cancel (CalibrationCancelled) is still raised. The first
start (steps 1 to 3) still fails the run.
src/core/verify.cpp: when adapt_gate_wait times out after 20 s, the
message now says which layer was reached (atomic and copy signal), that
the copy stream is not drained by the release, and dumps the
STRATA_VERIFY_TRACE breadcrumbs when that is on. Diagnostics only, #2188
stays open. The watchdog label is set only once a wait actually starts,
so a gate that is already open costs nothing extra (changed from the
PR, which set it on every call).
Tests: tools/test_calibrate.py 33 tests OK (3 new death tests).
Not covered: engine compile of verify.cpp is checked in the build step.
Problem: on a GP100 die (cc 6.0, no dp4a, no tensor cores) the verify window's
routed down projection for the native packs' IQ4_NL and Q2_0 rows (n_ff 640)
is latency bound. native_down_multi_kernel reads each weight as 2-byte loads,
one item at a time, straight from global memory.
Change: native_down_ws_kernel copies a block's rows (IQ4_NL 32, Q2_0 64, 11,520
bytes) into shared memory with 16-byte loads, then runs the same row dots on the
copy: same sub-warp, lane, items, float expressions and order. Only the integer
sum inside an item changes form (vmad byte selectors, prmt table lookups), the
same int8 x int8 products, so every output is native_down_multi_kernel's (expert
mode 0) bit for bit.
Gating: compiled only in the CUDA build with STRATA_EXPERIMENTAL_SM60 (not HIP,
not Metal). At run time only with STRATA_DOWN_WS=1 on a device of compute
capability exactly 6.0, IQ4_NL / Q2_0 down rows at n_ff 640, and not with
STRATA_OLD_IQ_MMVQ or STRATA_NO_SUB16_GU. Unset, the only change on the default
path is one cached boolean test per native_expert_grouped call.
docs/OLDER_GPUS.md lists the switch. The contributor's microbench numbers (one
PH402 die, Swift 1.5 IQ3_S, one-token window: IQ4_NL 72.6 us mode 8 and 54.5 us
mode 0 become 42.3 us; Q2_0 53.5 / 65.4 us become 38.6 us) were measured on a
four-GP100 rig, not here.
Tests: compile-only on sm_60 (CUDA 12.9) on P1004. The contributor's bitwise
check (native_expert_bench hash against modes 0 and 8) is not run here: it needs
a GP100 GPU run and the model's layers. The gate script is
phaseA-tests/gh027/gates/3867-down-ws.sh.
Not covered: no end-to-end measurement on a single P100 (16 GB cannot hold the
47 GB pack resident).
ROCm 7 and newer ship amd-smi and no rocm-smi, and the amdgpu product_name is empty on gfx1201, so a Radeon AI PRO R9700 run was filed
under 'unknown' with 'runtime: not measured'. bench_facts reads the market name from amd-smi static --asic, the VRAM from the amdgpu sysfs
file of the PCI address amd-smi list reports, the ROCm version from the install (or amd-smi -h), and keeps the probe's card name when the
server only says 'AMD Radeon'. Ported from PR #3519 by l33tm4st3r with its tests and the BENCHMARK_PROTOCOL.md text.
Changed while porting: the amd-smi branch also took macOS (os.name is posix there), which would have hidden the Apple chip; it now skips
darwin, and a test keeps the chip.
The "This computer" card said only "Fits fully in memory." for DeepSeek, with no word of the ~265 GB of disk. For a DeepSeek version it now says "Needs about {N} GB
of free disk space. What does not fit in memory is read from your SSD." (the spill line stays beside it when part of the model is read from the SSD), and the
selected version's line reads "{N} GB RAM + VRAM, plus about {disk} GB of disk". {disk} is not hard-coded: the catalog (tools/strata_mcp.py) gives each set
its `disk_gb`, the set's whole download (model files, n-gram table, drafter, metadata) plus the pack and room to work (dsv41_fit.DISK_HEADROOM_GB), so 2.00 and
2.25 differ. Same card and row markup, no design change; every other model keeps its card. Both sentences are fx80 and fx81 in the eight fix catalogs.
The Terminal's DeepSeek description makes no "fits fully in memory" claim (its lines already state the download and the free disk), so it is unchanged.
Older converters write the Engram scalars as engram.key_length and
engram.pad_id, and omit the sparse-attention geometry. The packer now
reads the old names only when the canonical engram.head_dim and
engram.pad_token_id are absent, and a new --config-json option fills the
omitted geometry from the source model's config.json. A value that
disagrees with metadata already in the GGUF stops the pack with a clear
message. The weights are never rewritten and a set that already has the
canonical names packs exactly as before.
Tests: LegacyModelMetadata (2 new, 5 in the file) in
tests/terminal/test_pack_gguf_engram.py.
Not covered: the author reports the older Q2_K set still needs dense
Q2/Q3 and Engram Q2_K tensor support in the engine; this does not add it.
The final response.completed / response.incomplete event and a
non-streaming /v1/responses reply now carry the same top-level timings
object that /v1/chat/completions has. The server already put timings in
the done dict; the Responses assembler dropped them.
Tests: UsageAndStatus.test_responses (stream and non-stream).
Unset timings leave the event unchanged; response.failed has none.
test_iq_pack imports without llama.cpp's gguf-py and skips the classes that build GGUF fixtures with a named reason; the MCP discovery tests use
a free port, not whatever Strata serves on 8080; the six conversation-cache tests add the repo root to sys.path so they run from the repo root
as the suite does. Ported from PR #3236 by Buddychek (tests only, no shipped code).
The 0.2.5 rewrite of the page dropped the 'Windows AMD: the driver resets' entry (v0.2.4.2 had it), although setup.py still tells a Windows gfx12
card owner to read it and tools/test_setup_amd.py (TdrPointer) checks for it, so that test failed on main. Restored from v0.2.4.2 with the
log and config locations worded for a source checkout.
Three 0.1.42 changes no page told. Ported by hand from PR #3244 by Buddychek (the PR conflicted with the rewritten pages), each sentence
checked against the code: /metrics/prometheus needs the API key like /metrics (serve/server.py _authorized), the prompt-speed row is
(prompt_tokens - reused) / prompt_ms (apps/host/server/monitor-dashboard.mjs), STRATA_VCVARS is read by setup.find_vcvars. The compiler
text now says Visual Studio 2022 (2019 is refused since 0.2.5, #3566) and went to docs/SOURCE_INSTALL.md.
conversation_detail::private_stream() and order_after_queued() kept one non-blocking stream and one event per device in
thread_local storage and never destroyed them. The layer split's later stages run each prompt chunk on a fresh
std::async thread (STRATA_HANDOFF_PIPE off, the default), and a checkpoint part is saved on that thread, so every
periodic checkpoint of every long cold prompt left a stream and an event behind on each later card: a few MiB of VRAM
per long prompt that never came back (#3470: the later card's free VRAM drops 2-6 MiB per long cold prompt, no plateau),
until cudaStreamCreate failed with "out of memory" (#3350: "conversation snapshot running-state copy: out of memory",
engine exit) on a card that was already nearly full.
A thread now hands its handles to a process-wide pool when it ends and the next thread takes them, so the number of
streams and events is the number of threads copying at once. Same stream, same flags, same contract; only the lifetime
changed. copy_streams_created() / copy_events_created() count what the process made.
conversation_copy_streams_test (CUDA, no model, skips without a device): 200 short-lived threads in a row create at most
one stream and one event; ten rounds of six concurrent threads at most six more; the same on a second device.
A conversation checkpoint is an optimization, like parking, but the writer treated a failed copy as fatal: "conversation
snapshot running-state copy: out of memory" (cudaErrorMemoryAllocation from the snapshot copy: the private copy stream's
first creation or the driver's staging on a nearly full card, not host RAM) made the prefill fail and the engine exit,
21 times in 4 h for one reporter and 16 in 34 h for another, each exit a reload and the in-flight requests lost.
Now, after a failed save, checkpoint_failure_is_survivable() clears the non-sticky allocation error and synchronizes the
device: when that is clean the checkpoint is skipped with a line on stderr and the request goes on (a layer split's stage
part is left empty, which already means "no checkpoint at this length"; the message-boundary / turn / root checkpoints
after a prompt chunk do the same). A device that is unwell (a sticky error, a GPU that stopped answering) still ends the
request exactly as before. Same edits in sycl/src/program/generate.cpp.
STRATA_TEST_COPY_FAIL_AT=N (tests only, like STRATA_TEST_SLOW_COPY_US): the N-th snapshot copy of the process fails with
cudaErrorMemoryAllocation, so the case can be made without a full card. Unset, nothing changes.
Not changed: "verify: instantiate: out of memory" (graph capture), which the reporter saw once.
CMakeLists.txt tested "NOT DEFINED CMAKE_CUDA_ARCHITECTURES" after enable_language(CUDA). Since CMake 3.18 that call
fills an unset variable with nvcc's default (sm_75 for CUDA 13), so the sm_120 default never applied: a hand build with
no option was an sm_75 build. The default is now set before enable_language, unless the user gave
-DCMAKE_CUDA_ARCHITECTURES or the CUDAARCHS environment variable, and the configure step prints a warning that says what
to pass for another card (a recommendation, not a refusal).
With -DCMAKE_CUDA_ARCHITECTURES=native, the one NVFP4 W4A4 unit (mmq_nvfp4_w4a4.cu) stayed on plain `native`, where
ptxas refuses the FP4 MMA on an RTX 50. `native` is now replaced by CMAKE_CUDA_ARCHITECTURES_NATIVE (the detected
architectures) before the 12x -> 12xa rewrite, as a list, so a mixed PC works too. docs/SOURCE_INSTALL.md says what
`native` and the other values build.
Release and setup builds pass their architectures explicitly and are unchanged. Configure matrix, CUDA 12.9 on Linux:
none -> 120 (+ the warning), CUDAARCHS=86 -> 86, 89, "75;86;120" (w4a4 120a only for the 12x entry), native -> detected.
Ported from PR #4029.
--adapt-decay took any number. Above 1 the adaptive tier's usage counts grow without end (inf after a few hundred
adaptations) and the swap gains turn NaN, which breaks the sort that orders the swaps; a negative value flips the counts.
The value is now checked where it is parsed (include/strata/program/adapt_decay.hpp, also in sycl/): a finite value in 0..1
is used exactly as before (0 and 1 included, the range the calibration profile already allows); anything else, NaN and inf
too, runs with the default 0.7 and says so on stderr. A warning instead of the PR's exit 2: a risky pick is reported,
the engine still starts.
adapt_decay_test (25 checks) covers the range, the fallback and finite counts after 100000 adaptations.
Ported from PR #1284 (the check; the refusal became a fallback).
The sandbox of a test start was deleted at once, so the PID records that show why a start failed were lost; only the last 64 KiB of the engine
log stayed. A failed start now also keeps run.json (without keys), the host, child and ready JSON records of the runtime (tokens cut to their
first 8 characters) and the host and server logs in .updates/logs/test-start-<version>-failed. The last three failed starts are kept.
Best effort, never hides the failure. Tests: tests/terminal/test_updater.py (FailedStartRecords).
The manual build in docs/INSTALL.md uses CMake's default generator (Unix Makefiles on Linux) in build/, and setup --build always asks for
Ninja, so cmake stopped with 'Does not match the generator used previously'. cmake_build now reads CMAKE_GENERATOR from the folder's
CMakeCache.txt; when it is not Ninja it says so in one line and removes the cache and CMakeFiles (no other file). A Ninja folder or a new
folder is untouched. Tests: tools/test_setup_build_dir.py.
A cmake -P script has no cmake_minimum_required, so CMake 3.x used the OLD behaviour and stopped at 'if given arguments ... IN_LIST: Unknown
arguments' (ctest ggml_tq_overlay with CMake 3.31). Checked with pip cmake 3.31.6 and 4.4.3 against the pinned llama.cpp: both print
'ggml_tq_overlay_test: ok'. A static test keeps the policy before the first IN_LIST in every cmake/*_test.cmake.
The script ended with a grep pipeline, so a failed compile exited 0 (and, before the ocloc check was added at the end, a good build exited 1).
A build.sh && next-step chain then ran after a failure. The status of cmake --build is kept and the script exits with it. The test runs the real
script with a stub cmake on PATH (good build: 0, failed compile: 1; the old script exits 0 on the failed compile).
tools/test_setup_golden normalized strata.exe but not strata-vision.exe, so all 25 baseline entries failed on Linux (48 subtests). VEXE
and HVEXE are mapped to <VEXE> and <HVEXE> like EXE, and the baseline records <VEXE>. test_nvfp4_local_pack imports pytest, which is not in
requirements.txt; under python -m unittest discover the module is now skipped with a reason instead of failing to import.
Substitutions were ordered by the length of their regular expression, and a user name wrapped in word bounds outranked the Strata or home
folder that contains it, so a report kept /home/<user>/Strata instead of <strata>. They are now ordered by the text they replace. Ported from
PR #3619 by ctr-137 with its test (Linux and Windows layouts).
The page said that starting a model never downloads and that nothing checks for updates in the background. Since 0.2.4.2 strata launch with a
DeepSeek model checks the signed feed (10 s at most, silent without a connection) and replaces an older installed DeepSeek engine package.
Ported from PR #4028 by pschaub (wording from issue #4022), checked against strata_cli/launch/runner.py upgrade_dsv41_engine.
engine_env put the ROCm library folders first on PATH, which only Windows' loader reads. On AMD Linux strata --families exited 127 on
libhipblas.so.3, so setup falsely said the engine does not run Qwen3.6. On Linux and elsewhere the folders now go first on LD_LIBRARY_PATH
(an existing value is kept); Windows keeps PATH. Ported from PR #3305 by xxxliqu1dxxx (resubmission of #2914), with its tests.
metal/setup_mac.py returned right after the engine step on --update while the model questions had been asked, so choosing a model the Mac did
not have ended with no download and no word. The update now names the missing model (or image encoder) and points at ./setup.sh --setup; an
update for a model that is installed prints what it printed before. Ported from PR #3848 by demetree, with its two tests (mocked; not run on a Mac).
adopt_user_defaults() set the token's default DACL with FA (FILE_ALL_ACCESS). That default also applies to new process objects, and the file
mask has no process query or set-information right, so a child could not be reopened by PID: the HighQoS helper timed out and
procs.start_token could not record the child. GA lets Windows map the mask to the object being created; the principals are unchanged.
Ported from PR #3428 by kossum (QoS reopen test). Test from PR #3399 by midhatn (a live child is identified and recorded) added; both fail
without the change and pass with it.
Co-authored-by: MIDHAT NASHAR <[email protected]>
_start() checked cfg['exe'] with Path.exists(), so a run config whose engine is component:strata-engine was refused as a missing file
before the server could resolve and verify it. Ported from PR #4027 by jubeater. A component reference is no longer tested as a file; model
files and literal engine paths are checked as before. ensure_engine_for, the Windows HIP runtime repair and the HIP architecture lookup
skip the source-build engine folder for a component. Eight tests (tools/test_setup_component_start.py).
Ported from PR #4011 by SuperDistance. project_version() read CMakeLists.txt with Path.read_text() and no encoding, which raised
UnicodeDecodeError on Windows code pages such as cp936. The test makes read_text default to each of 14 Windows code pages. The same crash was
reported in #3235 and #3414.
.git is not in the build context, so package_terminal.py had no commit to put in the archive and the documented source build stopped.
Ported from PR #3644 by robertoetcheverryr: ARG STRATA_COMMIT / ENV STRATA_BUILD_COMMIT as in docker/Dockerfile, and the two 'From source'
commands in docs/DOCKER.md pass --build-arg STRATA_COMMIT=$(git rev-parse HEAD).
Added: .dockerignore leaves out the folders docker/entrypoint.sh creates in a mounted data/ (install, state, home, Strata-data). A dangling
data/install/third_party/llama.cpp link in a clone that had been mounted made the Terminal stage fail with a copytree error (seen in the
review thread of #3644). Not built with docker here (no docker on this PC).
docker/build.sh staged only top-level tools/*.py, so the Terminal stage stopped at 'installer tool missing: tools/dsv41/tune.py'.
Ported from PR #3974 by hurui200320: the pathspec is tools/**/*.py, the no-.git fallback finds nested files, and the context test reads
INSTALLER_TOOLS and INSTALLER_SCRIPTS from packaging/package_terminal.py. Fixes the failure that images built from v0.2.4.1 and v0.2.4.2 had.
The portable lifeline watcher built a child's identity from `ps -o lstart=,stat=` and kept the whole line, so the token
changed with the process state (R while it starts, S asleep). The watcher compared its own reading with the token taken
at start, decided it was not the child it was told about and exited without ending it (test_lifeline "the child outlived
a killed server (watcher)", #3407). The identity is now the start time alone: field 22 of /proc/<pid>/stat on Linux
(a zombie has none), `ps -o lstart=` elsewhere; strata_control/procs.start_token drops the state column the same way.
A launch that is not a real Popen is not watched.
Telemetry._loop no longer ends on the first failing sample (a machine whose GPU or psutil source is unreadable): it
prints one line and keeps sampling, and NoLeakedSampler checks the thread itself instead of counting threads.
New test: the token is the same while the child runs (R) and sleeps (S); on the old code it fails on Linux
('...R' != '...S').
Ported from PR #3590.
A declared namespace.member tool (Codex MCP: mcp__filesystem.write_file) was accepted only as namespace__member (#3556).
Qwen3.8-Flash-Next also writes namespace=member (the <function= tag's equals sign ends up in the name), namespace_member
(one underscore dropped) and the bare member. Each is now an alias of the declared tool when exactly one declared tool
produces it and no declared name is the same; a name that matches nothing stays text (#3075), two namespaces that share
a member are not guessed, and a declared flat tool always wins. Requests without namespaced tools are unchanged.
serve/test_tool_name_alias.py (10 tests, 5 of the first 9 fail without the change; the tenth pins that a flat tool named
like the underscore alias wins), docs/DETAILS.md names the spellings.
Ported from PR #3972.
With --batch and conversation parking, a restore keeps the restored K/V (ConversationCache::retain) so the next park of
the main session copies only what changed. The reuse tracked a token count (unchanged_tokens), not whose tokens, and
copy_from_slot (a batch slot giving its conversation back to the main session) replaced the main session's K/V without
dropping the retained image. The next park reused the earlier conversation's K/V for the new one's prefix: a restore of
the new conversation then answered from the other one's history.
retain() now records the ids of the conversation the K/V belongs to; a park calls take_reuse_for(live), which reuses only
the prefix live shares with those ids (nothing when they share none or the ids are unknown), and one stderr line is
printed whenever a reuse is cut. copy_from_slot drops the retained image before it replaces the main session's K/V.
Same edits in sycl/src/program/generate.cpp. The guard only cuts reuse: at most one full copy of a parked conversation's
K/V, the saved bytes are otherwise identical; no flag, no default change.
Ported from PR #1503 onto the current tree (the first-stage ids sit after the stages field so the existing aggregate
initialisers keep their meaning). conversation_cache_test gains the cross-conversation cases of the PR and the logged
sequence (restore P, a slot hands Q to the main session, park Q): another conversation never gets the K/V, a diverging
one only the shared prefix on every stage, a slot hand-back drops it, unknown identity is never reused. Removing the
ids limit makes the first case fail.
A command on the allowlist used to run without a card with any option except -delete/-exec/-execdir/-fprint/-fls/-ok. Write modes (sort -o,
tree -o, uniq IN OUT), program-execution modes (sort --compress-program, rg --pre, git grep -O), date -s and tail -f got through, and so did
git config -e, git branch --set-upstream-to and git remote ... forms.
Each command now has a spec of the options and operands that only look (short and long, spelled in full); every other option, abbreviation
or attached form asks. git subcommands have the same specs, a branch/tag name is taken only with a list option, git config only reads with
--get. date takes only +FORMAT operands, uniq at most one operand, hostname none. less and more leave the list. Quoted texts are checked as
operands too. `du -sh` no longer trips the 'sh' word check.
Tests: apps/host/tests/gate-readonly.test.mjs (deny and allow cases for options, operands, spellings, quoting; every listed command asks for
an unknown option).
sample() applied the repeat / frequency / presence penalties to lg[id] for every token id of the last penalty_last_n
tokens of the history. A row shorter than the vocabulary (the MTP draft head picks among the first N token ids,
--draft-vocab N) made a history token with id >= N an out-of-bounds read and write on the host. A token that is not in
the row takes no penalty; the loop skips ids < 0 or >= lg.size(). A full-vocabulary row is bit-identical.
The loop and Sampling move to include/strata/qwen35/sampling.hpp (header-only) so a CPU unit test can run them:
src/qwen35/sampling_test.cpp (qwen35_sampling_test) checks nothing is written past a short row, that in-row tokens
take the penalty, and that a full row matches the unguarded loop bit for bit over 108 penalty/window mixes. Without the
guard the test dies with an access violation.
Co-authored-by: yehya <[email protected]>
cache_settle()'s store preload (DS4_SETTLE_STORE) read 2,243 hot experts into free store slots after every chunked prompt: 4,713 ms of a 4K prompt. Measured on the box (code fixture, prompt + 256 tokens, interleaved, 3 runs a side): with it on the decode is 13.9/14.4 s, with it off 14.2 s - no benefit - while the 4K prompt goes 31.3 -> 26.6 s (130.89 -> 153.78 tok/s). The swap loop above it is what the decode uses (it holds the CPU lane at ~5-7 k experts a run against ~14 k without), so that keeps its budget; DS4_SETTLE_STORE=1 restores the preload.
On the RTX 5070 (--agent-long, 22 tasks) split 36 kept agents at 6/10 (exact 4 and 7) and extraction at 585/606 facts (exact
585) for +11% prompt speed at 14K and +18% at 30K; the split-24 default passed 0/10 agent runs and found 556 facts, a tail of
8192 with split 24 found 574. KVA stays off unless STRATA_KVA=1. docs/KVA.md gives the numbers and the reasons; kva_test
checks the opt-in's split 36 / tail 2048 and the overrides. The geometry gate is unchanged.
The app toggle will write STRATA_KVA=1 for Qwen models in general; the split layer, the tail and the measured quality apply to
Qwen3.8-Flash-Next's shape only (48 layers, 512 experts). Any other shape (Qwen3.6-35B-A3B, ...) now logs "KVA configured but
NOT applied: this model's shape (L layers, E experts) is not one KVA was measured on" and reads exactly; the routine reasons
(short prompt, continuation) still come first. STRATA_KVA_ANY_MODEL=1 allows it, with one warning. kva_test and docs/KVA.md
cover both.
KVA was measured on the Flash-Next shape (48 layers, 512 experts) and the engine declines any other shape with a log reason, so the row would do nothing
for Qwen3.6-35B-A3B. Setup writes "arch" in a run config only for the families with an engine path of their own (Qwen3.6-35B-A3B: qwen35moe; DeepSeek:
deepseek41); every Qwen3.8-Flash-Next file (the original, Swift, Coder, Unsloth, Gyro, NVFP4) has none. The row is therefore shown for a model without an
"arch" on CUDA or HIP, in the app's Native engine settings and in the Terminal's Advanced table. The Radeon RX 9000 and DeepSeek rows are unchanged.
Tests: a Qwen3.6 config gets no KVA row (CUDA and HIP), the six Flash-Next family names do, and through the host fixture a model with an arch has no switch row.
The quality runs found lost records and invented rows in long extraction tasks and defects in agent loops with a ~10K
system prompt when most of the prompt was approximated. Opt-in now: unset (or 0) nothing runs, the draft layer included.
Switched on, the tail is 8192 tokens (STRATA_KVA_SUFFIX overrides), the safest configuration measured. A model other than
Qwen3.8-Flash-Next's geometry gets a one-line warning instead of a refusal. Kept: the untouched draft layer (cell floor
behind STRATA_KVA_DRAFT_FLOOR=1), the decline reasons, STRATA_KVA_SKIP=moe, the tests. docs/KVA.md says why it is off and
how to turn it on; kva_test checks unset = off and STRATA_KVA=1 = on with the 8192 tail.
The Speed row now follows the shared table's default: Faster long prompts reads "Off (recommended)" / "On" and its new sentence is "Reads very long prompts
faster by approximating part of the work. Off by default: answers about long documents can miss or mix up details." (replaces fx73; "Off (recommended)"
is fx79; translated in zh, ja, de, es, ru, pt, ko). The other two rows are unchanged ("On (recommended)" / "Off"). A STRATA_KVA=0 stored by an earlier
0.2.7 test build shows as Off. Still shown only where KVA applies (Qwen on CUDA or HIP). tests/speed-switches.test.mjs updated: the default launch has
no KVA entry, On writes STRATA_KVA=1, Off removes it, the old 0 shows as Off, and the other two switches are tested as before.
Owner decision (26 October, conditional on the last KVA variant results): KVA ships off. strata_control/speed_switches.py now gives every switch a default
and the environment it writes outside that default: KVA is default off and writes STRATA_KVA=1; the Radeon RX 9000 and DeepSeek switches stay on by
default and write their off entries as before. A launch in the default carries no entry. A file that holds STRATA_KVA=0 (the off switch of an earlier
0.2.7 test build) reads as off, which is what it is now; choosing Off cleans it up.
The Advanced table row says "off by default; set 1 to turn on" with the new sentence (`--set STRATA_KVA=1`; `off` or Del puts the default back).
tests/terminal/test_speed_switches.py covers the defaults, KVA on/off/old-zero, the rows and the comparison with the app's table.
The default stays the state-only form (byte-identical). With STRATA_KVA_SKIP=moe the GDN recurrence (gate projection, output
normalization, output projection) and the QSA attention (query side, selection, attention, output projection) run in full
for layers >= split, their outputs go into the residual, and the loop leaves before the MoE half; the expert transfers of
those layers are still skipped. The KVA line says "mode moe-skip". Motivation: the extraction losses of the state-only form
come from late-layer K/V and state built from a stale residual; this keeps the attention halves' contribution in it.
kva_test covers the switch.
fast_attn_enabled() was opt-in (DS4_FAST_ATTN=1); with the prompt chunk at 1024 it is worth far more than when
it was gated: measured on 4x R9700, 2.25 bpw, 8K code prompts, three warm prompts a load, the same ids
(5705e1164a) --
without it 1159 / 1180 / 1195 ppts (median 1180)
with it 1327 / 1354 / 1360 ppts (median 1354, +14.7%)
and after this commit the un-enved default reads 1314 / 1357 / 1369 (median 1357), i.e. identical to setting the
flag, so a deployment no longer needs any environment at all. The QK/PV sums are the matrix cores' order, which
the owner already approved for the prefill; the decode path keeps the scalar kernels. DS4_FAST_ATTN=0 turns it
off (it is a no-op on CUDA, whose dispatch is not compiled).
ring_mb_auto() returned a flat 400 MB on any box whose experts do not fit in VRAM, but on such a box the ring is exactly what lets a prompt queue its likely experts early. 400 -> 1200 MB measured on the box (settle off): the queue 3,200 -> 8,087 experts, the disk's effective rate 1.49 -> 2.44 GB/s, the compute's exposed read wait 11.4 -> 3.5 s, the 4K prompt 148.98 -> 190.98 tok/s; 16K 290.97 -> 315.01, 32K 316.92 -> 325.89. Decode unmoved in an interleaved A/B (17.54/16.39 against 17.61 tok/s). It is now a sixteenth of the VRAM room clamped to 400-1200 MB, so a smaller pool still loads; DS4_RING_MB still overrides.
c.A: the attention slab decides how many times each dense matrix is re-read a prompt. 8 -> 32 MB (STRATA_DS4_A_MB): 32K 325.89 -> 357.78 tok/s, 16K 315.01 -> 328.04, 8K +1.6%, 4K flat, decode unmoved (16.81 against 17.60 tok/s interleaved, noise-limited) for ~0.4 GB of VRAM.
Best measured prompt set with the two new defaults (DS4_SETTLE_MS=0;DS4_SETTLE_STORE=0, which is a prompt-only trade): 4K 195.18, 8K 258.47, 16K 328.04, 32K 357.78 tok/s. With the settle work at its default a 4K prompt is 31.3 s = 130.89 tok/s (was 88.30).
The dense GEMM was 11.1 s of a 4K prompt at hipBLAS's measured 4.97 TFLOPS (M 4096 N 2304 K 5120, the engine's call convention). A blocked WMMA kernel reaches 23.96 TFLOPS at that shape (128x128 tile, BK 32, 2-stage LDS, rocwmma; scripts/wmma2.cu) and the engine's own tuned kernel (BM 128 x BN 256, grouped tile order, raw wmma intrinsics) 20.8-20.9 at the wq_b shape and 38-41 at M 16384. gemm() now calls strata_pf_gemm_f16_ld with hipBLAS as the fallback (rows < 64, N < 512, K % 32, unaligned X: unchanged); STRATA_DENSE_PFGEMM=0 restores hipBLAS for an A/B. Same operands, the kernel's own k order (rounding-level).
Measured on the box (eng-028, 4E224E7F -> 467D7320, 4K prompt fixture p4k, chunk 16384): 46.4 -> 38.1 s, 88.30 -> 107.63 tok/s; with DS4_SETTLE_MS=0;DS4_SETTLE_STORE=0 27.5 s = 148.98 tok/s. 8K 134.66 -> 218.31, 16K 158.6 -> 290.97, 32K 178.9 -> 316.92 (settle off). Decode unchanged in an interleaved A/B (old 15.17/16.73, new 17.38/18.48 tok/s) with acceptance 151/164 = 92% (93/98/100/100/90 per position).
c.A (the attention slab, the reason each dense matrix is re-read 10-16x a prompt) takes STRATA_DS4_A_MB: 32 gives 107.63 -> 110.67, small because the expert reads now dominate.
The Models screen listed the shared sizes for DeepSeek but installed and loaded it at 8K whatever was chosen. Now the choice counts (8K stays the
default for a version that is not installed) and the list is DeepSeek's own: 8K to the limit the installed pack states in ds4.txt (max_position_embeddings,
read by the host as `contextMax`); a pack without it behaves as before (up to 256K). Past 256K the existing note slot under the model shows: "Not tested yet.
A context this long takes graphics memory from the model and reads long prompts slowly." (fx78 in the eight fix catalogs). Qwen's choices are unchanged.
src/lib/deepseek-context.ts is the app's copy of strata_control/dsv41_context.py; tests/terminal/test_dsv41_context.py compares the two.
The model reads 1,048,576 positions natively. Setup and `strata launch` offered DeepSeek the shared list that stops at 512K and dressed anything past
256K in Qwen's "rope scaling" words. Now:
- strata_control/dsv41_context.py: the choices are 8K, 32K, 64K, 128K, 256K, 512K and 1M (labelled "512K" and "1M"), limited to the pack's own limit
(max_position_embeddings in its ds4.txt) when it has one; a size past it (2M, or 1M on a pack that says 256K) is the one refusal; past 256K a plain
note says it is not tested yet and what it costs.
- setup.py: --context is checked before any download and again against the pack once it is built; the install tool refuses the same sizes.
- Terminal: the context step lists DeepSeek's own sizes (and the note beside them), a typed size is checked, `--context 512k` / `1m` and 524288 / 1048576
are accepted, plan and launch carry the note as a warning; Qwen's list and wording are unchanged.
- Engine: read, no change needed. strata-dsv41 takes --max-context as given (ds4_main.cpp) and sizes its buffers from it (ds4_model.cu); nothing clamps it
below the pack's limit, and the rope settings come from the pack. A test keeps it that way.
- docs/DEEPSEEK41.md: the cost (graphics memory per position taken from the expert cache, slow reading of very long prompts) and that 512K / 1M are untested.
Measured on a 5070 (IQ3_S) the cell floor cut decode after a 14K / 30K approximated prompt by 22 / 34 percent (drafts
accepted 100 -> 83 / 73 percent): a draft layer that reads only the exact tail loses its context. The approximated
chunks' draft cells are usable, so by default nothing about the draft layer changes: not the window, not the cells
built, not the attention. The floor machinery (MtpDrafter::set_cell_floor, window_ids_floor, KvaFloorRegistry) stays
behind STRATA_KVA_DRAFT_FLOOR=1 for research. kva_test now really covers window_start with floor 0, the registry and
wants_cell_floor.
The first default-on build set --mtp-window 2048 once per process whenever KVA was on, so short prompts (where KVA never
applies) lost draft context and decode speed. Now the draft layer keeps its window (32768 by default) and
MtpDrafter::set_cell_floor() names the first draft cell built from a real row: a KVA read raises it to the end of its
approximated chunks (they are neither built nor read), the draft attention starts at max(floor, pos - window)
(window_ids_floor, window_start.hpp), and KvaFloorRegistry gives every state a request continues from (live, checkpoint,
parked) the floor of its own tokens. A session KVA did not approximate has floor 0 and attends exactly as with
STRATA_KVA=0; with STRATA_KVA=0 or a process KVA is ruled out for the draft layer's code is 0.2.6's. The tail is 2048
whatever the window. Unit tests: window_start with floor 0 equals the plain window, the registry, wants_cell_floor.
- locale-gaps: the sentence added in 72c37ea2d (RTX 30/40/50, R9700, Strix Halo; "the line below says what to expect") had no key. fx77 in the eight fix catalogs.
- packaged-control: findControl was given an env without STRATA_COMPONENTS_DIR, so it fell back to the real per-user components folder and found the
updated Control installed on the PC (read only). The test now passes a scratch components folder.
Three rows under "Speed" in the native settings, in the existing row markup with the existing Select (On (recommended) / Off): Faster long prompts
(Qwen on CUDA or HIP; off writes STRATA_KVA=0), Faster answers on Radeon RX 9000 / AI PRO R9700 (Qwen on a gfx1200/gfx1201 card; off writes
STRATA_PCIE_FRAC_DEFAULT=old) and DeepSeek speed-ups (DeepSeek; off writes the six DS4_* = 0 of the contract). Each has a plain sentence saying what
turning it off does. A row is shown only where it applies, like the DeepSeek Speed row; the host decides from the model's file and the hardware report.
The switch is the model's file "env": off adds exactly its entries, on removes them, so a default launch is byte for byte what 0.2.6 started and the
Terminal, Control and the app reach the engine the same way (serve/server.py child_env). The host now shows and edits STRATA_* and DS4_* names (the
Strix Halo async entry is visible and kept); other names stay untouched. New strings are fx70..fx76 in the fix catalogs, translated in zh, ja, de, es,
ru, pt, ko. tests/speed-switches.test.mjs covers the default, each switch off and on again (same file), where each applies, and the catalogs.
strata_control/speed_switches.py holds the engine switch contract (Faster long prompts: STRATA_KVA=0; Radeon RX 9000 / R9700: STRATA_PCIE_FRAC_DEFAULT=old;
DeepSeek speed-ups: six DS4_* = 0), where each applies (Qwen on CUDA/HIP; Qwen on a gfx1200/gfx1201 card; DeepSeek) and the one-sentence effect text.
`strata launch` shows the environment names of the switches that apply as rows of the Advanced table (default "on", the sentence beside them); typing the
off value (`--set STRATA_KVA=0`) writes it into the run config's env, `off` takes it out again. A default launch carries none of them.
tests/terminal/test_speed_switches.py checks the names, the applicability, the config a launch writes, the rows, and that the app's table
(apps/host/server/native-settings.mjs) says the same.
The tail follows the draft layer: with KVA on and no --mtp-window given, the draft window becomes 2048 cells and the
fully computed tail 2112 (a given --mtp-window N sets the tail N + 64; a window of every cell, or STRATA_KVA_SUFFIX, is
handled and named). Layer splits, batch slots, --peer-device, pictures, the DFlash2 drafter, other model geometries and
the Metal build read exactly, with the existing "NOT applied: reason" line (quiet for continuation and short reads unless
STRATA_KVA=1). The reasons live in KvaConfig::decline and are unit tested (kva_test). Later chat turns continue from the
live session or a checkpoint and read their new tokens exactly over the cells the first read left; docs/KVA.md says so.
The slab rule (8 MB of q/o per device, 256 rows here) was set to keep the dense dequant's VRAM small; with a
1024-token chunk it dequantizes q_b/wo once per 256 rows - four times a chunk, 32 times an 8K prompt.
Measured on 4x R9700, 2.25 bpw, share 0.77, chunk 1024, 8K code prompt, ids 5705e1164a:
256-row rule 1204 / 1236 / 1250 ppts, decode 77.7 tok/s
chunk's 1024 1324 / 1353 / 1363 decode 78.1 (+9% prefill, decode unchanged)
The 1024 rows cost 268 MB a device (1% of the card). So high-share runs take min(chunk, 1024) rows and the
old rule stays elsewhere; DS4_ATTN_SLAB still overrides. The un-enved default verified at 1335 ppts /
77.8 tok/s with the ring at 253 MB.
With the 1024-token prompt chunk (the default where the cards hold most of the experts, abfe7d813) a big
staging ring is no longer the trade it was: every MB it takes is a tenth of an expert the decode keeps.
Measured on the same box, 8K code prompt + six 128-token decodes a load, ids 5705e1164a / 81d895366c:
2000 MB 1243 ppts / 73.2 tok/s
1000 MB 1287 / 77.5
500 MB 1314 / 78.5
256 MB 1324 / 78.4
128 MB 1338 / 78.3
So high-share runs take clamp(room/device * 0.008, 128, 512) MB (248 MB here) and the old deeper ring stays
for the cases it was measured for (share < 0.5, where the cards stream most of the experts). The un-enved
default verified at 1339 ppts / 78.1 tok/s, 42 s of summed GPU time in a 36 s pass.
prefill_chunk_default() returned 1024 only for a single-card run (own_default); a multi-card run whose cards
hold most of the experts fell through to 2048 (HIP skips the two 4096 branches) - and the rule's own datapoint
for 1024 was a 4 x R9700 run ('1024 370 tok/s, 8192 214'). Measured again here, 2.25 bpw, 4 cards, 8K code
prompt, --dspark auto, ids 5705e1164a, 4 prompts a load:
512 1110 / 1137 / 1142 ppts (median 1123)
1024 1229 / 1257 / 1270 (median 1243) <- this commit's default
2048 689 / 695 / 696 (median 692) <- what the engine picked before
8192 690 / 690 / 695 (median 690) <- what the 0.2.7 configs passed
Decode unchanged (73.2 against 73.2 tok/s medians, --dspark auto). So the owner's default 8K prompt goes from
692 to 1243 ppts (+80%) with no flag; --prefill-chunk still overrides.
DeepInfer's rule: verify a draft while the product of the confidences so far stays above the cut, instead of
only each draft's own floor. The 0.2.7 notes left it off ('not measured faster on code: H100 / owner PC'),
Project Maya runs 0.5 on a Radeon 8065S (+6.5% decode, the same tokens).
Measured on 4x R9700, DeepSeek 2.25 bpw, six 128-token decodes an arm in one load, same ids
(81d895366c): 71.7 -> 76.9 tok/s (medians) with STRATA_DS4_DSPARK_CUM=0.5, and a second pair 74.1/73.6;
the default with the rule on (no env) verified at 74.5 tok/s (6 decodes). CUDA keeps 0 (the rule was flat
there). STRATA_DS4_DSPARK_CUM still overrides either way.
The prompt path copies every streamed expert from the RAM copy into the pinned staging before its DMA (the
file's own note: 6.8 GB/s through the staging against 24-28 GB/s probes). Registering the copy lets it DMA
in place. Three interleaved pairs on 4x R9700, 2.25 bpw (share 0.72): the second 8K prompt of a process
12.0 / 11.9 / 11.9 s against 28.8 / 13.8 / 12.3 s unpinned; decode 35.42 / 35.50 / 34.9-36.0 against
35.43 / 35.45 / 34.9-36.0 tok/s, i.e. no cost; ids unchanged. Costs 23 s of load (registering a 50 GB
copy) - once a process, which for a server is the right trade. Auto only when the expert share is >= 0.5
(the CPU lane is then a minority of the decode, so the pinned pages' TLB cost cannot bite); DS4_COPY_PIN=0/1
forces it, and it stays off where the 12 GB cards' decode is CPU-lane-bound (the -15% measured there).
The RAM->staging memcpy runs on the IoQueue's threads; on a warm 8K prompt (stream stage 14.3 s, 12.8 GB a
device) they are busy 4.1 s, i.e. ~3 GB/s a device - a real share, and the thing to check in the 50-95 s
first-prompt runs (where the same counters are what to read first). DS4_PROF only.
The first prompt after a load is 2-9x slower than the second on the 4x R9700 2.25 bpw cell and it is not
the disk (0 MB read during it), the RAM store (removing it is worse), the dense GEMM shapes (q_b+rope
1.46 s) or the compute waiting (0 ms). The lead is the RAM->pinned-staging->DMA path (the file's own note:
6.8 GB/s through the staging against 24-28 GB/s probes); DS4_COPY_PIN=1 measured 47.9 -> 31.4 s on the
first prompt and 28.8 -> 12.0 s on the second, decode unchanged. This commit only adds the counters
(staging / DMA issue / slot wait) to the existing prompt stream line under DS4_PROF; no behaviour change.
No behaviour change on the 4x R9700 cell (126.8 GB of room / 4 = 31.7 GB a device, 8% -> 2000 MB, the value
measured in 76177ec34); on other card counts it is the per-device share that was meant.
The prompt path streams the experts the cards cannot hold. ring_mb_auto() gave 16 MB whenever the cards
held >= 1/1.3 of the experts, which is exactly this 2.25 bpw cell (126.8 GB of room, 156.3 GB of experts,
share 0.769): the stream then had no depth and the expert tiles waited 1.7 s inside a 5.2 s stage. A share
of the VRAM room (5%, 400-2000 MB) when any expert spills, 16 MB when they all fit (the 0.2.6 rule, kept
for the decode it was chosen for).
Measured, 4x R9700, 2.25 plain pack, 6 experts, --dspark off, 8192-token code prompt, ids 5705e1164a:
8K warm 632 -> 675-678 ppts (+6.8%, 5 prompts a run, 3 runs), agent decode 35.9 -> 35.3 tok/s (-1.7%,
4 runs). The profile of the warm pass: expert stage 5185 -> 4034 ms (kernels 3509 -> 2645 ms, the same
kernels now that the weights arrive in VRAM ahead of them, waits 1676 -> 1389 ms), everything else equal.
Also in this commit (opt-in, defaults byte-identical except the F16 activation path which is bit-identical):
- SetArgs::x16: gate/up take the layer's F16 copy of the activations (rows_f16 rounds exactly as the tile's
own smem store did) - half the activation bytes, 26% faster gate/up in the standalone bench
(box/expert_bench.cpp), +0.3% at the pass level (the stage is stream-bound).
- HIPK_CFG 6..11: tile variants for the expert GEMMs (64x128, 32x128, 128x256, 64x128/4 waves,
128/128+down 256, 64x256). 64x128 (cfg 6) is 21% faster than 128x128 in the bench and +0.9% in the pass.
- HIPK_DBG 1/2/3: switch the MMA / the decode off in the tiles (the breakdown in 09-marlin-kernel-diary.md).
- ds4_iq_hip.inc: a UTF-8 BOM on a source file removed.
It now ships on at 2.0 for a card of 13 GiB or less: +37.6% story / +35.2% code
decode over five interleaved pairs on an RTX 3060 12 GB, and it stays off on
bigger cards where it measured neutral to negative (l620 -1.9%/-0.5%, z440
-4.2%/+11.0%). It changes which experts run - the owner approved that trade
for 12 GB cards - and STRATA_ROUTE_PRIOR=0 turns it off.
The design was HIP-only: 80 resident blocks, a column slice of D each, two
grid barriers on an atomic counter, a barrier timeout so a card that cannot
hold all 80 blocks resident turns a hang into a refused start-up check. Ported
by giving grid_sync a CUDA branch (volatile generation counter + clock64() for
the timeout, scaled to SM clocks), guarding the gfx-arch refusal to HIP and
adding a masked-shuffle helper.
On the RTX PRO 5000 Blackwell 72 GB it passes its own check - '80 blocks on 110
multiprocessors, 0 token counts timed out, max relative difference 1.8e-07
against the default read' - and measurably does NOT win: story 201.4 -> 190.6
tok/s (-5.4%), code 245.6 -> 249.6 (+1.6%) over 3 interleaved pairs. That
matches the AMD note in the same file (an R9700 measured it neutral to 1%
slower), so it stays opt-in (STRATA_HC_PERSIST=1) and is not a default. The
port is kept as the starting point for a CUDA-tuned block mapping.
The router's top-10 is picked again with a logit bonus for the experts already
in VRAM (route_prior.cu); the weights stay the router's. Measured on an RTX
3060 12 GB, IQ3_XXS, 5 interleaved pairs of fresh engines: story 44.2 -> 60.9
tok/s (+37.6%), code 47.3 -> 64.0 (+35.2%); =1.0 gives +22.6% / +24.6%. Three
more pairs against the switch itself read +23.6% / +36.9%.
It changes which experts run, so it is a default only where the expert cache
cannot hold the model, and the owner approved that trade for 12 GB cards. It
is NOT a default on a bigger card: on an RTX PRO 5000 72 GB it measures
-1.9% / -0.5% and on 2x RTX A4000 -4.2% / +11.0% (3 pairs each); both were
re-run after this change and are unchanged (201.7 / 246.9 and 87.1 / 107.1
tok/s). STRATA_ROUTE_PRIOR=0 turns it off and any explicit value wins.
A draft for the owner: the one engine default that changed (the wide
hyper-connection up read on a 12.0 card, +1.9% story / +0.7% code decode on
the RTX PRO 5000, 5 pairs, bitwise), the four settings that are faster where
the PC supports them (PLE in RAM, the English draft vocabulary, --prefill-pipe
on a two-card split, q4_0 KV on 12 GB with its precision cost), and the list
of what was measured and rejected so it is not retried. No version number is
claimed: the owner decides what else 0.2.7 carries.
STRATA_GR_UP_WIDE (one warp a row, N/2 = 1,280 blocks instead of 160)
measured on the RTX PRO 5000 Blackwell 72 GB, IQ3_S, every expert in VRAM,
5 interleaved pairs of fresh engines: story 197.6 -> 201.3 tok/s (+1.9%),
code 243.5 -> 245.1 (+0.7%). A third run against the switch (default on vs
STRATA_GR_UP_WIDE=0) reads +1.7% / +0.8%. The produced token ids were
identical in all 6 pairs, and fused_gr_bench's per-T hashes are the plain
read's. The switch stays: =0 off, =1 on any card, STRATA_SM120_PERF=0 the
whole 12.0 set. Other cards are untouched.
setup writes DS4_ASYNC_T=4 into the config env for an integrated Radeon: the engine's mailbox path is sized for cards that hold
most of the experts and costs this box 31% on code (15.9-16.3 -> 20.1-21.3 tok/s), 16% on agent with identical accepted tokens,
9% on chat; the 2.25 set (no drafter) is unchanged. The same tokens come out. Discrete cards keep the engine default.
The engine's windowed decode uses the mailbox ("async") path for windows up to DS4_ASYNC_T (8 by default). That default was sized
for cards that hold most of the experts; on the 128 GB gfx1151 box the 6-row verify window is faster on the synchronous path.
Measured 2026-10-25 with the 2.00 bpw set and the shipped 0.1.51 engine (greedy, 256-token sets, three interleaved reps):
- code 15.87-16.33 -> 20.14-21.28 tok/s (+31%);
- agent, with identical accepted tokens a round (5.07): 15.93 -> 18.48 (+16%);
- chat 10.33 -> 11.30 (+9%);
- 2.25 code unchanged (11.2 -> 10.9): the drafter is unloaded there, so the extra syncs have nothing to hide.
The same tokens come out; only the scheduling changes, and the drafter's own single-token windows (T=1) stay on the async path.
dsv41_config_env(cards) writes DS4_ASYNC_T=4 into the generated config for an integrated Radeon only (a discrete card keeps the
engine default); STRATA_DSV41_ASYNC_T=n overrides the value and =8 restores the default. Tests:
tests/terminal/test_dsv41_async_t.py (5 cases). No engine rebuild; the signed dsv41 package is unchanged.
Docs: the STRIX_HALO and DEEPSEEK41 measured tables and the README row now read 19-21 tok/s code, 18 agent, 11 chat.
The trace origin was thread-local, and the prompt's later stages run on their
own thread, so their timestamps could not be compared with the first stage's.
One process-wide origin makes the hand-off visible.
Found with it on z440 (Q2_0, split 24, 16,383-token prompt, 8K chunks): stage
0 starts at 0 ms and 3,267 ms, stage 1 at 3,268 ms and 5,996 ms, read ends at
8,775 ms - the two stages do overlap chunk-for-chunk, and each takes ~2.7 s
per 8K chunk for its 24 layers. A split 30 run leaves stage 1 at 3.09 s a
chunk with 18 layers, so the per-chunk cost is not its layer count.
Same diagnostic as the request-step timestamps: with a layer split the chunk
line now says which stage is starting a chunk and when, so a prompt's stage
overlap can be read off the log. STRATA_TRACE only; no behaviour change.
Measured with it on z440 (Q2_0, split 24, 16,383-token prompt): serial reads
run stage 0's two 8,192-token chunks 3,332 ms apart and the whole read takes
8,788 ms; --prefill-pipe 1500 reads four 4,096-token chunks 1,582 ms apart and
the read takes 8,375 ms, so the stages overlap only partially at best.
The trace lines named the step a request reached but not how long it took to
get there, so a stall between two marks was invisible. Each line now carries
the time since the first traced step (thread-local). A diagnostic only: no
step, no order and no output changes without STRATA_TRACE set.
Found with it on z440 (Q2_0, layer split 24, 16,383-token prompt): the batched
read takes 8,758 ms while stage 0's own GPU timeline is 5,374 ms - stage 1 is
the slow one - and the expert-cache refill after the prompt costs a further
592 ms of the request.
Measured on the 128 GB box with the 2.00 bpw set: 4K prompt 78.7 -> 87.2 tok/s, 8K 108.9 -> 125.1, 32K 155.2 -> 176.9;
decode unchanged (16.53 -> 16.59). setup writes --prefill-chunk 8192 only for an integrated Radeon whose whole pool is at
least 64 GB; the escape hatch is STRATA_DSV41_NO_BIG_CHUNK=1. Also repairs UTF-8 damage in three docs from an earlier
PowerShell edit.
The engine's prompt-chunk rule returns 1,024 when a single card holds half or more of the experts; it was calibrated on
cards that hold most of them (4 x R9700). On the 128 GB Strix Halo (share 0.62, one shared pool) the bigger chunk wins:
measured with the 2.00 bpw set, engine 0.1.51, 2026-10-25: 4K prompt 78.7 -> 87.2 tok/s, 8K 108.9 -> 125.1, 32K 155.2 ->
176.9. Decode is unchanged (16.53 -> 16.59): the chunk buffers take ~5 GB from the GPU expert tier (81.9 -> 76.6 GB) and the
RAM store's hit rate rises (68.6% -> 76.8%) to compensate.
dsv41_prefill_chunk_args(cards, ram) writes --prefill-chunk 8192 for an integrated Radeon only, and only when the whole pool
(the RAM Windows sees plus the BIOS carve-out) is at least 64 GB; a discrete card or a smaller unified PC keeps the engine's
choice, and STRATA_DSV41_NO_BIG_CHUNK=1 turns it off. The chunk changes the prompt GEMMs' summation order, not the model: the
repo's chunk-size KL check reads 1024 0.1705 / 2048 0.1692 / 4096 0.1688 / 8192 0.1695. Test:
tests/terminal/test_dsv41_prefill_chunk.py (5 cases).
Also repairs UTF-8 damage in README.md, docs/STRIX_HALO.md and docs/MODELS.md from an earlier PowerShell edit (the middots,
em dashes and arrows had been double-encoded; no text changed). No engine rebuild is needed: the existing signed dsv41 package
accepts the flag.
The up projection of the hyper-connection read runs 160 blocks of 8 warps
(each warp takes 8 rows) on a 12.0 card, so 160 blocks of 192 SMs and ~6.7
warps an SM: the read is latency-bound at about 0.5 TB/s. This kernel takes
one row per warp and 2 columns x the 4 streams a block, so the grid is
N/2 = 1280 blocks (53 warps an SM). Every row's dot, xor tree, epilogue and
the cross-stream sum are gr_up_multi_kernel's, in its order, so the outputs
are bitwise the same; the token streams of both arms were identical in all
6 pairs measured.
Measured on l620 (RTX PRO 5000 72 GB, all experts in VRAM, IQ3_S, 3
interleaved pairs, fresh engine each): story 197.4 -> 201.3 tok/s (+1.9%),
code 243.0 -> 244.7 (+0.7%). Off by default (STRATA_GR_UP_WIDE unset); the
q8_1 tail of S26 needs the 16-column blocks, so the wide kernel is only used
when no q8_1 image is written.
The Windows AMD DeepSeek engine (feed member dsv41, windows-x86_64-hip-rdna3, gfx1151) ships with these
changes: the 0.2.6 copy-all rule is off on an integrated GPU (it copied 53 GB into RAM beside the GPU's 82 GB
in one 128 GB pool and an 8K prompt stalled; the in-place path is the 0.2.5 soak path), a Windows HIP package's
rocm/bin is on the engine's library path (setup, run config, app), the unified-memory fit plan counts the BIOS
carve-out, and --feed reaches the DeepSeek engine download. Docs, app text and tests included.
Measured on the Strix box (2.00 bpw, Faster, DSpark auto): 4K prompt 78.7 tok/s, 8K 108.9, code 17.6 tok/s,
chat 11.0. Clean-user end-to-end from a signed feed: install, download, pack, smoke, serve, chat, tool call,
agent round trip: 5/5.
strata launch --feed passed the feed to setup, but setup's parser had no --feed (getattr(a, 'feed', None)
was always None), so the flag was refused with 'Unrecognized arguments' and a local or non-default feed could
never serve the DeepSeek engine (found by the clean-user test on strix, 2026-10-25). It is now the same
option strata update has: a URL or a folder.
Found by the clean-user test on strix: 'strata launch --feed DIR' read the local feed for its own engine
check and then setup went to GitHub for the DeepSeek package, so a local (or any non-default) feed never
served the DeepSeek engine. setup's get_dsv41_engine/upgrade_dsv41_engine take the feed, install_dsv41
passes --feed, and runner.setup_args hands --feed to setup (a test in test_launch_dsv41).
The unified-memory plan counted only the RAM Windows sees (64 GB on the owner's box) and ignored the 64 GB
BIOS carve-out the GPU owns: 128 GB in one pool. uma_total_gb() adds the carve-out (or the GPU figure less
its shared part), so the plan now says about 109 GB of the 2.00 set's 132 GB stay in memory and 23 GB are read
from the SSD; the docs and the test fixtures use the 128 GB box. The 'measured' text says 128 GB unified too.
The Windows AMD DeepSeek engine package (feed member dsv41, windows-x86_64-hip-rdna3, arch gfx1151) is built from this
branch and installed as its own package; these are the code changes it needs.
- ds4_model.cu: on an integrated GPU the CPU lane's experts are no longer copied into RAM by the auto rule. The GPU's
expert tier and that copy are the same shared memory: on the 64 GB Strix Halo the GPU held 82 GB of experts, the rule
saw 62 GB 'free' and copied 53 GB, and an 8K prompt did not finish in 11 minutes. Found 2026-10-25 on strix; the
corrected build does the same prompt at 108.9 tok/s (4K 78.7), code answers 17.6 tok/s, chat 11.0. DS4_COPY_ALL=1
still forces the copy for a test.
- dsv41_engine.py: lib_dirs()/start_env() put the package's BUILD.json lib_dirs on PATH/LD_LIBRARY_PATH (a Windows HIP
package keeps hipBLAS/rocBLAS in rocm/bin, or a clean PC fails with 0xC0000135); the upgrade test-start uses it.
- setup.py: the dsv41 engine's library dirs come from the package (so the run config carries rocm/bin), the unsupported-card
text names Ryzen AI Max on Windows, and the manual package hint names the .tar.gz the release page actually has.
- runner.py: the same library dirs for the app/Terminal path.
- dsv41_fit.py: a unified-memory plan for the official IST-DAS sets (one pool, what stays in RAM, what is read from the SSD,
measured 14-16 tok/s on code, 10-11 on chat, 4K 78 / 8K 109), never a refusal; tools/strata_mcp.py uses it for the app
catalog's on-this-PC line.
- Docs: STRIX_HALO.md (DeepSeek section with speeds, driver 26.10.1, memory behaviour, limits), DEEPSEEK41.md, MODELS.md,
AI_REFERENCE.md, README.md, packaging/README.md (the Windows AMD package recipe).
- Tests: tests/terminal/test_dsv41_fit.py (7) and 4 new tests in test_dsv41_engine.py (lib dirs, feed selection for gfx1151).
The Windows commit limit decides how many experts an integrated GPU may hold (place_experts). With AMD driver 26.10.1 the limit
went from 74 GB to 133 GB on a Strix Halo, the engine put 9506 experts (86 GB) on the GPU instead of 8105, every byte of the
GPU's address space was gone, and the 32K prompt's later allocations spilled into RAM: free RAM fell to 0 and the engine died
("rocBLAS: No devices found"). The experts now leave 4 GiB + 0.2 MB a token of context above 16384 (DS4_UMA_GPU_RESERVE_GB sets
it, tests); only an integrated GPU on Windows is affected.
Strix Halo (gfx1151, 2.00 pack, shipped profile), driver 26.10.1:
32K prompt: killed (min free RAM 0) before, 113.2 tok/s with min free 13.6 GB now; 64K prompt (70000 context): 122.2 tok/s, min free 14.4 GB
4K / 8K prompt: 62.8 / 83.9 tok/s (64.9 / 86.2 with the 9506 experts: -3%), min free 13.4-13.8 GB
KL, code segment, 999 positions: 0.1020 (0.1028 before)
A smaller reserve (3 GiB + 0.1 MB a token) gained 0.0% on 4K / 8K / code decode and left 0.9 GB less free at 64K, so it was not taken.
R9700 1 card 2.00: 21.30 -> 21.82 tok/s (+2.4%), 4 cards 35.39 -> 37.00 (+4.5%), 3 pairs each, same generated ids. DS4_QUANT_SINGLE=1 / DS4_QUANT_NODEDUP=1 go back to one launch an item / no sharing.
The 512-bit i-quant rows (IQ2_XXS/XS/S, IQ3_XXS/S) assembled their grid words from scalar loads (~30 uops a 64-value chunk). Where cpu_gather_fast()
(Intel Sapphire Rapids and later) they now load them by one gather, 32-bit for IQ3 and 64-bit for IQ2 (STRATA_IQ_GATHER=1/0 force it; Zen 4 stays scalar:
its gathers are microcoded). The same words are loaded, so every output bit is unchanged. One token of an IQ3_XXS / IQ3_S expert takes the exact 512-bit rows (ggml-cpu's
single-token AVX2 dot bit for bit) with the gather, after a first-use bit check against this ggml; STRATA_NO_IQ512X1=1 turns it off.
Xeon 8468, native_gu_rows (gate+up of an expert), us: IQ3_XXS 1 token 295 -> 219, 2 tokens 462 -> 300, 4 tokens 558 -> 389; IQ3_S 415 -> 305, 603 -> 342, 562 -> 296;
IQ2_S 2 tokens 377 -> 235, 4 tokens 437 -> 327 (one IQ2_S token stays on ggml: 239 against 234). Output checksums identical with and without.
Test iq512_gather_parity (both lookup modes): exact rows == ggml vec_dot at 1-4 tokens, engine rows within float rounding.
After 'Update Strata' installed a new app version, the host that asked for it was ended by the installer, so the job stayed saved as running and the next start turned it into a red 'Interrupted operation' alert in Settings > Update and Diagnostics. The job now records the app version it started from, and a start under a different version marks it done ('Strata is updated to X').
strata update and the app's check reused the feed used last, including a loopback address or a folder from a test of the update feed; every later check on that PC then said 'No internet connection'. Only an https address is remembered now.
Without --adapt-async on the command line, the expert arena on one CUDA GPU (--serve, adaptive tier on, no --batch, no layer
split, no pipelined windows, no peer or remote tier) starts on the blocking tier, ignores 64 steady decode windows, measures 64,
and when the tier's wait is at least 5% of a window runs 64 windows on the asynchronous tier, going back (for the session) only if
that was more than 2% slower per token. One log line at each switch. --adapt-async 0|1 and STRATA_ADAPT_ASYNC=0 keep their meaning;
the identity gates run with --adapt-every 100000 and never reach it. Decision logic and its unit test are header-only
(adapt_async_default.hpp, adapt_async_default_test).
Window 2's residual stream was window 1's stale tail: stream copies and mtl_rt's deferred command buffer had no
ordering at sync points (memcpy, stream sync). Now a full barrier over all streams and the context, and the context
takes a lock (helper threads flush too). engram_dequant_rows wrapped n*18/32*hd bytes but strides rows by hd: rows
past 192 of 288 read unmapped memory (inf scales, router picked INT_MAX). The CPU lane (ggml-cpu rows) was compiled
out, so every non-resident expert contributed zero. Debug: DS4_DBG_ALL (every layer), DS4_DBG_EMB, DS4_DBG_EG.
init_pre writes T*8 floats - 32 bytes for T=4 - and its wrapper wrapped the enclosing 16 384-byte page to bind them. That aliasing
MTLBuffer overlaps the arena pages that hold h, and with it present h read back as 'valid at both ends, NaN in the middle' with
nan_range starting at exactly 4 096 - one wrapped page of floats. The work is trivial, so it now runs on the CPU through host_view:
no wrap, no alias, no dispatch.
Result with the fix: h at layer 0 is clean (zeros=40960, ninf=0, max=0) instead of ninf=4096 max=7.6e+07. The output is still
0,0,0,... so embed's write still does not reach h - but the contamination that had been confusing every probe of that buffer is
gone, and the two remaining readings (host_view vs cudaMemcpy on h between embed and layer 0) can now be compared cleanly.
Fixes this session, all measured: the 1.32 GB all-zero embedding file (my converter read the sparse trial shards), my own
20000-dispatch flush cap (it blocked every prompt of 256+ tokens), and now the init_pre page alias.
Targets, unmet: decode 8.94 tok/s vs 100, prompt 4000 tokens ~27 tok/s vs 1000, and no valid token has ever been produced.
Measured this session, in order of how much they narrow the problem:
* the pack's embedding table is real, and the engine hands embed correct rows (rows_in: first=4759 4759 4759 4759)
* ds4_embed's kernel receives the right arguments (args probe: dim=5120 hc=4)
* one Dev, one h: DBG devh prints id=0 h=0x100011a8000, the same address layer() reads
* the memory model is sound: writing through host_view round-trips through cudaMemcpy at 4 KiB and at 64 MiB (42 -> 42)
* both dispatch routes work: a constant written via Arg::B reads back 3 3 3 3, via a device pointer passed by value 7 7 7 7
* the Metal embed path is load-bearing: breaking its pipeline changes the exit code 0 -> 3
What is still wrong: h is partly NaN at layer 0 (nan_range=4096..40703, max=7.6e+07) with valid data at both ends, and
wholly NaN from layer 1, so the logits never resolve and every run prints 0,0,0,... Nine hypotheses were tested and
disproved today: the zeroed table (a real bug, fixed, not the cause), two queues, host_view vs device pointer, memory
coherence, the binding, a dead Metal branch, a wrong dim, a short write, and two Dev records.
Targets, measured, unmet: decode 8.94 tok/s vs 100; prompt 4000 tokens ~27 tok/s vs 1000. The 51.8 and 176 tok/s reported
earlier were dispatch-ENCODING rates and are withdrawn here as they were in the notes.
Next: locate the real d.h allocation (several probes landed on paths the engine does not take) and read that buffer at
three points in one run - at allocation, after embed, and at layer 0.
embed writes h = 0x144f24000
layer 0 reads h = 0x100011a8000
Both report device 0, and forcing --gpus 0 does not change it, so it is not a tensor-parallel artefact: the engine holds two Dev
records with two h allocations, embed fills one and layer() reads the other, and the bounce in ds4_model.cu (if (l.dev != cur)) is
meant to reconcile them but does not on this path.
That single pair of addresses explains every symptom: layer 0's input buffer was never written (hence zeros=40960), the first block
then produces NaN, and the logits never resolve. It also retires the last four hypotheses: the embedding table (real), the embed
kernel (verified: writes as_type<float>(0x4759 << 16)), two-queue ordering (shared Context now), and stale host reads (the probe
flushes).
Correcting the record: the 51.8 tok/s decode and 176 tok/s prefill reported earlier were dispatch-ENCODING rates and are withdrawn.
The real measured rates are 8.76 tok/s decode and ~27 tok/s at 4000 tokens, and no valid token has ever been produced. Targets remain
100 tok/s and 1000 tok/s: both unmet.
Next: print cur, l.dev and the two h addresses inside layer() - whichever differs is the buffer the embedding never reached.
Measured this stretch:
* the pack's embedding table is real (embed_pinned = 3d6b 3e6e 3e65 3dd7)
* embed itself is correct: immediately after the call h[0..3] reads 55296, which is exactly as_type<float>(0x4759 << 16),
the BF16 embedding value converted the way both ds4_embed and embed_kernel convert it
* yet at the top of layer 0 h is all zeros (zeros=40960 max=0), and block 1 turns that zero into 100% NaN
Nothing between the two probes writes h except init_pre (which writes pre). The remaining suspect is which Dev's h the layer reads:
the engine has a cross-device bounce (cudaMemcpyAsync through a host staging buffer) that ds4_metal_flush() does not cover, and a
stale l.dev under 'tensor parallel 2 -> 1' would produce exactly this.
Three hypotheses about the embedding are now retired with evidence: the zeroed table (a real bug, fixed, did not change h), two-queue
ordering between init_pre/embed and the rest (they now share one Context, no change), and the kernel writing nowhere (it writes
correctly). Targets remain unmet: decode 8.76 tok/s vs 100, prompt 4000 ~27 tok/s vs 1000, and no valid token has been produced.
State of the port, honestly: the engine runs, 40+ kernels dispatch, the pack and the 110 GB Engram are real, and it has NEVER
produced a valid token. Both targets are unmet: decode 8.76 tok/s (target 100), prompt 4000 tokens ~27 tok/s (target 1000).
What this commit adds:
* the drafter path: ds4_mx_gu/ds4_mx_dn (MXFP4 prompt gate_up/down, one runtime-ne kernel), ds4_markov_add, ds4_confidence, with
hosts and guards. --dspark on now runs end to end, but the confidence is NaN (its taps come from the trunk, which is NaN), so no
drafts are offered and there is no drafter speedup.
* token-boundary commits: seven host-read sites in ds4_model.cu flush the deferred command buffer before reading a device result.
This is what exposed that the earlier 51.8 / 176 tok/s figures were dispatch-ENCODING rates, not GPU rates: with the work inside
the timed window the real numbers are 8.76 tok/s decode and ~27 tok/s at 4K. Those earlier numbers are withdrawn.
* embed/init_pre now dispatch through the shared Context (they had their own Metal queue, so no ordering against the copies that
feed them) - measured, and it did NOT fix the blocker.
* the embedding blocker, measured at both ends: the pack's table is real (embed_pinned reads 3d6b 3e6e 3e65 3dd7), ds4_embed is
dispatched at ds4_model.cu:7490 and returns, and h[0..3] is 0 0 0 0 immediately afterwards, so the kernel writes nothing into the
buffer the engine reads. Behind it, block 1 turns a zero input into 100% NaN (a 0/0 in the first norm or router).
* env-gated debug probes (DS4_DBG_H, DS4_DBG_EMB, DS4_DBG_L) left in for the next session; they are inert without the variables.
Known edit hazard, learned three times: patching by a generic statement pattern lands in the wrong function. Patch by unique
context or by function name.
Removed the per-dispatch histogram from encode() (it cost a mutex and a string key per dispatch; measurement showed no gain, so the
~10 us a dispatch is genuine encode + engine host work). Reverted the two experiments that did not hold up: the q8_quant dedupe
(the repeated calls all differ - 0 dispatches saved) and skipping host_view() in cpu_addr (breaks the engine, 15 dispatches).
Prefill chunk sweep, 4000 tokens, pack-2.25-real: --prefill-chunk 16 -> 165.90 tok/s, 32 -> 175.89, 64 -> 176.26. The engine runs
16-token windows regardless, so the curve flattens: 176 tok/s is the 4K number and the target is 1000.
Measured state: decode 51.8 tok/s (target 100), prompt 4K 176.3 tok/s (target 1000). Neither met; the remaining levers are kernel
fusion for decode and a key-tiled attention kernel for prefill, both rewrites.
Two measurements tonight that matter:
1. Every prompt of 256 tokens or more died at exactly '20073 dispatches, 2 command buffers'. That is the deferred command buffer's 20,000
dispatch cap - my own constant from when I introduced batching - flushing mid-token. Raising it to 2,000,000 makes the long prompts run:
64 tokens 535.8 tok/s
128 tokens 556.3 tok/s
256 tokens 542.6 tok/s (31,508 dispatches, 1 command buffer)
1024 tokens 451.0 tok/s (126,020 dispatches, 1 command buffer)
4000 tokens 170.9 tok/s (492,254 dispatches, 1 command buffer)
The curve falls with length, so the 4K figure is attention scaling, not fixed overhead.
2. The per-kernel dispatch histogram now works (Pipeline carries its name; the earlier '<unnamed>' was a lookup bug):
ds4_q8_quant 4122 (36%)
ds4_iq_mv_k 1218
ds4_finish 1038
ds4_hc_pre_norm 486 / hc_split 480 / hc_mix 480 / hc_post_v 480
ds4_gemv_f32 474
~1,890 dispatches a token at 8.8 us each = the 16.7 ms a token, so decode speed is 1/(launches x 8.8us). 100 tok/s needs
~1,130 launches: dedupe q8_quant per layer (-380) and fuse the four hc_* passes (-320) gets there arithmetically.
Decode 51.8 tok/s (target 100, measured on pack-2.25-real), prompt 4K 170.9 tok/s (target 1000). Neither target met.
Measured (real 110 GB Engram pack, engine's own lines):
prompt 64 tokens -> 535.76 tok/s
prompt 128 tokens -> 556.31 tok/s
decode -> 48.3-50.0 tok/s
prompt 256+ -> aborts, router sentinel 2147483647 (no pps figure)
Targets are 100 tok/s decode and 1000 tok/s prompt: both roughly 2x away, neither met.
Added:
* sparse_attn_packed<G> (G=8/4/2/1) and the BF16 row variant, with the host wrapper that mirrors AttnRows and picks G - the 4K
abort at ds4_kernels.cu:2161 is gone.
* ds4_iq_mv widened to 16 accumulators; host chunks any job with ncol > 16 into 16-column jobs (this prompt needs ncol=96), and
Q5_K/Q6_K dense matrices get ds4_iq_mv_k instead of being fed to the IQ2-only decoder.
* route, gather_rows, combine kernels.
* Wrap extents made exact across 7 wrappers (moe_sort, swiglu, hc_split scale/base, sparse_attn rows, unpack_sel, engram_dequant);
the old 'bytes + 65536' form could overrun an arena chunk's edge.
Known bug, not fixed: for prompts of 256 tokens or more the attention SINK logits read NaN (sink[57] at the first failing window),
which poisons the softmax denominator, then q, then the router. 64 and 128 tokens are clean, so it looks like a path switch at the
256 boundary (few = T <= kAttnScratchT in sparse_attn) or an overrun into the sink buffer. That is what stands between the current
556 tok/s and a 4K figure.
Five conflicts, resolved keeping every f-024-rel behaviour (slots pin, long-prompt skip, HIPK, ZC fix) and maya2's:
- ds4_main.cpp, request struct: both fields kept (pin=N from f-024-slots and dspark_cum from maya2).
- ds4_main.cpp, request parser: both keys kept (pin, dspark_cum/spec_cum).
- ds4_main.cpp, prompt feed loop: slots' pin_at (the read stops at N for the shared-start checkpoint 3) and maya2's Stop handling (checkpoint 1 before the prompt, set_cancel, the stop check at the top of each feed) combined: the loop checks Stop first, then sizes the feed with the pin cut (n is no longer const).
- ds4_model.cu, forward chunk loop (two places): the skip's 'run.runs(k, il)' (an early chunk skipped past the yoco layer) kept, and maya2's halt_now() / halt_end(di, err) check added after it, so a skipped layer does no halt work.
win_submit (ds4_kernels.cu) is compiled to a no-op outside _WIN32 (#if defined(_WIN32) ... else (void) s), so it is already out of the Linux CUDA and Linux HIP builds.
Owner PC (RTX 5070 12 GB, Ryzen 5 7600, official 2.00 pack, shipped profile-2.00.bin, 3 alternating pairs, medians, off = every switch off):
4K prompt, off 92.64 tok/s: win_submit 99.12 (+7.0%, 3/3 pairs), flow pump 97.37 (+5.1%, 3/3), 64 Engram readers 96.64 (+4.3%, 3/3),
unbuffered Engram 85.56 (-7.6%, 0/3), all on 86.84 (-6.3%).
code decode, off 4.51: timer 4.51 (1/3), 64 readers 4.41 (2/3), unbuffered 4.67 (2/3, one pair 3.83), all 4.57.
mmap decode, off 1.95: cold-page prefetch 1.71 (-12%, 0/3).
Cold (no profile) agreed on win_submit and flow pump (+4.5%, +2.7%, 2/2) and showed the timer +12%, readers + unbuffered +22% on decode (2/2, off arm noisy).
Identity: --vram-experts 0, off vs all IDENTICAL (4K prompt, code decode); text equal to off in every warm arm.
Stop during a long prompt (H100, 8K and 16K): STOP answered in 0.12 / 0.16 s, the conversation before the prompt kept, the next request identical.
DSpark cum cut: H100 code 20.13 tok/s off vs 19.60 at 0.5, no gain.
ON: DS4_WIN_SUBMIT, DS4_FLOW_PUMP_WAIT, Stop, STRATA_DS4_LANE_FREE (a no-op on the physical-core lane).
OFF, switch turns it on: DS4_TIMER_MS=1, DS4_ENGRAM_THREADS=64, DS4_EG_DIRECT=1, DS4_IQ_PF_COLD=1, STRATA_DS4_DSPARK_CUM=0.5.
One conflict, docs/DEEPSEEK41.md (both sides edited the three-set paragraph and the file table). Resolution: f-024-rel's text kept (IQ1_M first, the model-card sentence, 533 GB for all three, the 'not a model to pick' Engram row), with the setup worker's sentence that setup recommends 1.75 where 2.00 is tight in memory or disk added to the 1.75 entry.
Context::run() committed a command buffer and blocked in waitUntilCompleted for EVERY dispatch. At ~60 dispatches a layer over 40
layers that is 3.3 us x 2400 plus a 133 us commit+wait each, and it showed up as exactly the 330 ms a token the engine reported.
Now a single dispatch is appended to a persistent command buffer and returned from without waiting; the buffer is committed when it
fills (20000 dispatches) or when the engine synchronises (strata::gpu's stream/device synchronise call ds4_metal_flush()).
Measured on the same fast harness (--vram-experts 0.05 --cpu-threads 2, ids 1,2, n=6):
before: prompt 5.43 tok/s, decode 3.00 tok/s, layers 330.1 ms/token
after: prompt 77.0 tok/s, decode 48.7 tok/s, layers 19.1 ms/token
with the experts resident in VRAM instead of streamed: decode 42.6 tok/s (layers 19.5 ms) - so residency is no longer the limit.
The remaining gap to 100 tok/s is the 1-token IQ2 matvec itself: with ncol == 1 the eight accumulators are idle and every weight is
decoded for a single dot, which is the 87.7 GB/s case measured earlier (versus 437 GB/s at an 8-token window). Fixing that is kernel
work, not runtime work.
One conflict, src/dsv41/ds4_model.cu (forward_chunks): f-024-rel has the long-prompt skip (kv_only layers), f-024-hipk added DS4_PROF stage marks around the same lines.
Resolution: the f-024-rel code kept as it is, and hipk's three prof_mark(d, 5/6/7) calls added at the same places (before the q/kv gemms, before the snapshot, after the kv_only return). They are no-ops with DS4_PROF unset.
Official harness completes with no abort: /tmp/r.out = '0,0,0,0,', zero 'no Metal dispatch', zero metal errors. Prompt 4 tokens in
5112 ms, decode 4 tokens in 18876 ms. The ids look degenerate and numerical parity is NOT established - the Engram in this pack is
zeros, and a wrong kernel would produce the same output - so what is proven is that the engine walks the entire path and emits ids,
nothing more.
Kernels added (each verified by the engine's next abort naming a later or different launch):
prompt expert GEMM: gate_up_kernel<T,NE>, down_kernel<T,NE> (ds4_iq.metal, with IQ3_XXS type 18 joining the decoders - this
pack mixes gate types 16/17 with down types 18/22)
mx_combine, engram_dequant (fmt 2), engram_gate, bf16_mv (engram wkv), compress, index_score, topk (radix with simd_ballot),
finish_kernel<true> (KV pack - the call site had no guard at all), finish rounding branch, gemv_f16_rows, block_max, hc_mean,
iota_rows, sampler_greedy (ds4_init.metal, 1521 lines)
mx_wait skipped under STRATA_METAL, like the pre-existing wait_seq no-op
Also fixed a real pre-existing bug: the 'payload + 65536 from the page base' wrap convention walked off the mapping when a small
buffer sat at an arena chunk's edge (flaky 'mtl: wrap 81920/98304 failed', exit 4, 2 runs in 4). ds4_finish.cpp now wraps exactly
the pages a buffer spans.
Not ported: sampler_kernel<0> (the sampled path; the harness is greedy), engram fmt 0/1 (this pack is fmt 2), bf16_mv's non-vector
variant. Expert types beyond 16/17/18/22 make the IQ kernels return false, so the CUDA launch names the line instead of a silent pass.
Same default as Maya #363. For the deepseek41 family only: Service.reasoning_budget() answers 32768 when the request has
no reasoning_budget_tokens / thinking_budget_tokens and the config has no reasoning_budget_tokens at all (a config value,
0 included, wins; a request's 0 turns it off). The existing wrap-up (#123) closes the thinking at the budget, and the
answer-room clamp still cuts it to leave the answer its share of max_tokens. Thinking off and other families unchanged.
Tests in serve/test_dsv41.py: the default, a request's 0 and own value, the config's value and 0, another family, and a
5000-token cap that is wrapped up and answered. serve suite: 1027 tests OK (15 skipped).
/v1/messages and /v1/chat/completions with max_tokens 200-600 and the thinking on spent all of max_tokens on the
thinking: the 0.1.45 answer room starts at max_tokens 1024 and a client that gave no budget of its own got no text block.
For the deepseek41 family only, with no budget from the client or the config, the thinking is now capped at max_tokens / 2
for max_tokens from 48 up to 1023 (the existing wrap-up, </think> and the answer follow in the other half). A client's
own reasoning_budget_tokens / budget_tokens wins, "answer_room": false turns it off, thinking off is unchanged, other
families (Qwen) are untouched. Maya #363 closes the thinking in the engine instead; the server's #123 budget already does
that for every engine, so only the short-cap gap needed closing.
Tests (serve/test_dsv41.py SmallCap, fake engine that thinks without end): max_tokens 200, 400 and 600 on both APIs return a
non-empty text block with the thinking wrapped up at half; thinking off answers directly; a client budget wins; a tool
call under a short cap still parses; another family keeps the old behaviour. serve suite: 1025 tests OK (15 skipped).
Layers 20-39 of DeepSeek-V4.1-Flash read a prompt's earlier tokens only through layer 20's compressed KV and index
keys, and each looks back one 128-token window more than the layer before. So for the answer and for the rings the
decode reads they need their own work only on the last (40 - 20) x 127 + 128 = 2,668 rows. A prompt of ~3,700 or more
new tokens on one GPU now runs its earlier chunks through layers 0-19 plus layer 20's kv_only part (the hyper-connection
pre-norm, the compressor latent, the index keys, the compressed KV) and nothing of the later layers' attention or
experts; the chunk the cut falls in runs layers 20-39 from the cut row on; later chunks run every layer. The expert flow
plans nothing for the skipped (chunk, layer)s. STRATA_DS4_PROMPT_SKIP: 1 (default) the skip, 2 only the last 128 rows
(DeepSeek's own serving; answers shift a little), 0 off. Same approach as Maya #318.
H100 emulation (64 GB cgroup, 12 GiB VRAM, store 41 GB, 3.6 GB/s SSD cap): 16K prompt 69.3 s -> 51.5 s (236 -> 318
tok/s, two alternating pairs), 4K 32.5 s -> 30.7 s; 800 teacher-forced positions after a 4,200-token skipped prompt: KL
0.00029 against the full work, top-1 against the text 88.11% against 87.98%, ppl 1.5110 against 1.5111.
--score gets DS4_SCORE_PREFIX=N (the segments joined, the first N tokens read as one prompt with the last row's logits,
the rest scored) so the skip can be scored; tools/dsv41/test_prompt_skip.py runs it against a pack when one is named.
The Maya notice file now says later Maya changes are covered by the same notice.
A prompt that does not continue the conversation in the model first sets that one aside (position, window rings,
compressor partial groups, each KV source's compressed KV and index keys, its checkpoints), then takes back the kept
conversation it does continue, in RAM or from the slot folder, also across restarts. The latest checkpoint the prompt
still agrees with is used (end of the last prompt, before it, the anchor of the first prompt of a run, and the shared
start the server marks with pin=N). The .meta format and the prefix matching are in ds4_slotmeta.hpp with a GPU-free
test (ds4_slot_test). STRATA_DS4_REUSE=0 reads every request from token 0 (reference for tests). Off unless
STRATA_DS4_SLOT_KEEP=1.
f-dsv41-rel was based on the 0.2.2-era tree (merge-base 1f727f4b5), so the merge brings it onto 7ddc17a38 without dropping 0.2.3 / 0.2.3.1.
Conflict resolutions (0.2.3.1's behaviour kept everywhere that is not DeepSeek):
- packaging/package_engine.py: build() keeps 0.2.3.1's check_reader / allow_stale_vision / allow_no_reader and adds require_dsv41; both the
picture-reader check and the strata-dsv41 presence check run; the CLI call passes both sets of arguments.
- packaging/ENGINE-COMPONENT.md: 0.2.3.1's cuda13 / cuda12 rows (vision built per variant), with strata-dsv41.exe added to the cuda13 contents.
- serve/server.py: the parser is self.parser_cls (OutputParser, or the DeepSeek DSML parser) and keeps 0.2.3.1's declared_only argument.
- strata_cli/launch/cmd.py: the --plan output prints 0.2.3.1's drafting line and then the DeepSeek placement lines.
- strata_cli/launch/runner.py: plan() returns 0.2.3.1's dict with "drafting", plus "placement" for multi-file (DeepSeek) models;
drafting_plan_line() kept; dsv41_engine() and placement() added.
ds4_bounds_scan, ds4_moe_sort_count and ds4_moe_sort_fill (ds4_kernels.cu:1938-1943) with the host wrapper that launches them
in order on one command buffer (serial dispatch, so call order is execution order). kPicks = 8: idx is [T][8], q -> (q/8,
q%8); one threadgroup a virtual slot, a block-wide prefix over 8 warps, row_tok/pos written only by the filling lanes.
One Metal fact paid for: simd_ballot returns a simd_vote that does not cast to an integer here. The counts come instead from
simd_prefix_exclusive_sum (a lane's offset inside its simdgroup) and simd_sum (the group's total) - which is a cleaner
formulation than CUDA's ballot+popc anyway. AOT: 20,160-byte AIR, rc=0.
Correction worth recording: the engine's current gate is NOT moe_sort. It still stops at ds4_kernels.cu:1915, i.e. the Engram
row decoder (engram_dequant_tok) - I attributed :1915 to moe_sort two briefs ago and that was wrong. The routing trio is ported
and wired but not yet reached; its turn comes after the Engram decode.
ds4_hc_split is the SCALAR path of ds4_kernels.cu:268: lane 0 of a 32-thread group does the whole token in a local 8x8 matrix -
sigmoid pre/post, row softmax with max shift, +eps, column normalise, then iters-1 alternating row/column normalisations. That is
the same mathematics as the CUDA warp-lane version for hc == 4, without the shuffles: correctness first, the lane-parallel form can
follow if it ever shows in a profile.
The first attempt segfaulted (rc 139) and the cause is worth remembering: scale and base were memcpy'd as if they were plain host
pointers, but every operand the engine hands a kernel is a device buffer - reach it through strata::gpu::host_view() and pass it as
a wrapped Buffer with its offset. With that fixed the engine advanced from :1908 to :1915.
Queue, nine items, each verified by the engine advancing: init_pre, embed, cublasSgemm, finish, hc_post, hc_mix_scale, hc_mix,
hc_split -> next is engram_dequant_tok (:1915), the Engram row decoder whose maths was verified standalone earlier.
Three more kernels, each confirmed the only way that matters: the engine moved past the launch line it had been stopping at.
* hc_post_inplace (ds4_kernels.cu:1677) - h[t][k][e] = post[t][k]*y[t][e] + sum_j comb[t][j][k]*h_old[t][j][e]
* hc_mix_scale (:1692) - mixes[t][i] *= rsqrt(sum(h[t])^2 / hcdim + eps)
* hc_mix (:1901) - mixes[t][m] = (h[t] . fn[m]) * rsqrt(mean(h[t]^2) + eps)
all in ds4_init.metal with host glue in ds4_finish.cpp (which is now the home of the small kernels: it owns the Context, the
pipeline cache and the wrap/offset helpers). Queue so far, seven items, every one verified by the engine advancing:
init_pre, embed, cublasSgemm, finish, hc_post, hc_mix_scale, hc_mix -> next is hc_split (:1908).
hc_split is NOT a transcription: its softmax/sinkhorn is expressed in warp-lane geometry (shfl_xor masks 1,2 for rows and 4,8
across columns over lanes 0..15, for hc == 4) and it loops 'iters' times. A wrong lane mapping yields a plausible but skewed
combination matrix, so it needs a parity test against the CUDA version on synthetic mixes before it is trusted.
ds4_finish.cpp binds the engine's x as a wrapped buffer with an offset, passes Ds4Finish as a constant struct whose pointer fields
are the engine's gpuAddresses, and dispatches one launch with grid = n_vec (the kernel's blockIdx.x is the threadgroup index).
It declines the mrope (rd > 0) and rounding (quant != 0) branches and the KV pack, so those still abort with the launch line
rather than being skipped.
Verified the only way that matters: the engine no longer stops at :1894, it stops at ds4_kernels.cu:1921 (hc_post_inplace_kernel).
Queue so far, each step confirmed by the engine advancing: init_pre, embed, cublasSgemm, finish.
Two plumbing lessons: ds4_finish.cpp had to be added to STRATA_DS4_SOURCES (strata_ds4 does not use the engine-source glob), and
scp with a relative remote path put it nowhere - absolute paths from now on.
ds4_finish (ds4_kernels.cu:1894) with its Ds4Finish struct and the block reduction: warp_sum over the 256-thread group
(simd_shuffle_xor 16,8,4,2,1 into red[8], barrier, second warp_sum, barrier), the RMS path (inv = rsqrt(tot/len + eps), then
r *= inv * w[e]) and the store (BF16 round to nearest even into dst, or back into x).
AOT gate: xcrun metal -c -> 10,032-byte AIR, rc=0. Two Metal constraints cost a round each and are worth remembering:
'long long' is not supported (use 'long'), and a kernel parameter must be a buffer/constant reference, not a plain value.
Not wired yet: the host side still has to bind the Ds4Finish struct (pointers as 64-bit gpuAddresses, strides as long), and the
mrope (rd > 0) and rounding (quant != 0) branches are not written - the host wrapper will decline those until they are, so the
engine names the line rather than skipping a step.
ds4_model.cu:2737 is the engine's first dense call: cublasSgemm(OP_T, OP_N, n_mix, T, hcd, &one, fn, hcd, h, hcd, &zero,
mixes, 32) - mixes[t][m] = h[t] . fn[m], both operands F32, alpha 1, beta 0. ds4_sgemm_tn (one thread an output element) plus
metal_host::sgemm_tn wrap the engine's buffers and dispatch it; cublas_v2.h calls it for exactly that case and keeps aborting
for everything else, so an unported shape fails loudly rather than computing something wrong.
With that the engine passes the dense product and names its next need: finish_kernel<false> (the plain variant - RMS norm,
mrope, optional in-place rounding, no KV pack) at ds4_kernels.cu:1898.
Queue so far, each step verified by the engine advancing: init_pre, embed, cublasSgemm, then finish_kernel<false>.
The DeepSeek engine now loads a real 48,786-tensor pack on the Mac, plans the staging (102.26 GB of experts in unified memory,
5,604 on the CPU lane), and executes until it needs a kernel that has no dispatch. Two are ported:
* init_pre (ds4_kernels.cu:1865) - a three-line initialiser
* embed (ds4_kernels.cu:1873) - the BF16 gather with hc copies
each as MSL in metal/src/engine/msl/ds4_init.metal plus host glue in metal/src/engine/ds4_host.cpp (wrap the engine's own device
buffer through strata::gpu::host_view + mtl::Context::wrap, then Dispatch), wired into the shared call site under
#if defined(STRATA_METAL). The engine advanced past both and now asks for finish_kernel at :1896.
Two infrastructure fixes were needed and are in here:
* runtime.cpp: StreamImpl's encoder/command-buffer state is touched from the loader and the compute path, so two encoders could be
open on one command buffer and Metal asserted ('A command encoder is already encoding'). A std::recursive_mutex around
encoder()/flush()/encode_copy()/encode_fill()/encode_dispatch() fixes it - recursive because encode_copy calls flush (a plain
mutex deadlocked, which cost one round).
* ds4_host.cpp compiles its MSL relative to the context's kernel_dir (metal/kernels), so the engine's own kernels are
../src/engine/msl/<file>.
Known: this build has STRATA_NATIVE_EXPERTS=OFF, so the engine routes all experts to VRAM ('this build has no ggml-cpu'); the
CPU-lane path is untested on the Mac.
metal/src/engine/ds4_ctx.cpp owns one mtl::Context, allocates through Context::alloc and remembers gpuAddress -> Buffer, so the
engine's device memory can live in the same residency set as the dispatches. Off unless STRATA_DS4_METAL_CTX=1, which keeps the
current behaviour and makes the two paths comparable the moment a pack exists. Build rc=0.
It exists because strata::gpu (what metal_compat's cudaMalloc maps to) is a CUDA-shaped facade with its own device state: a dsv41
launch through mtl::Context would otherwise read a foreign residency set. Flipping cudaMalloc/cudaFree to this under the env var
is the two-line change, and it is deliberately left for a session that can test it against a loaded pack.
metal/src/engine/ds4_host.cpp compiles ds4_quant.metal once through mtl::Context::compile, wraps the engine's buffers with
cx.wrap(strata::gpu::host_view(...)), builds a Dispatch (args = buffer, inline ldx, q, d; grid = n/32; tg = 32) and runs it with
cx.run(). Build rc=0 - the port's integration path is now code, not a plan.
It also exposed the last architectural question, which the comments record: the Qwen engine allocates and dispatches entirely
inside mtl::Context (cx.alloc / cx.wrap / cx.run) and uses strata::gpu only for a purge helper, so a dsv41 kernel must be given
the gpuAddress of a Buffer allocated by the SAME Context that encodes its dispatch. Until metal_compat's allocations go through
that Context, the wrapper compiles but would dispatch into a foreign residency set - the plan for that is written up in
RIGS/m1002.md rather than guessed at here.
metal/tests/dsv41/iq2s_t8_parity.mm: block 82 B (f16 d, 32 index bytes, 32 sign bytes, 8 qh, 8 scale), grid index assembled
from qs[4ib+l] plus two bits of qh[ib] (<< (8-2l) & 0x300), signs read at qs+32 in the block, scale alternating db[l/2] -
transcribed from ggml's dequantize_row_iq2_s. Verified: 65,536 rows x 5,120 k x 8-token window, rel L2 2.86e-07, PASS.
Rate 515.71 GB/s effective (64.46 GB/s of raw weight traffic) - the fastest of the three expert formats and 89 % of the
579 GB/s streaming ceiling. With IQ2_XXS (437.5) and IQ2_XS (467.8) the 89 GB of IQ2 weight in the 2.25bpw set now has
verified Metal kernels, which is 165-195 tok/s of headroom on expert traffic alone.
metal/tests/dsv41/iq2xs_t8_parity.mm: block 74 B (f16 d, 32 uint16 qs, 8 scale bytes), grid = iq2xs_grid[qs & 511],
signs = ksigns_iq2xs[qs >> 9], two scales a group (sc & 0xf, sc >> 4) - transcribed from ggml's dequantize_row_iq2_xs and
verified against the CPU reference: rel L2 2.83e-07 at 24 MB and 2.88e-07 at 97 MB, PASS.
Rate: 117.4 GB/s effective on 24 MB (launch-dominated) and 467.8 GB/s effective on 97 MB, which is in line with IQ2_XXS's
437.5. IQ2_XS is 24 GB of the 2.25bpw set and IQ2_XXS 49 GB, so the two largest expert formats are now ported.
metal/tests/dsv41/iq2xxs_t8_parity.mm decodes the weights ONCE per 8-token window and applies them to eight accumulators, which is
how the engine runs decode (DSpark's draft/verify rounds are 4-5 tokens, the prompt window is 8). Same tables, same decode as
the verified one-token kernel, and the reference checks it: relative L2 2.91e-07 on sampled rows, PASS.
The number that matters: 437.52 GB/s effective (54.69 GB/s of raw weight traffic) against 87.71 GB/s for the warp-per-row form.
At 2.33 GB of expert reads a token that is ~5.3 ms of expert time, i.e. ~188 tok/s of headroom - the first measurement that says
the 50 tok/s target is reachable, and it says the expert path must be the T-window form from the start, as the repo's moe_mm is.
metal/tests/dsv41/iq2xxs_gemv_parity.mm decodes IQ2_XXS exactly as ggml does (block 66 B, db = d*(0.5 + (aux32[1] >> 28))*0.25,
grid = iq2xxs_grid[aux8[l]], signs = ksigns_iq2xs[(aux32[1] >> 7l) & 127]) and dots it with the activations. Measured against a
CPU reference that uses the same tables: 65,536 rows x 5,120 k, relative L2 2.88e-07, PASS, 87.71 GB/s.
Two findings that shape the expert path:
* the group headers advance 4 bytes and read 8, so an 8-byte uint2 load is misaligned on odd groups. Both vector-load variants
failed (L2 1.38 and 0.985) and were slower; byte assembly is the correct access pattern, as in CUDA's own decoder.
* 87.7 GB/s is 4.8x slower than the BF16 GEMV's 425 GB/s for the same traffic, because the grid decode costs per byte. At 2.33 GB
of expert reads a token that caps decode near 38 tok/s - below the 50 target - so the expert path must be tiled (one weight read
feeding several tokens, as the repo's moe_mm does) rather than one row per warp.
That is a measured design constraint now, not an assumption.
metal/src/engine/msl/ggml_grids.h carries kmask_iq2xs[8], ksigns_iq2xs[128], iq2xxs_grid[256], iq2xs_grid[512], iq2s_grid[1024],
iq3xxs_grid[256] and iq3s_grid[512] as MSL constant arrays. metal/tools/gen_grids.sh produces them from
third_party/ggml/ggml-common.h (3cf03257f, GGML_TABLE_BEGIN/END blocks), so the Metal expert kernels use bit-identical tables
to CUDA instead of a hand copy - the same discipline that caught the half-rounded-scale error in the Q8 kernel.
Every expert format in the 2.25bpw set (IQ2_XXS 49 GB, IQ2_XS 24 GB, IQ2_S 16 GB, IQ3_XXS) is grid-based, so nothing in the
expert path could be written without these.
metal/tests/dsv41/q8_gemv_parity.mm: Q8_0 blocks (32 int8 + one f16 scale) times F32 activations, one warp a row, lane-strided
blocks. 16,384 rows x 5,120 k, relative L2 8.11e-08, PASS, 173.52 GB/s and 326.62 GFLOP/s (0.514 ms for 89 MB).
The first run measured 1.85e-04 and looked like a kernel bug; it was the CPU reference using the exact float scale while the
kernel reads the half-rounded scale from the file. Matching the rounding took it to 8e-08 - the same class of drift that would
quietly cost quality if the port did not reproduce the file's arithmetic exactly.
Three kernels are now verified with reproducible parity tests: the Q8 activation quantiser, the BF16 GEMV and this one.
metal/tests/dsv41/bf16_mv_parity.mm compiles the MSL version of the BF16 matrix-vector product (ds4_gguf.cu:1043's
mx_mv_kernel, single job: one warp a row, lane-strided k, warp reduction), dispatches it on the GPU and compares against a
double-precision reference: 2048 rows x 5120 k, relative L2 1.80e-07, PASS.
It also prints throughput: 83.19 GB/s and 83.2 GFLOP/s, 0.252 ms for 21 MB of BF16 weights. That is a naive kernel (no
threadgroup tiling) against a GPU that streams 579 GB/s in rt_bench, so 7x of headroom remains - and the number matters:
2.33 GB of expert reads a token at 83 GB/s caps decode near 36 tok/s, which is exactly why the expert path has to use the
tiled multi-token kernels (the repo's gemv/moe_mm family) rather than a warp-per-row matvec.
A second measurement lesson: per-row relative error of ~1e-3 on a 5120-term float dot is expected cancellation, not a bug;
the standard measure is relative L2 over the result, which is what both parity tests now use.
metal/src/engine/msl/ds4_quant.metal carries the Q8 quantiser that metal/tests/dsv41/q8_quant_parity.mm proved bit-exact (4096
blocks, 131,072 values, 0 differences). It is the first dsv41 kernel in the engine's own MSL directory, next to the Qwen
engine's mtp/ple kernels, so the runtime loader will find it by name when the launch is wired.
The comment records the driver quirk that cost the debugging: threadgroup_position_in_grid arrived as 0 in every threadgroup
on this Mac, so the block index comes from thread_position_in_grid / 32.
metal/tests/dsv41/q8_quant_parity.mm compiles the MSL version of q8_quant_kernel (ds4_gguf.cu:201) at run time, dispatches it
on the GPU and compares every int8 against the same arithmetic on the CPU. 4096 blocks, 131,072 values: 0 differ, worst
scale error 4.6e-04 (half rounding). PASS.
Two bugs found on the way, both in the test harness rather than the kernel: indexing with thread_position_in_grid where the
threadgroup id was meant, and passing ldx = 1 where the stride between blocks of 32 values is 32 (which made block b read a
window starting at b instead of 32b - the giveaway was block 1 reproducing block 0's values shifted by one).
This is the pattern every remaining kernel follows: an MSL function, a pipeline from it, the args as buffers, one dispatch,
and a CPU reference in the same program.
strata_ds4_launch is a macro that passes __FILE__ and __LINE__ to strata_ds4_launch_at, so the abort message reads
'src/dsv41/ds4_iq.cu:1234' rather than a generic line. Each run of the engine is then an exact work item instead of a guess
about which of the 146 launches is next, and the message is machine-readable from a log. Build rc=0, strata_ds4 = 1,412,984 B.
Two things made this possible:
* the launch syntax. 146 kernel launches (119 in the sources, 27 in the vendor tree) are now
strata_ds4_launch(kernel, grid, block, shared, stream, args...), a call a C++ compiler can parse, with a macro that forces
the global name past argument-dependent lookup and a body that stops the run naming the kernel until a Metal dispatch exists.
Every source then compiles as C++ through include/metal_compat - .cu files included, since the Metal project has no CUDA
language and was silently skipping them.
* the last CUDA names the kernel files needed: __grid_constant__, cudaDataType, __umulhi/__mulhi, the bitwise atomics,
__halves2half2 / __half2half2 / __low2half / __high2half / __low2float / __high2float / __hadd2 / __hmul2 / __hfma2 /
__hneg2, the occupancy query and the stream-capture status.
The EXL3 legacy path (v1.1.8's kernels, NVIDIA before sm_80) is defined in metal/src/engine/exl3_legacy_stub.cpp: four of its
seven entry points abort with their name, three answer as the declarations require. The 2.25bpw text model does not use EXL3.
Result: build rc=0, 0 undefined symbols, strata_ds4 = 1,410,552 bytes, Mach-O arm64, and it runs (its own usage text). Next
is the run-time loop: a pack reports the first kernel it needs, implement it in MSL, repeat.
The DeepSeek host sources now build inside the Metal project as their own executable, with the .cu files compiled as C++ through
include/metal_compat. ninja reaches step 23/35 and ds4_model.cu (8,232 lines, host only) builds; the first failure is a <<<>>>
launch in ds4_gguf.cu, which is the port's real boundary.
The measured size of the remaining work: 101 kernel launches to convert to Metal dispatches and 78 kernel bodies to write in MSL
(text only; the vision tower adds 21 and is optional). ds4_model.cu has zero of both.
Two gotchas from building it: run_locked.sh returns exit code 69 for builds that are fine (run ninja directly), and the Xcode
licence line on stderr is noise - the toolchain compiles without accepting it. Run with -j8 directly, the host sources compile in
about a minute.
A full-list compile (-ferror-limit=0) found 350 errors behind the 20-error display cap, and they collapsed into twelve missing
families. With dim3, __threadfence_system, __funnelshift_r/l, rsqrtf, __expf/__logf/__sinf/__cosf, sincosf, __ldg, __ldcv,
__float2int_* and __shfl_up_sync in cuda_device_builtins.h, plus the C math names in scope, the counts are
ds4_iq 52, ds4_gguf 42, ds4_kernels 132, ds4_vision 27 (was 66/67/185/32) AND the distinct-names list is now EMPTY: what is
left is not a missing name but device construct - the <<<grid, block>>> launch syntax and shared-memory declarations that a
C++ compiler cannot parse as written.
That is the boundary of the host-compile audit, which was its purpose: the API surface is complete, and the remaining work is
the MSL implementation itself.
Twenty rounds of 'compile ds4_iq.cu as C++ and fix the first name it cannot find' produced the complete list, and it is small:
__byte_perm and the bit/int intrinsics, the SIMD video instructions with their exact per-lane semantics (__vsub4,
__vsubss4, __vadd4, __vaddss4, __vcmpeq4, __vcmpgtu4, __dp4a), the rn float helpers, integer min/max, __align__, the
vector types beyond float2/float4, the CUDA device built-ins for the host compile (threadIdx/blockIdx/__syncthreads/__shfl_*
- labelled as host-only, nothing there is a kernel), and wmma in CUDA 12's tag style whose every call aborts naming the
simdgroup-matrix equivalent.
ds4_model.cu stays at 0 errors. The four kernel sources are at the compiler's 20-error display cap; what is left behind it
is mma_sync's signature and single scalars (__float2int_rn), not structure. Next is the MSL work itself.
ds4_numerics.cuh names __NV_E4M3/__NV_SATFINITE and __half_raw, ds4_q6k.cuh names __fmul_rn/__ushort_as_half, and both kernel
files include mma.h: cuda_bf16.h now carries the fp8 enums and includes cuda_fp16.h (for __half_raw), cuda_fp16.h carries the
five rn helpers and __ushort_as_half, and mma.h says what it is (the wmma surface does not exist on Metal; the token path
does not use it, and an instantiation fails loudly).
State after this: ds4_model.cu 0 errors; the four kernel files are at the compiler's 20-error limit with a shorter list each -
the remaining class is the device code itself (block decoders, shuffles, shared memory), which is the MSL work.
The kernel sources' single remaining error was cuda_fp16.h not existing; with it, vector_types.h and cuda_fp8.h in place they
get past the wall and show their real errors - ds4_iq 7, ds4_gguf 7, ds4_kernels 20, ds4_vision 20 - which is now the work
queue for the Metal port. ds4_model.cu stays at 0.
Both copies (this branch and the box) carry the box's version of cuda_fp16.h; an earlier local edit had failed on a path and
this commit syncs them.
The last layer of the compat surface: the CUDA function-space qualifiers (__host__/__device__/__global__/__shared__/
__constant__/__forceinline__/__launch_bounds__), cpu_set_t's CPU_ZERO/CPU_SET for the macOS branch of ds4_os.hpp,
CUBLAS_GEMM_DEFAULT_TENSOR_OP, a cudaMemcpy2D that reports failure rather than copying nothing, and a cublasGemmEx that takes
any argument list (the engine passes cuBLAS's full 19-argument spelling) and aborts naming kernels/dense_mm.metal.
With that, src/dsv41/ds4_model.cu - 8,232 lines, host only - compiles against metal/include/metal_compat on the M5 Max. The
four kernel sources are one error each, and the kernel work itself (MSL bodies for 27 + 76 + 33 device functions, plus the
dense GEMM and a pitched copy at the runtime) is what remains before a first token.
The include had landed between '#ifndef NOMINMAX' and '#include <windows.h>', so on a Mac it was never reached - that is why
the audit still showed every Linux name undeclared. Moved after '#pragma once': ds4_model.cu goes from 20 to 15 errors.
The 15 left are four CPU_ZERO/CPU_SET uses (macOS has no cpu_set_t macros) and the __host__/__device__ qualifiers that
cuda_runtime.h does not define yet; both are a few lines and are the next step, then the Linux surface of the engine's
storage helpers is done.
ds4_os_apple.h maps O_DIRECT (-> 0, F_NOCACHE does the work), posix_fadvise (-> 0), mincore (-> ENOSYS, 'unknown' is the
answer the header's contract already allows), MADV_HUGEPAGE and SYS_mbind (no-ops) and the CPU-affinity calls (no-ops on a
kernel that schedules on its own), and ds4_os.hpp includes it at the top under __APPLE__.
State: included, but ds4_model.cu still reports the same 20 errors at the same uses, so something in the include order is
not taking effect yet - the next session's first job is one -E on ds4_os.hpp to see which definitions win. Not claiming it
works.
Compiling src/dsv41/*.cu against metal/include/metal_compat as C++ is how the DeepSeek-on-Metal port is scoped, and it is
much smaller than assumed: ds4_model.cu is 8,232 lines of *host* code with zero device markers, and the six non-kernel
sources already compile. What was missing at the compat layer:
* cuda_bf16.h - __nv_bfloat16/__nv_bfloat162/__nv_fp8_storage_t and the four conversions the engine uses
(__float2bfloat16_rn, __bfloat16_as_ushort, __nv_cvt_float_to_fp8, __nv_cvt_fp8_to_halfraw), E4M3 with round-to-nearest.
* cublas_v2.h - the nine GEMM calls (7 cublasGemmEx, 2 cublasSgemm) and their enums. The dense GEMM itself is NOT wired:
it aborts with a message naming kernels/dense_mm.metal (which exists and passes its parity test) rather than returning a
silently wrong result.
After these, ds4_model.cu's remaining 20 errors are all in ds4_os.hpp - the Linux-only OS surface (O_DIRECT,
posix_fadvise, mincore, MADV_HUGEPAGE, SYS_mbind, cpu_set_t/sched_setaffinity) - which needs a __APPLE__ branch.
T tokens in tiles of 16 and rows in groups of 8 (activation slice and accumulators in L1): 7.7-9 ns per block and token on IQ2/IQ3 rows against
28-60 for ggml's dot (x3.6-6.6), IQ1_M 16 ns (x1.8); bit-identical to ggml (ds4_gg_bench --exact at 7, 8, 16, 33, 64 tokens, all six types).
gg_cpu_gate_up takes up to 64 tokens a call (and any number in pieces); ds4_gg_bench --expert times the whole expert pipeline.
strata_kernels_cpu compiled expert.cpp / q2_avx2.cpp / iq_avx*.cpp with hard-coded -mavx512f / -mavx2, so on an arm64 host the
build died with 'unsupported option -mavx512f for target arm64-apple-darwin27.2.0'. STRATA_CPU_KERNELS_X86 now decides:
on x86 the same sources and ISA flags as before, elsewhere the portable files plus src/kernels/cpu/iq256_portable.cpp, which
defines iq256_supported() = false and iq256_rows() = abort (callers check the first), so ds4_gg_cpu.cpp keeps ggml-cpu's
rows - which have their own NEON kernels on Apple Silicon.
Verified on the M5 Max: configure rc=0, build rc=0, ds4_gg_bench runs IQ3_XXS at 83.9 GB/s through ggml-cpu's rows (the
x86 EPYC does 60.5), 0 of 1,567,344 values differ in bits against ggml and 7.14e-09 against a double reference. The x86 path
is untouched: the guard only selects which sources join the library.
The layer-major chunked prompt path only ran with --threads on / --tp > 1. --prefill-chunk N (N > 0) now enables it on one card too (kChunkedPromptDefault stays false until the decode gate is signed off). A chunk's routing bumps the layers' heat and CPU-lane hits, and cache_settle() runs the swap rule in rounds after the prompt so the decode that follows finds the hot experts in VRAM (8K prompt: 1699 swaps, 5.6 s). If the chunk buffers do not fit, the chunk is halved down to 256 rows, then windows of 16 with one line saying so. ds4_main: --window 0 (default) feeds a whole prompt when the chunk path is on.
The rows are ggml-cpu's vec_dot bit for bit at any token count, so the default cannot change an answer; a one-time per-type check against the
ggml-cpu build in use turns them off if they ever differ (DS4_GG_IQ256=0 never, =1 always). EX rows: 5 and 6 tokens in one decode pass
(were 4+rest), independent integer chains for 1-2 tokens, IQ1_M: delta dot from per-call pair sums of the activation, group signs from one
shifted load. The main engine's iq256_rows hash is unchlanged (phaseA-tests/cpul/hash_iq256.cpp).
The main engine's iq256_* rows keep their float order (EX=false, the default; hash of their outputs on all types and 1..8 tokens is
identical before/after, phaseA-tests/cpul/hash_iq256.cpp). The new iq256x_rows (EX=true) sums as ggml-cpu does - block scale f16(d)*y.d, the
format constant after the sum, ggml's hsum - so a token's dot no longer depends on how many tokens share the pass (the greedy ids of a
DSpark run changed with DS4_GG_IQ256=1 because the old order differed in the last bits). IQ1_M (29), single-token in ggml, gets a
multi-token AVX2 row. ds4_gg_bench --exact: 0 differing bits for IQ1_M, IQ2_XXS/XS/S, IQ3_XXS/S at 1..8 tokens.
Conflict decisions (win side = f-dsv41-win, port side = f-dsv41-port):
- ds4_iq.cu: kept port. Its IQ1_S/IQ1_M/IQ4_NL kernels are verified bit-exact on HIP gfx1201, sm_86 and sm_120; win's
IQ1_M table load and supported() list are a subset (IQ1_S, IQ4_NL added).
- ds4_kernels.cu/.cuh: kept port's engram_dequant fmt int (0 MXFP8, 1 Q6_K, 2 IQ4_NL) and its iq4nl_value, replacing
win's q6k flag and iq4nl_value; same arithmetic.
- ds4_model.cu: kept port's eg_fmt vector and its Engram type check (unknown type exits 2 instead of segfaulting);
win's eg_q6k is gone, the duplicate IQ4_NL row branch from the textual merge is removed.
- ds4_packinfo.cpp: kept port (device types derived from iq::supported()).
- iq_expert_test.cpp: kept port (hip_or_cuda.h); win's tests/dsv41/hip_as_cuda.h removed (nothing else used it).
The IQ1_M test-d placement is the same bytes in both versions.
- CMakeLists.txt: kept port's test-language block (HIP, CUDA or CXX for the CPU-only build); win's
_strata_dsv41_test_lang variable dropped.
- pack_gguf.py: kept port's type table (IQ1_S/IQ1_M comment) and combined both missing-Engram checks
(has_part from win's split Engram layout AND port's dense-override names).
Deriving their engram.gate from their k*q and transposing their wkv scored 0.409 then 2.996 KL - both my arithmetic, not
their data (their k/q/wkv are stored transposed and the gate is not k*q in the engine's layout). --gate-kv-from <pack>
copies those two tensors from a pack that is known good, so a pack can carry their ple table and nothing else of theirs
differ. That is how the comparison the plan asks for was finally made (numbers in EACH-PC-PLAN/R9700-BOX.md, step 2):
their 110.6 GB IQ4_NL table scores KL 0.16787 / ppl 3.627 against 0.16502 / 3.588 for our 203 GB MXFP8 one - +0.003 KL,
inside this engine's run-to-run placement noise - while saving 92 GB of disk.
Also fixes the fold's reads: Gguf's tensor offsets are relative to the shard's data section (pack_gguf adds data_start),
so the fold must add it too; without that every tensor came out as NaN.
tests/dsv41/{iq,tq}_expert_test.cpp included <hip/hip_runtime.h> unconditionally, which no CUDA toolkit ships - every CUDA
box had to pass a hand-written shim on -I (found on l620 and z440, 2026-10-23, and reported as the one thing worth
upstreaming). include/strata/dsv41/hip_or_cuda.h now provides the handful of HIP names those tests use on a HIP build and
maps them onto the CUDA runtime otherwise (hipMalloc/Free/Memcpy, device sync, the error type and the events).
Verified on r9700 that the HIP build is untouched: incremental build rc=0, ds4_iq_test --type 29 ALL OK, ds4_tq_test ALL OK.
Their engram_k/q/wkv are transposed against the engine's layout (wkv [6144, 25600] against [25600, 6144], k/q [5120, 4]
against a [4, 5120] gate), so the fold now transposes and writes the engine's shape. The pack builds (their IQ4_NL table,
all files present) and runs without crashing, but it scores KL 0.40915 - and that is this tool's bug, not their data: a
direct read of their BF16 engram_k through Gguf's offsets already comes out as NaN, while the engine reads the same tensors
through a pack fine, so the fault is in how this code turns an offset into bytes (the Q5_K dequant inherits it).
Do not quote 0.409 as the IQ4_NL-vs-MXFP8 comparison; that measurement is still open. Next step is to read the tensors the
way pack_gguf writes them into tensors.txt and check one row against the engine's own dump.
Their sets keep the Engram as three per-layer tensors (blk.N.engram_k/q/wkv) while the engine wants engram.gate (k*q
folded) and engram.wkv; their engram-IQ4_NL.gguf carries only the ple table, so a pack built around their Engram stopped at
'blk.1.engram_kv.weight: missing'. engram_fold.py folds the gate and renames the kv into a dense_requant-style folder, and
pack_gguf now treats a tensor a --dense-override supplies as present (it checked the quant's shards and the Engram set only).
Result: packs/ista_2.00_native - 48784 tensors, their IQ4_NL table, Engram layers 1,14, 2896 Engram reads for a 32-token
prompt, no crash.
NOT yet a measure of their table: it scored KL 0.40915 / ppl 4.769 against 0.16502 / 3.588 for the same set with our MXFP8
table, because their engram_k/q/wkv are stored TRANSPOSED relative to the engine's layout (their engram_wkv is [6144, 25600]
against [25600, 6144]; their k/q [5120, 4] against the [4, 5120] gate). Copying the bytes as they are yields a valid pack
with the wrong data. engram_fold.py still needs the transpose (gate in float; the Q5_K kv dequantised, transposed, written
BF16) - Phase 1.1's 'load their layout as published'.
It was a hand-kept copy of the supported() switch and fell behind when IQ1_S, IQ1_M and IQ4_NL got kernels: the 2.00 set's
60 IQ1_M expert tensors (59.5 GB) were reported as 'no GPU kernel ... the CPU lane takes them' while the kernels were in the
same binary and ds4_iq_test passed them. That is exactly the line an operator trusts before a download, so it now walks
ids 0..63, asks iq::name()/iq::supported(), and adds MXFP4 and the trellis types by hand (they are not in that switch).
Verified: packs/ista_2.00bpw_9e2528f and packs/h225 both report 'no GPU kernel: none'. Also fixes the intake's eval call
(session16.sh takes the pack directory name, not a path - the 2.00 score had failed with an empty runs/s16_<path>.tsv).
Found and fixed on p620, the only box with their engram-IQ4_NL.gguf (110.6 GB). Before this, a pack built around their
file loaded for 38 s and then died in Model::Impl::eg_issue at the first token (3 of 3 runs, addr2line 0x6d3450).
- Model::Impl::load: one guard line and exit(2) for an Engram tensor whose type the reader does not handle, naming the
layer and the type (verified with a pack retyped to IQ4_XS, rc 2; a missing embed tensor also exits 2). The guard stays
for the next unknown type.
- eg_q6k becomes eg_fmt per layer: 0 MXFP8 ([scale, 32 codes] blocks, 264 B a row), 1 ggml Q6_K (210 B a 256), 2 ggml
IQ4_NL (18 B a 32: an fp16 d and 16 nibble bytes), 3 an old pack's split FP8. The load loop sniffs the type from the
pack's tensor table, eg_issue builds the row descriptor from it (144 B a row, off + id*144, no scale tensor), and
EgReader::read() needed no new branch: its contiguous, scale-free path already served Q6_K.
- ds4_kernels.cu: engram_dequant_kernel and engram_dequant_tok_kernel take the format and add the IQ4_NL branch, a literal
port of ggml's dequantize_row_iq4_nl (fp16 d per 18-byte block, low nibbles 0-15, high 16-31, kvalues_iq4nl).
- The Engram row slot stays cols*(hd+hd/32) for every format, so existing packs are byte-identical: --pack-info on the
MXFP8 h225 pack and on a Q6_K zero pack is unchanged, and a Q6_K pack's 32-token ids are identical old against new.
Verified with their real table: the pack that segfaulted runs 32 tokens in 1692-1711 ms (18.1-18.3 tok/s) and produces
coherent Python (GGML_TYPE_Q2_0 = 42, GGML_TYPE_NAMES: dict[int, str]); 64 tokens on the code prompt, DSpark off 17.16
and on 23.36 tok/s (44 of 53 drafts accepted). IQ4_NL decode is bit-exact against ggml on 8 of 8 sampled rows, and the row
addresses match the pack's tensor table (data_start 320, 144 B a row).
Not covered: no non-zero MXFP8 table existed on the box to re-check that path in-engine (the Q6_K and synthetic-zero MXFP8
packs were), there is no in-engine float dump against the Python reference (the proof is formula- and address-level), and
the legacy split-FP8 format (3) is kept but untested.
1. CMakeLists: -DSTRATA_BUILD_DSV41=ON was silently dropped when no GPU backend was on, because the whole DSV41 block
sits inside if(STRATA_ENABLE_CUDA OR STRATA_ENABLE_HIP). Found on p1004 (2026-10-23), whose P100 is not a DeepSeek
target but which is exactly the box the CPU-lane A/B belongs on. A new branch builds ds4_gg_bench with no GPU
runtime at all. Verified on r9700 with both backends off: configure rc=0, build rc=0, the benchmark runs IQ3_XXS at
60.5 GB/s (ggml-cpu) against 64.1 (AVX2 rows, x1.06) and 8.02e-09 against a double reference.
2. tools/dsv41/ista_intake.py: the pack step now uses the recipe that actually builds a loadable pack for these sets -
--engram-meta (the UD set's first shard carries the Engram hash parameters, without it pack_gguf sees no Engram layers
and refuses per_layer_token_embd), --mmproj, the combined --skip-tensors regex, and the DSpark noise token and markov
rank (pack_gguf fills the rest of the geometry from the drafter but not those two, so a pack came out with
dspark_noise_token_id 0). Verified end to end on the 2.25 set: the intake rebuilds packs/h225's content exactly -
same 48794 tensor lines, --pack-info 'files: 10, all present', 'no GPU kernel: none', dspark geometry 5/128799/256/128/3.
Found and implemented on z440's CUDA bring-up, applied here and re-verified on HIP.
ds4_iq.cu: decode<T>/deq8<T> for kIQ1_S, kIQ4_NL, kIQ1_M, plus iq1_codes() (iq1s_grid_gpu's 4-bit nibbles -> the int8 codes
8*(n-1)+-1 through two __vsub4) and iq1m_d() (IQ1_M carries its F16 d in the top nibbles of its four u16 scales - there is no d
field). kQK<T> is 32 for IQ4_NL as well as MXFP4; load_tab uses iq1s_grid_gpu / kvalues_iq4nl; DS4_IQ_SWITCH and
supported() cover the three types, so they leave the CPU lane and join the device path.
tests/dsv41/iq_expert_test.cpp: make_mat now writes IQ1_M's scale where the format keeps it (bytes 49/51/53/55, high nibbles).
Without that the derived d is random and roughly 1/32 of the blocks come out NaN, which std::max then hides.
Parity (z440, one A4000, ds4_iq_test --type 19,20,29): token gate_up 6.70e-08 / 2.04e-08 / 5.97e-08, token down 3.62e-08 /
4.71e-08 / 3.71e-08 against ggml's to_float and a double reference; CPU lane 3.4e-8..8.2e-8; F16 prompt tiles inside their
own tolerances. Bandwidth: IQ4_NL 376 GB/s (84% of the A4000's 448, bandwidth-bound), IQ1_S 214, IQ1_M 202 (both
compute-bound at 0.92-1.10e12 weight-values/s against IQ2_XXS's 1.22e12).
Not covered: no end-to-end run with a pack that uses these expert types (none on z440); a combined-seed run put IQ1_M's
light-path gate_up at 1.10e-2 against its 1e-2 tolerance, which is the int8 activation rounding of that path (it does not
move with tensor cores off), not the dequant.
1. ds4_tq.cu:262 used the two-argument __shfl_xor, which is HIP-only; the rest of the tree already uses
__shfl_xor_sync(0xffffffffu, v, lane), which both backends have.
2. tests/dsv41/{tq,iq}_expert_test.cpp were forced to LANGUAGE HIP, so a CUDA configure failed at generate time
('The language HIP was requested for compilation but was not enabled'). The language now follows the enabled
backend: HIP, CUDA, else CXX.
Verified on r9700 that the HIP build is untouched: incremental rebuild rc=0, ds4_iq_test --type 17 ALL OK, the engine
--pack-info clean. Verified on z440 (CUDA 12.9, sm_86) by the bring-up agent: HEAD builds, ds4_iq_test ALL OK for
IQ2_XXS/XXS/IQ2_XS/IQ2_S/IQ3_XXS/IQ3_S/IQ4_XS at 5e-8 against ggml, 272-369 GB/s on one A4000. IQ1_M, IQ1_S and
IQ4_NL still have no CUDA kernel (CPU lane only) - that is the Phase 5 kernel list.
The release runbook promises the pack/requant builder checks free space before it writes (DEEPSEEK-RELEASE.MD 9.6 item 5);
it did not. --dry-run now prints 'space: N GB planned, M GB free on <fs>' and returns 3 when 98% of the free space does
not cover the plan, so 'requant_dry && requant' cannot start a run that will fill the disk; a real run raises instead.
With --resume only the bytes still missing are counted against the free space.
Verified on the box with two recipes that do not fit /data today: rq3 (224.3 GB planned against 128.6 GB free) and rq_gu3
(251.2 GB against 128.6 GB) both refuse with the numbers printed, rc 3 in dry-run. pack_gguf's own output is a few MB
(tensors.txt, engram.bin, tokenizer) and needs no guard.
The rule from DEEPSEEK-RELEASE.MD 6.5.2 - every ggml type works, a tensor whose type has no GPU kernel stays on the CPU
lane with one warning instead of stopping the load - is now implemented for IQ1_S (19), IQ1_M (29) and IQ4_NL (20):
- iq::known(t) (a layout exists: block_values/block_bytes/name) is separated from iq::supported(t) (a device kernel
exists). type_row_bytes/type_block use known(), so a pack lays out such a tensor instead of refusing it;
gg::type_on_device(t) is what the VRAM tier tests.
- the CPU lane takes them: gg_cpu_type() covers 10-14, 16-23 and 29 (ggml-cpu's row dots).
- a layer whose experts are such a type is flagged cpu_only at load, printed once ('no GPU kernel for it, its experts
stay on the CPU lane') and excluded from the VRAM budget in place_experts().
- iq::block_values(IQ4_NL) = 32 (QK4_NL): the only block here that is not 256 values. Without it the CPU lane read
IQ4_NL at the wrong row stride (2.4e-02 and 1.4e+05 error against the reference).
Verified: ds4_iq_test --type 29,19,20 runs the CPU lane against ggml's own decode and a double reference - IQ1_M
1.6e-08 / 4.6e-08, IQ1_S 5.4e-08 / 4.0e-08, IQ4_NL 3.1e-08 / 3.2e-08, ALL OK; the device types still pass; --pack-info
reports the layer and role for each such tensor. A full model with real IQ1_M weights has not been run (none on the
box): the arithmetic of the lane is what is verified.
Today a pack built by another tool had dspark_block_size 0: DSpark ran one token a round and was worth +0.8% instead of
+28-77%. The report now shows block_size / noise_token_id / markov_rank / the drafter's expert counts, the mtp tensor
count, and the Engram layer ids, with a <- marker when the block size is 0 or there is no drafter. Verified on h225,
c0q3hQ6_K, c0q2d5 and mxfp4 (all: block 5, noise 128799, rank 256, 1242 mtp tensors, engram layers 1,14).
The pack is pointers: ds4.txt + tensors.txt + the weights where model_dir says. Today /home/niko/dsv41-c0/t3 was empty
and the c0q3 pack (all IQ3_XXS) described itself perfectly while being unable to load. --pack-info now stats every file,
compares the length with the largest tensor end, prints MISSING/SHORT lines and exits 2 when any is bad; a healthy pack
prints 'files: N, all present and long enough'. Verified: h225 10/10, c0q3hQ6_K 19/19, c0q2d5 12/12, c0q3 fails with
six MISSING lines.
The IST-DAS 2.25bpw GGUF differs from ours in five ways that stopped pack_gguf, each found by building its pack:
- geometry(): ALIASES for the keys they spell differently (kv_source_layer_ids, index_source_layer_ids, the indexer's
candidate_*, dspark_target_layer_ids, engram_layer_ids). Without the source ids the engine refused the pack
('layer 2: compress ratio 2 without a KV source of that ratio before it').
- --engram-meta FILE: their GGUF carries no Engram hash parameters (no ple.*/engram.layers, no token map, no
multipliers/primes/offsets - those are tensors of integer type in their file); the parameters come from a GGUF of the
same model that still has them, the tables from --engram.
- integer tensors (I8/I16/I32/I64/F64) are left out with a line instead of stopping the build (engram_token_map.weight,
engram_primes.weight, ...).
- --skip-tensors REGEX: their attn_output_a is one interleaved 3-D tensor ([8,1024,4096] ne) that this engine's
o_groups slices cannot take; our Q5_K requant of the same BF16 weights supplies it through --dense-override, which
now also *adds* a tensor the quant leaves out (it could only replace before).
- the DSpark geometry (dspark_block_size / n_routed_experts / num_experts_per_tok / target_layer_ids / markov_rank) is
filled from the drafter's dflash.* keys and the markov tensor's shape, and --dspark-noise-token sets the model's noise
token (128799). Without it d_block stayed 0 and DSpark ran one token a round: h225 measured 16.77 tok/s off vs
25.93 on (3.16 tokens a round) after the fix, 1.01 before.
strata-dsv41 --pack-info --pack <dir> reads only ds4.txt and tensors.txt: no device, no weights, no lock. It prints the
layer x role type map with an expert's bytes and a token's MB (top-k picks), the type totals, all-experts GB at the mean
bpw, the dense side split into embed / head / mtp / vision / engram, and every tensor whose type has no device kernel
(IQ1_S, IQ1_M, IQ4_NL today) with the layer and role - the H1 rule's report, and the first line of every benchmark log.
Verified on c0q3 (208.1 GB at 3.06 bpw, 3.251 GB a token - the document's figures) and c0q3d5.
- ds4_iq.cu: IQ2_XS (type 17) through the same kernels as IQ2_XXS: Ty<17> (74 B, 1024-word grid),
decode<17> and deq8<17>, block_bytes/name/dispatch. ds4_iq_test compares the token path, the window
path, the prompt sets and the CPU lane against ggml's own to_float + a double reference: worst
4.6e-08 (token gate_up), 4.6e-08 (down), 8.9e-03 (F16 tiles), CPU lane 3.5e-08. IQ2_XXS / IQ2_S /
IQ3_XXS / IQ3_S / IQ4_XS pass the same test unchanged.
- ds4_model.cu: gg_type_of takes 17; embed.weight may be any ggml block type (Q5_K / Q4_K / Q6_K ...)
through gg_row_to_bf16 (ggml's to_float, the same BF16 rounding as the Q8_0 path) - the IST-DAS sets'
Q5_K token_embd needs no dense override now.
- gg_dot_bench: the A/B's pass criterion was |x-y|/|x| on single outputs, which fails on cancellation
while the two paths agree to a few ulps; it now also compares 64 sampled rows against a double
reference scaled by the L1 norm (worst 1.1e-08) and requires |x-y| <= 1e-6 of the output scale.
A/B at 32 threads: single token x1.03 (IQ2_XXS) / x1.13 (IQ2_S); 4 rows x2.4 / x2.3.
- CMake: ds4_iq_test target; ds4_tq_test and ds4_tq_test_portable link ds4_gg_cpu.cpp (the ggml CPU lane
ds4_iq_cpu.cpp refers to since 49e82263a, missing from those targets).
- setup.py: the nine pinned shards of unsloth/DeepSeek-V4.1-Flash-1017-GGUF UD-Q8_K_XL (revision f924b94f, sizes and SHA-256 from the reference run; the DSpark drafter and picture reader pinned but not fetched), MULTI_FILE_FAMILIES, install_dsv41: plain warnings for the 508 GB download, RAM and disk (a refusal only where the disk cannot hold the files), the strata-dsv41 check before any download, --gguf-dir for files you already have, pack_gguf.py, config with --pack --max-context 8192 --eos 1 --tp 2 and the cards; menu entry and a --check line.
- strata_control/dsv41_fit.py: where the model goes on the cards and in RAM, from the measured run (6,240 experts in graphics memory predicted on 4 x 32 GB, 6,197 measured), the n-gram table's SSD note.
- tools/strata_mcp.py and strata_cli/launch: the model is in the list; strata launch --plan prints the placement (layer blocks per card, experts on the cards and on the CPU, RAM, disk, the SSD note); --model-path takes the nine shards where they are; parallel, pictures and tuning are said, not passed to an engine that does not take them.
- tests: tools/test_setup_dsv41.py (22), tests/terminal/test_launch_dsv41.py (13).
- tools/strata_tokenizer.py: the deepseek-v3 pre-tokenizer (llama.cpp's DEEPSEEK3_LLM list as a sequence of splits) chosen by the pack's tokenizer.json 'pre'; special tokens decode to their literal text (the DSML token is not in the byte alphabet). 20 of 20 reference prompts tokenize to the ids llama.cpp made.
- serve/dsml.py: DSML calls (the GGUF template's spelling) translated as they stream into the form OutputParser reads, so tool streaming, tool_choice and calls in the thinking work as for the other families; string=true values stay strings.
- serve/server.py: Service.set_family('deepseek41'): end token from the pack, DsmlOutputParser, forced-call opening in DSML, effort words low/medium/high/max/1..100, model name and card sampling.
- serve/test_dsv41.py (39 tests, fake engine) and fixtures from the GGUF's own template and the llama.cpp reference prompts.
Maya's tools/ds4/chat_template.jinja lost tool associations and later system instructions; the GGUF's embedded
template matched DeepSeek's encoding.py byte for byte on 9 cases (notes-dsv41-ref.md).
Co-authored-by: mw00 <[email protected]>
CMake option STRATA_BUILD_DSV41 (default OFF, adds the target strata-dsv41 and a private static kernel lib; no
existing target changes). The vendored exl3 / sampler / window-grouping kernels live in namespace strata::dsv41k so
no symbol is shared with strata_kernels. Missing CUDA names for HIP come from a force-included private header
(include/strata/dsv41/hip_extra.h), the IQ CPU lane stays the stub branch, MAYA_* env aliases are gone.
New --score IDS.json --score-out RUN.tsv: teacher-forced top-64 log-probabilities in tools/quant_quality.py's
row format (the gate against the llama.cpp reference). MIT notice added to every ported file.
Co-authored-by: mw00 <[email protected]>
Co-authored-by: needmorevram <[email protected]>
Verbatim copy of mw00/project-maya src/ds4, tests/ds4_*, tools/ds4 and the exl3 / sampler
kernel sources it needs, MIT, with their LICENSE and a NOTICE. Not built yet; the next
commits adapt it to Strata's tree.
Co-authored-by: needmorevram <[email protected]>
tools/trellis/tqenc.{hpp,cpp}: shift-trellis Viterbi (group-min ACS), tail-biting by a rotated free pass or the exact
search over all start groups, fp16 scale refinement, importance weights. tqenc_test.cpp checks the streams, the
Viterbi against a naive DP, planted solutions, bit-exact decode against tq_states.hpp and the CPU kernels.
The CPU decoder gets K = 5 and 4 (ids 224, 225, experimental v1); existing types and layouts are unchanged.
2026-10-20 14:36:38 +02:00
929 changed files with 137908 additions and 12498 deletions
description:"In the app: Settings > About. In the Terminal: `strata --version`. From source: the About tab of the web page, or `engine` in `GET /v1/status`, or the git commit."
Strata runs the Qwen3.8-Flash-Next mixture-of-experts model (and its Coder, Swift 1.5, Unsloth, Gyro and NVFP4
variants, and Qwen3.6-35B-A3B) on a normal PC: one NVIDIA or AMD graphics card plus system RAM on Windows or Linux,
or an Apple Silicon Mac. Version 0.2.3, engine 0.1.48. It has three parts:
Version 0.2.8, engine 0.1.54. The public model families are DeepSeekV4.1-Flash-1017,
Qwen3.8-Flash-Next and Qwen3.6-35B-A3B. DeepSeek V4.1-Flash runs in the app and in the Terminal
(`strata launch --family deepseek41 --model ist-das-2.00bpw`; sets in
[docs/MODELS.md](docs/MODELS.md#deepseek-v41-flash)). Read **[docs/AI_REFERENCE.md](docs/AI_REFERENCE.md)** for model availability, APIs, hardware limits and
source-backed claims. Articles and videos should use **[docs/MEDIA_KIT.md](docs/MEDIA_KIT.md)**.
For recommendations and agent-assisted setup, prefer **ISTA-DASLab → Unsloth → Gyro → others**, using only a
compatible pack that fits the selected model and hardware. ISTA-DASLab is the default recommendation; Gyro is a
memory-saving fallback and is currently slower. Honor an explicit user choice. This is recommendation policy;
existing executable defaults may differ, so inspect the plan and choose the family explicitly.
Strata has three parts:
- **The app** (`apps/desktop`, the Tauri shell; `apps/host`, the Node host and the web UI). It is for people who do
not want a terminal. The installer has the engine inside.
not want a terminal. One installer per system ([latest release](https://github.com/Niko1221/Strata/releases/latest)):
`--output json`. `strata launch --status` shows what is serving, `strata launch --stop` stops it, and
`--speed faster|full` (DeepSeek), `--output json`. `strata launch --status` shows what is serving, `strata launch --stop` stops it, and
`strata launch --show-key` prints the API key.
4. Connect the user's apps: `strata connect` lists them, `strata connect <app> --check` sends one prompt and one tool
call the way that app does.
Keep the server on `127.0.0.1` unless the user asks for more. Beyond it an API key is recommended: `--access lan` makes one for you, and `--no-api-key` serves without one (a warning, never a refusal; on a public tunnel the warning is stronger). The old route (clone
the repository and run `START-HERE.bat --yes` or `./setup.sh --yes`) still works and is the fallback when the
Terminal cannot run. Strata's MCP server ([docs/MCP_SERVER.md](docs/MCP_SERVER.md)) offers the same steps as tools.
Keep the server on `127.0.0.1` unless the user asks for more. Beyond it an API key is recommended: `--access lan` makes one for you, and `--no-api-key` serves without one (a warning, never a refusal; on a public tunnel the warning is stronger). Source setup (`START-HERE.bat`, `setup.sh`) is the fallback for a developer or for a system with no release package; it is
documented in [docs/SOURCE_INSTALL.md](docs/SOURCE_INSTALL.md). Those wrappers exist in a source checkout, not in the
released app. Do not tell an app user to find a `.bat` file.
Source setup does not make an unsupported model or backend compatible. Strata's MCP server ([docs/MCP_SERVER.md](docs/MCP_SERVER.md)) offers the same steps as tools.
# The AVX kernels exist for x86 only. On aarch64 (DGX Spark / GB10) portable.cpp stands in for them and
# native experts run on ggml-cpu's NEON dot products.
if(NOTCMAKE_SYSTEM_PROCESSORORCMAKE_SYSTEM_PROCESSORMATCHES"^(x86_64|AMD64|amd64|x64|X64|i[3-6]86|x86|X86)$")# an unset processor stays the x86 build, as before
set(STRATA_CPU_X86ON)
else()
set(STRATA_CPU_X86OFF)
endif()
if(STRATA_ISA_FLOORANDNOTSTRATA_CPU_X86)
message(WARNING"STRATA_ISA_FLOOR is an x86 older-CPU build and does not apply to ${CMAKE_SYSTEM_PROCESSOR}; ignored")
set(STRATA_ISA_FLOOR""CACHESTRING""FORCE)
endif()
set(CMAKE_CXX_STANDARD20)
set(CMAKE_CXX_STANDARD_REQUIREDON)
set(CMAKE_CXX_EXTENSIONSOFF)
@@ -81,6 +92,31 @@ if(STRATA_HIP_GFX906)
endif()
option(STRATA_ENABLE_HIP"Build the HIP targets for wave32 AMD GPUs (gfx1100, gfx1201; gfx1012/gfx1030 unvalidated)"OFF)
option(STRATA_PREFILL_MMQ"Build the optional GGML MMQ prefill path for HIP"OFF)
# The rows one verify / batch window holds (include/strata/kernels/window_limits.hpp): 8, the 0.2.7 window, or 16 (four pipelined
# groups of 16 slots = 64 slots on a four-card split, docs/BATCHING.md). Only the CUDA and HIP engines are built from
# these sources with a wide window; the SYCL and Metal builds never read this variable and keep 8. A window of up to
# 8 rows runs the kernels it always ran, whichever is chosen (the wide instantiations are extra code). "auto" = 16 for
# the HIP engines (a HIP code object loads on first use, so the extra kernels cost nothing until a wide window forms)
# and 8 for CUDA, which loads every kernel at start (CUDA_MODULE_LOADING=EAGER: the wide kernels would take VRAM from
# the expert cache of every card); -DSTRATA_WINDOW_MAX_T=16 builds the wide CUDA engine.
set(STRATA_WINDOW_MAX_T"auto"CACHESTRING"rows of a verify / batch window: 8, 16 or auto (HIP 16, CUDA 8)")
<p align="center"><a href="https://github.com/Niko1221/Strata/releases/download/v0.1.10/Pagoda.mp4"><img src="docs/media/pagoda-preview.webp" width="720" alt="A voxel pagoda garden that Strata's model wrote, running in the browser"></a><br>
<sub>Ein Voxel-Pagodengarten aus einem einzigen Prompt, auf einer RTX 5070 mit Strata (IQ3_S, 128K Kontext) ·
<a href="https://github.com/Niko1221/Strata/releases/download/v0.1.10/Pagoda.mp4">ganzes Video (49 s)</a></sub></p>
<a href="https://buymeacoffee.com/strataengine"><img src="https://img.shields.io/badge/Buy_me_a_coffee-FFDD00?logo=buymeacoffee&logoColor=000" alt="Strata auf Buy Me a Coffee unterstützen"></a>
</p>
Strata lässt **[Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)** laufen, ein Modell mit 125
Milliarden Parametern, für das man sonst einen Server braucht, auf einem normalen PC. Es chattet, schreibt Code, liest
Bilder und arbeitet mit deinen Apps und Coding-Agents zusammen. Deine Chats und Dateien bleiben auf deinem PC.
**Neu hier?** Sieh dir an, [was du brauchst](#was-du-brauchst), und [installiere dann die App](#option-a-die-strata-app).
**Kostenlos, Open Source und ganz in deiner Hand.** Nutze Strata auf dem Desktop oder mit deinen liebsten Coding-Tools. Kein Konto nötig.
- **Läuft auf einem Gaming-PC.** 32 GB Arbeitsspeicher (RAM) und Grafikspeicher zusammen reichen für den Anfang.
- **Passt zu deinen Werkzeugen.** Strata spricht die OpenAI- und die Anthropic-API, deshalb verbinden sich Claude Code, Codex CLI, OpenCode und andere Agents damit.
- **Liest Bilder.** Das ist standardmäßig eingeschaltet.
- **Kostenlos.** MIT-Lizenz, kein Konto, keine Cloud.
- **Privat.** Keine Telemetrie. Was Strata herunterlädt, wann es das Internet nutzt und wie es Downloads prüft, steht in [docs/SAFETY.md](docs/SAFETY.md).
## Modelle
Version 0.2.3 · Engine 0.1.48
Wir empfehlen **zuerst ISTA-DASLab**, danach **Unsloth** und andere Anbieter, soweit ein passendes Modell in deinen Speicher passt.
In der **Desktop-App** und im **Terminal**. Was nicht in den Arbeitsspeicher passt, wird von deiner SSD gelesen. Darum brauchst du viel freien Speicherplatz.
| Set (IST-DAS) | Download | Freier Speicherplatz nötig |
| --- | ---: | ---: |
| 1.75 bpw | 239 GB | etwa 245 GB |
| 2.00 bpw | 257 GB | etwa 265 GB |
| 2.25 bpw | 274 GB | etwa 280 GB |
Läuft auf NVIDIA RTX 30, 40 und 50 (Windows und Linux) und auf AMD Radeon AI PRO R9700 (Linux). Das Setup lädt die Dateien und die DeepSeek-Engine herunter und startet dann das Modell.
| | NVIDIA, ältere Karte oder Treiber älter als 580 | [`Strata-0.2.3-windows-cuda12.exe`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-windows-cuda12.exe) |
| | AMD Radeon RX 9000 Serie oder Radeon AI PRO R9700 | [`Strata-0.2.3-windows-hip.exe`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-windows-hip.exe) |
| | AMD Radeon RX 7000 Serie oder Ryzen AI Max (Strix Halo) | [`Strata-0.2.3-windows-hip-rdna3.exe`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-windows-hip-rdna3.exe) |
| | NVIDIA, ältere Karte oder Treiber älter als 580 | [`Strata-0.2.3-linux-nvidia-cuda12.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-nvidia-cuda12.AppImage) |
| | AMD Radeon RX 9000 Serie oder Radeon AI PRO R9700 | [`Strata-0.2.3-linux-amd-rx9070.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-amd-rx9070.AppImage) |
| | AMD Radeon RX 7000 Serie | [`Strata-0.2.3-linux-amd-rx7900.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-amd-rx7900.AppImage) |
| | AMD Ryzen AI Max (Strix Halo) | [`Strata-0.2.3-linux-amd-strixhalo.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-amd-strixhalo.AppImage) |
Installiere die App, wähle ein Modell und leg mit dem Chat los. Es gibt einen Installer pro System. Beim ersten Start lädt die App die Engine für deine Grafikkarte herunter. [Hilfe zum Installer](docs/INSTALL.md#the-app).
**Wenn du unsicher bist:** Bei NVIDIA nimmst du `cuda13`, wenn dein Treiber Version 580 oder neuer ist, sonst `cuda12`.
`cuda12` ist für ältere Karten und ältere Treiber. Die Dateien `hip` und `amd` sind für AMD-Karten. Für andere AMD-Karten
(RX 6000) gibt es noch keine fertige App: siehe [AMD-Karten](docs/AMD_HIP.md).
Das Skript prüft deinen PC, lädt das Terminal herunter, kontrolliert die SHA-256-Prüfsumme und entpackt es in deinen
Home-Ordner. Es braucht keine Administratorrechte. Öffne unter Windows danach ein neues Terminalfenster. Dann startest du
Strata:
Öffne unter Windows danach ein neues Terminal. Dann:
```sh
strata launch
```
Es stellt dir eine Frage nach der anderen: Modell, Kontext, Bilder, wer sich verbinden darf. Danach lädt es die Engine für
deine Grafikkarte herunter (signiert und vor der Nutzung probeweise gestartet), dazu das Modell, und beginnt zu laufen.
Für Skripte und KI-Agents gibt es Flags:
```sh
strata launch --model q2_0 --yes --no-tui
```
Wenn du das Installationsskript vor dem Ausführen lesen willst, lade es zuerst mit `curl -fsSL <address> -o install.sh`
herunter, lies es und führe dann `sh install.sh` aus. Alle Bildschirme und Flags: [docs/TERMINAL_LAUNCH.md](docs/TERMINAL_LAUNCH.md).
### Option C: deine KI richtet es ein
Füge das in Claude Code, Cursor, Codex, GitHub Copilot oder einen anderen Coding-Assistenten ein:
```text
Set up Strata on this PC for me: https://github.com/Niko1221/Strata - follow docs/AI_SETUP.md in that repository.
```
Der Assistent prüft deine Grafikkarte, deinen Arbeitsspeicher und den Speicherplatz, wählt das passende Modell, installiert
und startet es und sagt dir, wie du deine Apps verbindest. KI-Werkzeuge können Strata auch über den
[MCP-Server](docs/MCP_SERVER.md) installieren, starten und stoppen.
Du willst es lieber selbst bauen? Siehe [Aus dem Quellcode bauen](#aus-dem-quellcode-bauen).
## Modelle
Die App und `strata launch` empfehlen ein Modell passend zu deinem Speicher. Es ist ein einziges Modell in mehreren
Größen, mehr oder weniger stark komprimiert. Kleinere Größen sind schneller. Größere Größen sind etwas klüger.
### Wenig Speicher? Probier Gyro
[Gyro](https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF) (von agentionai) ist die kleinste Möglichkeit, dieses Modell auszuführen.
| Größe | Benötigter Speicher (RAM + VRAM) | Download | Was es ist |
| --- | ---: | ---: | --- |
| **Gyro-S** | 27,6 GB | 58,5 GB | Für PCs, auf denen nichts anderes läuft. Die Qualität ist geringer als bei IQ2_XS. |
| **Gyro-M** | 35 GB | 92,0 GB | Ungefähr die Qualität von IQ3_XXS bei etwa 20 % weniger GPU-Speicher. Es schreibt Antworten langsamer als Q2_0: 16 % auf einer RTX 3060 mit 12 GB, 21 % auf einer R9700. |
Gyro läuft auf NVIDIA- und AMD-Karten (bei AMD experimentell) auf Intel-Arc-Karten unter Linux (getestet auf einer Arc A750) und ab 0.2.2 auf Macs mit Apple Silicon (Gyro-S auf einem Mac mini mit 32 GB: 9 Tokens/s, die Experten werden von der SSD gelesen). Details und Messwerte:
[docs/GYRO.md](docs/GYRO.md). Starte es mit `strata launch --model gyro-s`.
### Die anderen Größen
| Dein Speicher (RAM + VRAM) | Nimm | Hinweise |
| --- | --- | --- |
| **32 GB** | **Gyro-S**, oder den **[Coder](docs/MODELS.md#coder)** für Code | Der Coder braucht 32 GB RAM und behält die Hälfte der Experten, deshalb ist er außerhalb von Code schwächer |
| **40 GB** | **[Swift 1.5](docs/MODELS.md#swift-15)** IQ2_XS (39,2 GB), oder **Gyro-M** (35 GB) | Swift denkt kürzer nach, bevor es antwortet. Gyro-M kommt der Qualität von IQ3_XXS nahe, schreibt Antworten aber 16-21 % langsamer als Q2_0 |
| **64 GB** | **IQ3_S** (54,8 GB), das Original oder Swift 1.5 | Am langsamsten, am besten. Es erreicht in den veröffentlichten Tests das volle Modell |
| **96 GB oder mehr** | **UD-IQ4_XS** von Unsloth (etwa 72 GB RAM) bis **Q8_0** (etwa 136 GB), 4 bis 8 Bit | Näher am vollen Modell. Bei weniger RAM liest Strata einen Teil von der SSD, das ist viel langsamer |
Auch mit weniger RAM, als eine Zeile nennt, kann es funktionieren: Strata wechselt von selbst in seinen
[Modus für wenig RAM](docs/MODELS.md#a-big-graphics-card-and-little-ram). Rechne mit weniger Tempo, mit einer kleinen
Grafikkarte mit deutlich weniger.
Weitere Auswahl:
- **[Swift 1.5](docs/MODELS.md#swift-15)** ist eine Feinabstimmung, die kürzer nachdenkt, bevor sie antwortet, bei etwa gleicher Qualität und gleichem RAM-Bedarf.
- **[NVFP4](docs/NVFP4.md)** ist NVIDIAs 4-Bit-Version für NVIDIA-Karten. Sie braucht etwa 80 GB RAM, um alle Experten im RAM zu halten.
- **[Qwen3.6-35B-A3B](docs/QWEN36.md)** ist ein kleineres, anderes Modell, das Bilder liest und mehrere Anfragen gleichzeitig beantworten kann.
- **[OrcaRouters Uncensored IQ3_XXS](docs/MODELS.md#orcarouter-uncensored-iq3_xxs)** richtest du von Hand ein.
Alle Größen, Downloads und was wo läuft: [docs/MODELS.md](docs/MODELS.md).
## Wie schnell ist es?
Gemessen mit einer Grafikkarte mit 12 GB (RTX 5070, Ryzen 5 7600, 64 GB RAM). Ergebnisse anderer Karten findest du hier:
[Tempo jeder Größe](docs/MODELS.md#how-fast-is-each-size), [Ergebnisse der Community](docs/COMMUNITY_BENCHMARKS.md).
| Modell | Antworten erscheinen mit | Liest deinen Prompt mit |
Die Balken haben in beiden Tabellen denselben Maßstab. Eine Karte mit mehr VRAM ist schneller. Eine RTX 3090 (24 GB) sollte
etwa 100-140 Tokens pro Sekunde schreiben. Die Engine-Version hinter jeder Zahl steht in [DETAILS.md](docs/DETAILS.md#speed-measured).
## Benutzung
### In der App
Die App hat einen **Chat**, einen **Monitor** für das Modell sowie für GPU, CPU und RAM, eine Seite **Modelle** und die
**Einstellungen**. Der Chat kann in einem Ordner auf deinem PC arbeiten. Im Modus **Manuell** fragt er nach, bevor er eine
Datei ändert oder einen Befehl ausführt. Mit `/loop 10m <what to do>` wiederholst du eine Nachricht alle 10 Minuten, solange
der Chat geöffnet ist.
Der Chat kann das Web mit DuckDuckGo (oder Bing) durchsuchen, ohne Konto. Unter **Einstellungen, Tools** kannst du einen
Suchschlüssel hinzufügen. Die Suche ist die einzige Funktion, die etwas von deinem PC sendet (die Suchbegriffe).
Um zwei Modelle auszuführen, schalte **Mehrere Modelle ausführen** ein und gib jedem Modell eine eigene Karte desselben
Herstellers.
<p align="center"><img src="docs/media/app-chat.png" width="900" alt="The Strata app: a chat where the agent built a small voxel pagoda garden, with the page open in the preview on the right"><br>
<sub>Ein Chat in der Strata-App: Der Agent hat einen kleinen Voxel-Pagodengarten als eine HTML-Datei geschrieben, und die Seite öffnet sich in der Vorschau neben dem Gespräch</sub></p>
### Im Terminal
<p align="center"><img src="docs/media/terminal-monitor.png" width="900" alt="The Strata Terminal monitor overview while the model writes a reply"><br>
<sub><code>strata monitor</code>, während das Modell antwortet: Geschwindigkeit, GPU, Speicher und die letzten Anfragen</sub></p>
strata update --check # is there a new engine? (strata update installs it)
strata doctor # check the installed engine files
```
`strata --help` listet alles auf. Starten und Bereitstellen laden nichts herunter. Nur die Installation, `strata launch`
(wenn du ein Modell wählst) und `strata update` tun das. Um Strata von deinem Handy oder einem anderen PC aus zu erreichen,
nutze `strata launch --access lan`. Der Befehl erzeugt einen API-Schlüssel und startet in einem Netzwerk nicht ohne einen.
Ein Browser öffnet sich nur, wenn du `--open` angibst. Anleitung: [docs/TERMINAL_LAUNCH.md](docs/TERMINAL_LAUNCH.md). Der ganze Client: [docs/TERMINAL.md](docs/TERMINAL.md).
### Mit KI-Agents
Ein Agent installiert Strata mit dem Einzeiler aus Option B und führt
`strata launch --model NAME --yes --no-tui` aus. Die Schritte, Flags und Exit-Codes stehen in [docs/AI_SETUP.md](docs/AI_SETUP.md),
die Regeln für das Repository in [AGENTS.md](AGENTS.md).
### Andere Apps verbinden
| App-Typ | Basis-URL |
| --- | --- |
| OpenAI-kompatibel (die meisten Apps und Agents) | `http://127.0.0.1:8080/v1` (beliebiger API-Schlüssel, beliebiger Modellname) |
| Anthropic-API (Claude Code und andere) | `http://127.0.0.1:8080`, ohne `/v1`. Diese Apps hängen `/v1/messages` selbst an. Claude Code: `ANTHROPIC_BASE_URL=http://127.0.0.1:8080` |
| OpenAI Responses API (Codex CLI) | `http://127.0.0.1:8080/v1/responses` ([Einrichtung](docs/DETAILS.md#the-responses-api-and-codex-cli)) |
`strata connect` gibt die genauen Einstellungen für OpenCode, Codex CLI, Claude Code, oh-my-pi, pi, Hermes Agent und OpenClaw aus.
`strata connect opencode --write` schreibt die Datei nach einer Rückfrage und legt eine Sicherung an.
### Bilder
Bilder sind für Modelle, die sie lesen können, standardmäßig eingeschaltet. Hänge ein Bild im Chat an oder füge es ein, oder
sende es aus deiner App. Der Bildleser läuft auf dem Prozessor, wenn deine Grafikkarte weniger als 24 GB hat (einige Sekunden
pro Bild, ohne Speicher der Karte zu belegen), und ab 24 GB auf der Grafikkarte. Die Regel gilt für NVIDIA und AMD gleich.
Du kannst Bilder pro Modell ausschalten oder den Leser mit `--vision off`, `--vision gpu` oder `--vision cpu` wählen.
Die Zahlen hinter der Regel: Auf einer Radeon AI PRO R9700 brauchte der Bildleser der Karte 57 ms für ein Bild mit 260 Tokens,
der des Prozessors 3,1 s. Auf Karten mit 16 GB (Tesla P100 und RTX A4000, Linux) kostete der Bildleser der Karte 3-5 % des
Texttempos, weil er etwa 1,4 GB belegt, die sonst das Modell nutzen würde. Auf dem Mac liest die GPU, egal wie viel Speicher er hat (0,45 s für ein Bild mit 260 Tokens auf einem M4 mit 32 GB, etwa 1,2 GB des gemeinsamen
Speichers); der Prozessor braucht dort etwa 50 s pro Bild und liest deshalb nur mit `--vision cpu`.
### Gut zu wissen
- Das Denkniveau (**aus, niedrig, mittel oder hoch**) stellst du im Chat-Menü oder unter "reasoning effort" in deiner App ein. Aus ist am schnellsten.
- Strata beantwortet standardmäßig eine Anfrage nach der anderen. Für mehrere gleichzeitig nutze `--parallel N` in `strata launch` oder
`"parallel": 2` in der Konfiguration ([BATCHING.md](docs/BATCHING.md)). Auf einer Karte mit 12 GB wird dadurch jede Antwort langsamer.
- Strata liest die erste Nachricht eines Chats komplett, etwa 1 Minute pro 30.000 Tokens. Folgenachrichten starten in Sekunden.
- Wo deine Chats gespeichert werden und die ganze API: [docs/INSTALL.md](docs/INSTALL.md#where-things-are-stored), [docs/DETAILS.md](docs/DETAILS.md#using-it).
## Aus dem Quellcode bauen
<details>
<summary>Windows, Linux und Mac aus dem Repository</summary>
[Lade Strata herunter](https://github.com/Niko1221/Strata/archive/refs/heads/main.zip) und entpacke es, oder nutze `git clone`.
**Windows:** Doppelklick auf `START-HERE.bat`. **Linux:** Führe `./setup.sh` im Strata-Ordner aus.
**Mac (Apple Silicon, macOS 15 oder neuer):** Installiere einmal die Command Line Tools von Apple mit `xcode-select --install`
(Xcode selbst brauchst du nicht). Öffne das Terminal im Strata-Ordner und führe `./setup.sh` aus. Das Skript baut die
Mac-Engine auf deinem Mac (etwa 7 Minuten), stellt dieselben Fragen, lädt das Modell herunter und startet es. Beim nächsten
Mal startet `./setup.sh` sofort. Welche Größe zu deinem Mac passt: [docs/MAC.md](docs/MAC.md).
Der Installer erkennt deine Karte und richtet die passende Engine ein. Er fragt, welches Modell und welche Größe du willst,
wie viel Kontext (wie viel Text sich das Modell merkt) und ob es Bilder lesen soll. Drück Enter für die empfohlene Antwort.
Dann lädt er das Modell herunter (etwa 70 GB) und startet es. Wenn der Download abbricht, führe das Skript erneut aus, dann
geht es weiter. Dein Browser öffnet die Strata-Seite unter `http://127.0.0.1:8080`.
> Während das Modell startet, kann dein PC 1-3 Minuten langsam sein oder nicht reagieren (beim ersten Mal am längsten). Strata
> lädt 35-55 GB in deinen RAM und sperrt einen Teil davon für die Grafikkarte. Warte ab und schließe das Fenster nicht.
Starte beim nächsten Mal wieder `START-HERE.bat` (oder `./setup.sh`). Es startet sofort und lädt nichts doppelt herunter.
Schließ sein Fenster, um das Modell zu beenden. `UPDATE.bat` (`./update.sh`) aktualisiert Strata, ohne es zu starten.
Aktualisieren, Docker, mehrere Karten, wohin die Dateien kommen und alle Optionen: [docs/INSTALL.md](docs/INSTALL.md).
Um deinen PC zu messen, führe `BENCH.bat` (`./bench.sh`) aus. Das dauert 5-10 Minuten und schreibt einen Bericht, den du teilen
**Optional: mehr Tempo auf einer Karte, wenn das ganze Modell in den RAM passt.** Zwei Einstellungen zum Einschalten von
RubenTorney (#1904, #1905). Sie helfen, wenn du eine Grafikkarte nutzt und dein RAM alle Experten aufnimmt, zum Beispiel bei
Q2_0 oder IQ3_XXS auf einem PC mit 64 GB. Öffne `strata-<model>.json` im Strata-Datenordner ([wo](docs/INSTALL.md#where-things-are-stored)) und ergänze:
-`"--adapt-async", "1"` in der Liste `"args"`. Der Expertencache wird dann in einem Hilfsthread aktualisiert, statt die
Antwort anzuhalten. Gemessen auf einer RTX 5070 (Windows, Q2_0): +4-5 % beim Decodieren, +7-10 % bei Code und langen
Dokumenten. Die Antworten können sich leicht von der Standardeinstellung unterscheiden und bleiben von Lauf zu Lauf gleich.
-`"STRATA_PIN_BY_NEED": "1"` im Block `"env"` (Windows). Windows kann nur etwa die Hälfte des RAM für die Grafikkarte sperren,
und diese Einstellung sperrt zuerst die Experten, die die Karte am dringendsten braucht. Gemessen auf der RTX 5070: +4,4 % beim Lesen des Prompts.
```json
"args":[...,"--adapt-async","1"],
"env":{"STRATA_PIN_BY_NEED":"1"}
```
Starte Strata nach der Änderung neu. Zum Rückgängigmachen entfernst du die beiden Einträge.
Folge den Anweisungen, um ein Modell herunterzuladen und zu starten. [Terminal-Anleitung](docs/TERMINAL_LAUNCH.md).
</details>
## Wie funktioniert das?
<p align="center">
<img src="docs/media/terminal-monitor.png" width="900" alt="Strata-Terminal-App mit Modellaktivität, Tempo, Speicher und letzten Anfragen">
<br><sub>Behalte Modell und Hardware im Blick, während es läuft.</sub>
</p>
Modelle wie dieses laufen normalerweise auf Servern mit Hunderten Gigabyte Grafikspeicher. Deine Grafikkarte hat 12-24 GB.
Strata bringt das Modell auf deinen PC, indem es die Arbeit auf den ganzen Rechner verteilt. Stell dir eine Bibliothek vor.
<a id="build-from-source"></a>
- Das Modell ist ein Team aus 24.576 kleinen Spezialisten ("Experten"), wie 24.576 Nachschlagewerke. Für jedes Wort braucht es nur 10 davon.
- **Deine Grafikkarte ist der Schreibtisch.** Er ist klein und schnell. Die etwa 4.000 am häufigsten genutzten Bücher liegen
dort aufgeschlagen.
- **Dein RAM sind die Regale im Lesesaal.** Alle 24.576 Bücher stehen dort. Braucht ein Wort ein Buch, das nicht auf dem
Schreibtisch liegt, holt dein Prozessor es sofort aus dem Regal, während die Karte weiterarbeitet.
- **Deine SSD ist das Archiv.** Sie enthält eine Nachschlagetabelle von 29 GB. Das Modell liest pro Wort nur wenige Zeilen
daraus, deshalb muss sie nicht im RAM liegen.
[Aus dem Quellcode bauen](docs/SOURCE_INSTALL.md) · [Docker](docs/DOCKER.md) · [Vollständige Installationsanleitung](docs/INSTALL.md)
<p align="center"><img src="docs/media/how-it-works.svg" width="860" alt="The model's 24,576 experts: the busiest on the graphics card, all of them in RAM, a lookup table on the SSD"></p>
<a id="what-you-need"></a>
Ein kleiner Helfer rät außerdem die nächsten Wörter. Das große Modell prüft sie alle auf einmal und behält die richtigen. Du
bekommst dieselbe Antwort, 1,6-1,8-mal früher. Lange Texte werden in großen Stücken gelesen, bis zu 8.192 Tokens auf einmal,
mit über 1.000 Tokens pro Sekunde.
## Hardware
<p align="center"><img src="docs/media/guess-and-check.svg" width="860" alt="A small helper guesses the next words; the big model checks them all at once and keeps the right ones"></p>
| **Apple Silicon · M1 oder neuer** | [Mac-Setup](docs/MAC.md) |
| **AMD Ryzen AI Max · Strix Halo** | [Strix-Halo-Setup](docs/STRIX_HALO.md) |
Die ausführliche Erklärung: [docs/HOW_IT_WORKS.md](docs/HOW_IT_WORKS.md). Jeder Teil mit seinen Zahlen:
[die Details](docs/DETAILS.md#how-it-works) und das [Paper](docs/paper/Strata-Paper.pdf).
<a id="using-it"></a>
## Etwas ist schiefgegangen?
## Tools
- **Mein PC ist beim ersten Start von Strata eingefroren.** Strata lädt das Modell in den RAM. Warte ab und schließe das
Fenster nicht. Nach 10 Minuten immer noch eingefroren? Starte den PC neu, schließe andere Programme und versuch es noch
einmal, oder wähle eine kleinere Größe.
- **Es ist beim Herunterladen oder Installieren stehen geblieben.** Starte die App oder `strata launch` noch einmal, oder
führe `START-HERE.bat` (oder `./setup.sh`) erneut aus. Es macht dort weiter, wo es aufgehört hat.
- **Es ist sehr langsam, die Festplattenanzeige blinkt ständig, oder es meldet "the engine stopped unexpectedly".** Dein PC
hat nicht genug freien RAM. Schließe andere Programme (Browser brauchen viel) oder wähle eine kleinere Größe wie Q2_0 oder IQ2_XS.
- **Es meldet, dass es keine Engine für diesen PC gibt.** Öffne in der App **Einstellungen, Engine** und drücke **Engine
aktualisieren**. Führe im Terminal `strata update` aus. `strata doctor` zeigt, welche Engine-Dateien installiert sind.
- **Es meldet, dass Port 8080 schon belegt ist.** Strata läuft bereits. Such das Fenster.
**Claude Code · Codex CLI · OpenCode · oh-my-pi · pi.dev · Hermes Agent · OpenClaw**, und weitere über die OpenAI- und Anthropic-kompatiblen APIs.
Weitere Probleme und Lösungen: [docs/TROUBLESHOOTING.md](docs/TROUBLESHOOTING.md). Du kommst nicht weiter? Öffne ein
[Issue](https://github.com/Niko1221/Strata/issues). Hänge in der App die Datei aus **Einstellungen, Engine, Diagnose, Protokolle anzeigen, Protokolle speichern** an. Füge im Terminal die Ausgabe von `strata doctor` ein. Hänge beim Start aus dem Quellcode die
Datei `strata-<model>.log` aus dem Strata-Ordner an. Ein Sicherheitsproblem meldest du bitte privat: [SECURITY.md](SECURITY.md).
Führe **`strata connect`** aus, um die Einrichtung für dein Tool zu sehen. [Verbindungsanleitung](docs/TERMINAL_LAUNCH.md#connect-your-apps-strata-connect).
## Danksagung und Lizenz
<a id="guides"></a>
<a id="how-fast-is-it"></a>
<a id="how-does-it-work"></a>
<a id="something-went-wrong"></a>
Das Modell ist [Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) vom Qwen-Team. Komprimiert wurde es von
Strata ist kostenlos und steht unter der MIT-Lizenz. Ich baue es und teste es auf jeder Hardware, die ich bekommen kann. Wenn es auf deinem PC gut läuft, hilft mir ein Kaffee,
weiter neue Karten zu testen und das zu beheben, was ihr meldet. Es besteht kein Druck. Strata zu nutzen und zu melden, was dir auffällt, hilft genauso viel.
**Hilf mit, lokale KI für alle offen zu halten.** Ein Kaffee hilft, Entwicklung und Tests auf mehr Hardware zu finanzieren.
<a href="https://buymeacoffee.com/strataengine"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="50"></a>
<a href="https://buymeacoffee.com/strataengine"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me a Coffee: Strata unterstützen" height="48"></a>
Du möchtest Strata sponsern? Schreib mir über [GitHub](https://github.com/Niko1221).
[MIT-Lizenz](LICENSE) · [Danksagung und Modelllizenzen](docs/HOW_IT_WORKS.md#credits) · [Datenschutz](docs/SAFETY.md)
<p align="center"><a href="https://github.com/Niko1221/Strata/releases/download/v0.1.10/Pagoda.mp4"><img src="docs/media/pagoda-preview.webp" width="720" alt="A voxel pagoda garden that Strata's model wrote, running in the browser"></a><br>
<sub>Un jardín con una pagoda de vóxeles, hecho con un solo prompt en una RTX 5070 con Strata (IQ3_S, contexto de 128K) ·
<a href="https://buymeacoffee.com/strataengine"><img src="https://img.shields.io/badge/Buy_me_a_coffee-FFDD00?logo=buymeacoffee&logoColor=000" alt="Apoya Strata en Buy Me a Coffee"></a>
</p>
Strata ejecuta **[Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)**, un modelo de 125 mil millones
de parámetros que normalmente necesita un servidor, en un PC normal. Conversa, escribe código, lee imágenes y funciona
con tus aplicaciones y tus agentes de programación. Tus chats y tus archivos se quedan en tu PC.
**¿Eres nuevo?** Mira [lo que necesitas](#lo-que-necesitas) y luego [instala la aplicación](#opción-a-la-aplicación-strata).
**Gratis, de código abierto y tuyo para ejecutar.** Usa Strata en tu escritorio o con tus herramientas de programación favoritas. No necesitas cuenta.
- **Funciona en un PC gaming.** Con 32 GB entre RAM y memoria gráfica ya puedes empezar.
- **Funciona con tus herramientas.** Habla las API de OpenAI y de Anthropic, así que Claude Code, Codex CLI, OpenCode y otros agentes se conectan a él.
- **Lee imágenes.** Viene activado por defecto.
- **Gratis.** Licencia MIT, sin cuenta, sin nube.
- **Privado.** Sin telemetría. Lo que Strata descarga, cuándo usa internet y cómo comprueba las descargas está en [docs/SAFETY.md](docs/SAFETY.md).
## Modelos
Versión 0.2.3 · motor 0.1.48
Recomendamos **primero ISTA-DASLab**, después **Unsloth** y otros proveedores, siempre que haya un modelo compatible que quepa en tu memoria.
En la **aplicación de escritorio** y en el **Terminal**. Lo que no cabe en la memoria se lee desde tu SSD, así que hace falta mucho espacio libre en disco.
| Conjunto (IST-DAS) | Descarga | Espacio libre necesario |
| --- | ---: | ---: |
| 1.75 bpw | 239 GB | unos 245 GB |
| 2.00 bpw | 257 GB | unos 265 GB |
| 2.25 bpw | 274 GB | unos 280 GB |
Funciona en NVIDIA RTX 30, 40 y 50 (Windows y Linux) y en AMD Radeon AI PRO R9700 (Linux). La instalación descarga los archivos y el motor de DeepSeek, y después inicia el modelo.
| | NVIDIA, tarjeta más antigua o controlador anterior a 580 | [`Strata-0.2.3-windows-cuda12.exe`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-windows-cuda12.exe) |
| | AMD Radeon serie RX 9000 o Radeon AI PRO R9700 | [`Strata-0.2.3-windows-hip.exe`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-windows-hip.exe) |
| | AMD Radeon serie RX 7000 o Ryzen AI Max (Strix Halo) | [`Strata-0.2.3-windows-hip-rdna3.exe`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-windows-hip-rdna3.exe) |
| | NVIDIA, tarjeta más antigua o controlador anterior a 580 | [`Strata-0.2.3-linux-nvidia-cuda12.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-nvidia-cuda12.AppImage) |
| | AMD Radeon serie RX 9000 o Radeon AI PRO R9700 | [`Strata-0.2.3-linux-amd-rx9070.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-amd-rx9070.AppImage) |
| | AMD Radeon serie RX 7000 | [`Strata-0.2.3-linux-amd-rx7900.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-amd-rx7900.AppImage) |
| | AMD Ryzen AI Max (Strix Halo) | [`Strata-0.2.3-linux-amd-strixhalo.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-amd-strixhalo.AppImage) |
Instala la aplicación, elige un modelo y empieza a chatear. Hay un instalador por sistema. En el primer arranque, la aplicación descarga el motor para tu tarjeta gráfica. [Ayuda con el instalador](docs/INSTALL.md#the-app).
**Si no estás seguro:** con NVIDIA, elige `cuda13` si la versión de tu controlador es la 580 o posterior, y `cuda12` en caso
contrario. `cuda12` es para tarjetas y controladores más antiguos. Los archivos `hip` y `amd` son para tarjetas AMD. Las
demás tarjetas AMD (RX 6000) todavía no tienen una aplicación ya preparada: mira
<img src="docs/media/app-chat.png" width="900" alt="Aplicación de escritorio de Strata con un chat y una vista previa del código en vivo, una junto a otra">
<br><sub>Chat, archivos y vista previa en vivo en una sola ventana.</sub>
</p>
- **Windows:** ejecuta el `.exe`. Se instala para tu usuario y no necesita permisos de administrador. Si Windows muestra
"Windows protegió tu PC", elige **Más información** y luego **Ejecutar de todas formas** (el instalador todavía no está
firmado).
- **Linux:** haz el archivo ejecutable (`chmod +x`) y lánzalo. Necesita FUSE 2: `sudo apt install libfuse2` (Ubuntu 22.04)
o `libfuse2t64` (Ubuntu 24.04 y posteriores).
- **Mac:** abre el `.dmg` y arrastra Strata a Aplicaciones.
<a id="option-b-the-terminal-one-line"></a>
En la primera pantalla, la aplicación analiza tu PC y recomienda un modelo adecuado. Pulsa Instalar. El modelo es una
descarga grande (unos 40 a 70 GB) y, si se interrumpe, continúa donde se quedó. Al cerrar la ventana, Strata y el motor se
detienen.
### Aplicación de Terminal
### Opción B: la Terminal (una línea)
¿Prefieres la terminal? Instala Strata y ejecuta **`strata launch`**.
La Terminal es el comando `strata`, pensado para scripts, servidores y agentes de IA.
El script analiza tu PC, descarga la Terminal, verifica su SHA-256 y la descomprime en tu carpeta personal. No necesita
permisos de administrador. En Windows, abre una ventana de terminal nueva después. Luego arranca Strata:
En Windows, abre una terminal nueva después. Luego:
```sh
strata launch
```
Te hace una pregunta cada vez: modelo, contexto, imágenes, quién puede conectarse. Después descarga el motor para tu tarjeta
gráfica (firmado y arrancado como prueba antes de usarlo) y el modelo, y empieza a servir. Para scripts y agentes de IA, usa
opciones:
```sh
strata launch --model q2_0 --yes --no-tui
```
Si quieres leer el script de instalación antes de ejecutarlo, descárgalo primero con `curl -fsSL <address> -o install.sh`,
léelo y luego ejecuta `sh install.sh`. Todas las pantallas y opciones: [docs/TERMINAL_LAUNCH.md](docs/TERMINAL_LAUNCH.md).
### Opción C: que lo instale tu IA
Pega esto en Claude Code, Cursor, Codex, GitHub Copilot u otro asistente de programación:
```text
Set up Strata on this PC for me: https://github.com/Niko1221/Strata - follow docs/AI_SETUP.md in that repository.
```
El asistente revisa tu tarjeta gráfica, tu RAM y tu disco, elige el modelo adecuado, lo instala y lo arranca, y te explica
cómo conectar tus aplicaciones. Las herramientas de IA también pueden instalar, arrancar y detener Strata mediante su
[servidor MCP](docs/MCP_SERVER.md).
¿Prefieres compilarlo tú? Mira [Compilar desde el código fuente](#compilar-desde-el-código-fuente).
## Modelos
La aplicación y `strata launch` recomiendan un modelo según tu memoria. Es un solo modelo en varios tamaños, más o menos
comprimido. Los tamaños pequeños son más rápidos. Los grandes son algo más inteligentes.
### ¿Poca memoria? Prueba Gyro
[Gyro](https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF) (de agentionai) es la forma más pequeña de ejecutar este modelo.
| Tamaño | Memoria que necesita (RAM + VRAM) | Descarga | Qué es |
| --- | ---: | ---: | --- |
| **Gyro-S** | 27,6 GB | 58,5 GB | Para PC donde no cabe nada más. La calidad es menor que la de IQ2_XS. |
| **Gyro-M** | 35 GB | 92,0 GB | Calidad parecida a IQ3_XXS con un 20 % menos de memoria de GPU. Escribe las respuestas un 16 % más despacio que Q2_0 en una RTX 3060 de 12 GB y un 21 % más despacio en una R9700. |
Gyro funciona en tarjetas NVIDIA y AMD (en AMD es experimental) en tarjetas Intel Arc con Linux (probado en una Arc A750) y, desde 0.2.2, en Mac con Apple Silicon (Gyro-S en un Mac mini de 32 GB: 9 tokens/s, con los expertos leídos del SSD). Detalles y mediciones:
[docs/GYRO.md](docs/GYRO.md). Arráncalo con `strata launch --model gyro-s`.
### Los demás tamaños
| Tu memoria (RAM + VRAM) | Elige | Notas |
| --- | --- | --- |
| **32 GB** | **Gyro-S**, o el **[Coder](docs/MODELS.md#coder)** para código | El Coder necesita 32 GB de RAM y conserva la mitad de los expertos, así que es más flojo fuera del código |
| **40 GB** | **[Swift 1.5](docs/MODELS.md#swift-15)** IQ2_XS (39,2 GB), o **Gyro-M** (35 GB) | Swift piensa menos tiempo antes de responder. Gyro-M se acerca a la calidad de IQ3_XXS, pero escribe las respuestas un 16-21 % más despacio que Q2_0 |
| **48 GB** | **IQ3_XXS** (47,0 GB) | Más lento, mejor |
| **64 GB** | **IQ3_S** (54,8 GB), el original o Swift 1.5 | El más lento, el mejor. Iguala al modelo completo en las pruebas publicadas |
| **96 GB o más** | **UD-IQ4_XS** de Unsloth (unos 72 GB de RAM) hasta **Q8_0** (unos 136 GB), de 4 a 8 bits | Más cerca del modelo completo. Con menos RAM, Strata lee una parte desde el SSD, que es mucho más lento |
Con menos RAM de la que indica una fila también puede funcionar: Strata pasa solo a su
[modo de poca RAM](docs/MODELS.md#a-big-graphics-card-and-little-ram). Espera que sea más lento, y mucho más lento con una
tarjeta gráfica pequeña.
Más opciones:
- **[Swift 1.5](docs/MODELS.md#swift-15)** es un ajuste fino que piensa menos tiempo antes de responder, con una calidad y un uso de RAM casi iguales.
- **[NVFP4](docs/NVFP4.md)** es la versión de 4 bits de NVIDIA para tarjetas NVIDIA. Necesita unos 80 GB de RAM para tener todos los expertos en RAM.
- **[Qwen3.6-35B-A3B](docs/QWEN36.md)** es un modelo más pequeño y distinto que lee imágenes y puede atender varias peticiones a la vez.
- **[El Uncensored IQ3_XXS de OrcaRouter](docs/MODELS.md#orcarouter-uncensored-iq3_xxs)** se configura a mano.
Todos los tamaños, descargas y lo que cabe en cada equipo: [docs/MODELS.md](docs/MODELS.md).
## ¿Qué velocidad tiene?
Medido con una tarjeta gráfica de 12 GB (RTX 5070, Ryzen 5 7600, 64 GB de RAM). Resultados con otras tarjetas:
[velocidad de cada tamaño](docs/MODELS.md#how-fast-is-each-size), [resultados de la comunidad](docs/COMMUNITY_BENCHMARKS.md).
| Modelo | Las respuestas aparecen a | Lee tu prompt a |
Las barras usan la misma escala en las dos tablas. Una tarjeta con más VRAM es más rápida. Una RTX 3090 (24 GB) debería
escribir unos 100-140 tokens por segundo. La versión del motor detrás de cada cifra está en [DETAILS.md](docs/DETAILS.md#speed-measured).
## Uso
### En la aplicación
La aplicación tiene un **Chat**, un **Monitor** del modelo y de tu GPU, CPU y RAM, una página de **Modelos** y los
**Ajustes**. El chat puede trabajar en una carpeta de tu PC. En modo **Manual** pregunta antes de cambiar un archivo o
ejecutar un comando. Escribe `/loop 10m <what to do>` para repetir un mensaje cada 10 minutos mientras el chat esté abierto.
El chat puede buscar en la web con DuckDuckGo (o Bing), sin cuenta. Puedes añadir una clave de búsqueda en **Ajustes,
Herramientas**. La búsqueda es la única función que envía algo fuera de tu PC (las palabras que buscas).
Para ejecutar dos modelos, activa **Ejecutar varios modelos** y dale a cada modelo su propia tarjeta de la misma marca.
<p align="center"><img src="docs/media/app-chat.png" width="900" alt="The Strata app: a chat where the agent built a small voxel pagoda garden, with the page open in the preview on the right"><br>
<sub>Un chat en la app de Strata: el agente escribió un pequeño jardín de pagoda de vóxeles en un solo archivo HTML, y la página se abre en la vista previa junto a la conversación</sub></p>
### En la Terminal
<p align="center"><img src="docs/media/terminal-monitor.png" width="900" alt="The Strata Terminal monitor overview while the model writes a reply"><br>
<sub><code>strata monitor</code> mientras el modelo responde: velocidad, GPU, memoria y las últimas solicitudes</sub></p>
strata update --check # is there a new engine? (strata update installs it)
strata doctor # check the installed engine files
```
`strata --help` lo lista todo. Arrancar y servir no descargan nada. Solo lo hacen la instalación, `strata launch` (cuando
eliges un modelo) y `strata update`. Para llegar a Strata desde tu móvil u otro PC, usa `strata launch --access lan`. Crea
una clave de API y no arranca en una red sin ella. No se abre ningún navegador salvo que añadas `--open`. Guía:
[docs/TERMINAL_LAUNCH.md](docs/TERMINAL_LAUNCH.md). El cliente completo: [docs/TERMINAL.md](docs/TERMINAL.md).
### Con agentes de IA
Un agente instala Strata con el comando de una línea de la opción B y ejecuta
`strata launch --model NAME --yes --no-tui`. Los pasos, las opciones y los códigos de salida están en [docs/AI_SETUP.md](docs/AI_SETUP.md),
y las reglas del repositorio en [AGENTS.md](AGENTS.md).
### Conecta tus otras aplicaciones
| Tipo de aplicación | URL base |
| --- | --- |
| Compatible con OpenAI (la mayoría de aplicaciones y agentes) | `http://127.0.0.1:8080/v1` (cualquier clave de API, cualquier nombre de modelo) |
| API de Anthropic (Claude Code y otros) | `http://127.0.0.1:8080`, sin `/v1`. Estas aplicaciones añaden `/v1/messages` por su cuenta. Claude Code: `ANTHROPIC_BASE_URL=http://127.0.0.1:8080` |
| API OpenAI Responses (Codex CLI) | `http://127.0.0.1:8080/v1/responses` ([configuración](docs/DETAILS.md#the-responses-api-and-codex-cli)) |
`strata connect` muestra los ajustes exactos para OpenCode, Codex CLI, Claude Code, oh-my-pi, pi, Hermes Agent y OpenClaw.
`strata connect opencode --write` escribe el archivo después de preguntar y guarda una copia de seguridad.
### Imágenes
Las imágenes vienen activadas por defecto en los modelos que pueden leerlas. Adjunta o pega una imagen en el chat, o envíala
desde tu aplicación. El lector de imágenes funciona en el procesador cuando tu tarjeta gráfica tiene menos de 24 GB (unos
segundos por imagen, sin quitar memoria a la tarjeta) y en la tarjeta gráfica a partir de 24 GB. La regla es la misma para
NVIDIA y AMD. Puedes desactivar las imágenes por modelo, o elegir el lector con `--vision off`, `--vision gpu` o
`--vision cpu`.
Las cifras detrás de la regla: en una Radeon AI PRO R9700, el lector de la tarjeta tardó 57 ms en una imagen de 260 tokens y
el del procesador 3,1 s. En tarjetas de 16 GB (Tesla P100 y RTX A4000, Linux), el lector de la tarjeta costó un 3-5 % de
velocidad en texto, porque ocupa unos 1,4 GB que de otro modo usaría el modelo. En un Mac lee la GPU, tenga la memoria que tenga (0,45 s para una imagen de 260 tokens en un M4 de 32 GB, unos 1,2 GB de la memoria compartida);
el procesador tarda unos 50 s por imagen, así que solo lee con `--vision cpu`.
### Conviene saber
- Elige el nivel de razonamiento (**desactivado, bajo, medio o alto**) en el menú del chat o en "reasoning effort" de tu aplicación. Desactivado es lo más rápido.
- Strata responde a una petición a la vez por defecto. Para varias a la vez, usa `--parallel N` en `strata launch` o
`"parallel": 2` en la configuración ([BATCHING.md](docs/BATCHING.md)). En una tarjeta de 12 GB, esto hace más lenta cada respuesta.
- Strata lee entero el primer mensaje de un chat, unos 1 minuto por cada 30.000 tokens. Los mensajes siguientes empiezan en segundos.
- Dónde se guardan tus chats y la API completa: [docs/INSTALL.md](docs/INSTALL.md#where-things-are-stored), [docs/DETAILS.md](docs/DETAILS.md#using-it).
## Compilar desde el código fuente
<details>
<summary>Windows, Linux y Mac desde el repositorio</summary>
[Descarga Strata](https://github.com/Niko1221/Strata/archive/refs/heads/main.zip) y descomprímelo, o haz `git clone`.
**Windows:** haz doble clic en `START-HERE.bat`. **Linux:** ejecuta `./setup.sh` en la carpeta de Strata.
**Mac (Apple Silicon, macOS 15 o posterior):** instala una vez las Command Line Tools de Apple con `xcode-select --install`
(no hace falta Xcode). Abre la Terminal en la carpeta de Strata y ejecuta `./setup.sh`. El script compila el motor para Mac
en tu Mac (unos 7 minutos), hace las mismas preguntas, descarga el modelo y lo arranca. La próxima vez, `./setup.sh` lo
arranca enseguida. Qué tamaño cabe en tu Mac: [docs/MAC.md](docs/MAC.md).
El instalador detecta tu tarjeta y prepara el motor adecuado. Pregunta qué modelo y qué tamaño quieres, cuánto contexto
(cuánto texto mantiene el modelo en mente) y si debe leer imágenes. Pulsa Enter para aceptar la respuesta recomendada.
Luego descarga el modelo (unos 70 GB) y lo arranca. Si la descarga se detiene, vuelve a ejecutar el script y continúa. Tu
navegador abre la página de Strata en `http://127.0.0.1:8080`.
> Mientras el modelo arranca, tu PC puede ir lento o dejar de responder durante 1-3 minutos (más tiempo la primera vez).
> Strata carga 35-55 GB en tu RAM y bloquea una parte para la tarjeta gráfica. Espera y no cierres la ventana.
La próxima vez, ejecuta de nuevo `START-HERE.bat` (o `./setup.sh`). Arranca enseguida y no descarga nada dos veces. Cierra
su ventana para detener el modelo. `UPDATE.bat` (`./update.sh`) actualiza Strata sin arrancarlo. Actualización, Docker,
varias tarjetas, dónde van los archivos y todas las opciones: [docs/INSTALL.md](docs/INSTALL.md).
Para medir tu PC, ejecuta `BENCH.bat` (`./bench.sh`). Tarda 5-10 minutos y escribe un informe que puedes compartir
**Opcional: más velocidad en una tarjeta cuando todo el modelo cabe en la RAM.** Dos ajustes que se activan a mano, de
RubenTorney (#1904, #1905). Ayudan cuando usas una sola tarjeta gráfica y tu RAM guarda todos los expertos, por ejemplo con
Q2_0 o IQ3_XXS en un PC de 64 GB. Abre `strata-<model>.json` en la carpeta de datos de Strata ([dónde](docs/INSTALL.md#where-things-are-stored)) y añade:
-`"--adapt-async", "1"` a la lista `"args"`. Así la caché de expertos se actualiza en un hilo auxiliar en vez de detener la
respuesta. Medido en una RTX 5070 (Windows, Q2_0): +4-5 % en decodificación, +7-10 % en código y documentos largos. Las
respuestas pueden diferir un poco de las del ajuste por defecto y son iguales de una ejecución a otra.
-`"STRATA_PIN_BY_NEED": "1"` en el bloque `"env"` (Windows). Windows solo puede bloquear aproximadamente la mitad de la RAM
para la tarjeta gráfica, y este ajuste bloquea primero los expertos que la tarjeta necesita antes. Medido en la RTX 5070: +4,4 % en lectura del prompt.
```json
"args":[...,"--adapt-async","1"],
"env":{"STRATA_PIN_BY_NEED":"1"}
```
Reinicia Strata después del cambio. Para deshacerlo, quita las dos entradas.
Sigue las indicaciones para descargar e iniciar un modelo. [Guía del Terminal](docs/TERMINAL_LAUNCH.md).
</details>
## ¿Cómo funciona?
<p align="center">
<img src="docs/media/terminal-monitor.png" width="900" alt="Aplicación de Terminal de Strata con la actividad del modelo, la velocidad, la memoria y las peticiones recientes">
<br><sub>Observa tu modelo y tu hardware mientras se ejecuta.</sub>
</p>
Los modelos como este suelen ejecutarse en servidores con cientos de gigabytes de memoria gráfica. Tu tarjeta gráfica tiene
12-24 GB. Strata hace que el modelo quepa repartiendo el trabajo por todo tu PC. Imagina una biblioteca.
<a id="build-from-source"></a>
- El modelo es un equipo de 24.576 pequeños especialistas ("expertos"), como 24.576 libros de consulta. Cada palabra necesita solo 10 de ellos.
- **Tu tarjeta gráfica es el escritorio.** Es pequeño y rápido. Los unos 4.000 libros que más se usan están abiertos encima.
- **Tu RAM son las estanterías de la sala de lectura.** Los 24.576 libros están ahí. Cuando una palabra necesita un libro que
no está en el escritorio, tu procesador lo coge enseguida de la estantería mientras la tarjeta sigue trabajando.
- **Tu SSD es el archivo.** Guarda una tabla de consulta de 29 GB. El modelo lee unas pocas filas por palabra, así que no
hace falta que esté en la RAM.
[Compilar desde el código fuente](docs/SOURCE_INSTALL.md) · [Docker](docs/DOCKER.md) · [Guía de instalación completa](docs/INSTALL.md)
<p align="center"><img src="docs/media/how-it-works.svg" width="860" alt="The model's 24,576 experts: the busiest on the graphics card, all of them in RAM, a lookup table on the SSD"></p>
<a id="what-you-need"></a>
Un pequeño ayudante también adivina las siguientes palabras. El modelo grande las comprueba todas a la vez y se queda con
las correctas. Obtienes la misma respuesta, entre 1,6 y 1,8 veces antes. Los textos largos se leen en trozos grandes, de
hasta 8.192 tokens a la vez, a más de 1.000 tokens por segundo.
## Hardware
<p align="center"><img src="docs/media/guess-and-check.svg" width="860" alt="A small helper guesses the next words; the big model checks them all at once and keeps the right ones"></p>
| **Intel Arc · Linux** | [Instalación en Intel](docs/INTEL_ARC.md) |
| **Apple Silicon · M1 o posterior** | [Instalación en Mac](docs/MAC.md) |
| **AMD Ryzen AI Max · Strix Halo** | [Instalación en Strix Halo](docs/STRIX_HALO.md) |
La explicación más larga: [docs/HOW_IT_WORKS.md](docs/HOW_IT_WORKS.md). Cada parte con sus cifras:
[los detalles](docs/DETAILS.md#how-it-works) y el [artículo](docs/paper/Strata-Paper.pdf).
<a id="using-it"></a>
## ¿Algo ha ido mal?
## Herramientas
- **Mi PC se congeló la primera vez que arrancó Strata.** Carga el modelo en la RAM. Espera y no cierres la ventana. ¿Sigue
congelado después de 10 minutos? Reinicia el PC, cierra otros programas e inténtalo de nuevo, o elige un tamaño más pequeño.
- **Se detuvo durante la descarga o la instalación.** Abre de nuevo la aplicación o `strata launch`, o vuelve a ejecutar
`START-HERE.bat` (o `./setup.sh`). Continúa donde se quedó.
- **Va muy lento, la luz del disco parpadea sin parar, o dice "the engine stopped unexpectedly".** Tu PC no tiene bastante
RAM libre. Cierra otros programas (los navegadores gastan mucha) o elige un tamaño más pequeño, como Q2_0 o IQ2_XS.
- **Dice que no hay motor para este PC.** En la aplicación, abre **Ajustes, Motor** y pulsa **Actualizar motor**. En la
Terminal, ejecuta `strata update`. `strata doctor` muestra qué archivos del motor están instalados.
- **Dice que el puerto 8080 ya está en uso.** Strata ya se está ejecutando. Busca su ventana.
**Claude Code · Codex CLI · OpenCode · oh-my-pi · pi.dev · Hermes Agent · OpenClaw**, y más, mediante las API compatibles con OpenAI y Anthropic.
Más problemas y soluciones: [docs/TROUBLESHOOTING.md](docs/TROUBLESHOOTING.md). ¿Sigues atascado? Abre una
[incidencia](https://github.com/Niko1221/Strata/issues). En la aplicación, adjunta el archivo de **Ajustes, Motor,
Diagnóstico, Ver registros, Guardar registros**. En la Terminal, pega la salida de `strata doctor`. Si lo ejecutas desde el
código fuente, adjunta `strata-<model>.log` de la carpeta de Strata. Para un problema de seguridad, avísanos en privado:
[SECURITY.md](SECURITY.md).
Ejecuta **`strata connect`** para ver las instrucciones de tu herramienta. [Guía de conexión](docs/TERMINAL_LAUNCH.md#connect-your-apps-strata-connect).
## Créditos y licencia
<a id="guides"></a>
<a id="how-fast-is-it"></a>
<a id="how-does-it-work"></a>
<a id="something-went-wrong"></a>
El modelo es [Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next), del equipo de Qwen. Lo comprimieron
y Unsloth. Strata usa partes de [llama.cpp / ggml](https://github.com/ggml-org/llama.cpp). Todos los créditos:
[docs/HOW_IT_WORKS.md](docs/HOW_IT_WORKS.md#credits). Strata es de código abierto con la [licencia MIT](LICENSE). Algunas
partes y todos los modelos tienen sus propias licencias ([cuáles](docs/HOW_IT_WORKS.md#license)).
## Ayuda
## Contribuir
¿Has encontrado un error, o quieres enviar un cambio o un benchmark? Lee primero [CONTRIBUTING.md](CONTRIBUTING.md). Explica
qué incluir para que otros puedan comprobar tu trabajo.
[Todas las guías](docs) · [Solución de problemas](docs/TROUBLESHOOTING.md) · [Informar de un error](https://github.com/Niko1221/Strata/issues/new/choose) ·
Strata es gratis y tiene licencia MIT. Lo construyo y lo pruebo en el hardware que consigo. Si te funciona bien en tu PC, un café me ayuda
a seguir probando tarjetas nuevas y a arreglar lo que nos cuentas. Sin presión. Usarlo y contar lo que encuentras ayuda igual de mucho.
**Ayuda a mantener la IA local abierta a todos.** Un café ayuda a financiar el desarrollo y las pruebas en más hardware.
<a href="https://buymeacoffee.com/strataengine"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="50"></a>
<a href="https://buymeacoffee.com/strataengine"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me a Coffee: apoya Strata" height="48"></a>
¿Te interesa patrocinar Strata? Escríbeme a través de [GitHub](https://github.com/Niko1221).
[Licencia MIT](LICENSE) · [Créditos y licencias de los modelos](docs/HOW_IT_WORKS.md#credits) · [Privacidad](docs/SAFETY.md)
<a id="option-c-let-your-ai-set-it-up"></a>
<sub>Para herramientas de IA y lectores automáticos: <a href="AGENTS.md">referencia técnica</a> · <a href="docs/AI_SETUP.md">instrucciones de instalación</a>.</sub>
<p align="center"><a href="https://github.com/Niko1221/Strata/releases/download/v0.1.10/Pagoda.mp4"><img src="docs/media/pagoda-preview.webp" width="720" alt="A voxel pagoda garden that Strata's model wrote, running in the browser"></a><br>
<sub>Un jardin de pagode en voxels, généré en une seule requête sur une RTX 5070 avec Strata (IQ3_S, contexte de 128K) ·
<a href="https://buymeacoffee.com/strataengine"><img src="https://img.shields.io/badge/Buy_me_a_coffee-FFDD00?logo=buymeacoffee&logoColor=000" alt="Soutenir Strata sur Buy Me a Coffee"></a>
</p>
Strata fait tourner **[Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)**, un modèle de 125 milliards
de paramètres qui demande normalement un serveur, sur un PC ordinaire. Il discute, écrit du code, lit des images et
fonctionne avec vos applications et vos agents de code. Vos conversations et vos fichiers restent sur votre PC.
**Nouveau ici ?** Regardez [ce qu'il vous faut](#ce-quil-vous-faut), puis [installez l'application](#option-a--lapplication-strata).
**Gratuit, open source, et à vous de l'exécuter.** Utilisez Strata sur votre bureau ou avec vos outils de programmation préférés. Aucun compte nécessaire.
- **Tourne sur un PC de jeu.** 32 Go de RAM et de mémoire graphique réunis suffisent pour commencer.
- **Fonctionne avec vos outils.** Il parle les API d'OpenAI et d'Anthropic, donc Claude Code, Codex CLI, OpenCode et d'autres agents peuvent s'y connecter.
- **Lit les images.** C'est activé par défaut.
- **Gratuit.** Licence MIT, pas de compte, pas de cloud.
- **Privé.** Aucune télémétrie. Ce que Strata télécharge, quand il utilise internet et comment il vérifie les téléchargements est expliqué dans [docs/SAFETY.md](docs/SAFETY.md).
## Modèles
Version 0.2.3 · moteur 0.1.48
Nous recommandons **d'abord ISTA-DASLab**, puis **Unsloth** et d'autres fournisseurs, lorsqu'un modèle compatible tient en mémoire.
| **Gyro · agentionai** | [Gyro-S, Gyro-M](https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF). Utilise moins de mémoire, mais plus lent. |
Dans l'**application de bureau** et dans le **Terminal**. Ce qui ne tient pas en mémoire est lu depuis votre SSD : il faut donc beaucoup d'espace disque libre.
Fonctionne sur les NVIDIA RTX 30, 40 et 50 (Windows et Linux) et sur l'AMD Radeon AI PRO R9700 (Linux). L'installation télécharge les fichiers et le moteur DeepSeek, puis démarre le modèle.
Mesuré avec le jeu 2.00 bpw sur Strata 0.2.5. « Sortie » est la vitesse de la réponse, « Prompt » la vitesse à laquelle Strata lit votre texte.
| Matériel | Sortie | Prompt |
| --- | ---: | ---: |
| RTX 5070 12 GB, 64 GB de RAM | 4 à 6 tokens/s | environ 110 tokens/s (4K) |
| RTX 3060 12 GB, 46 GB de RAM | 9 à 11 tokens/s | environ 155 tokens/s (4K) |
| 1x Radeon AI PRO R9700, 251 GB de RAM | 21 à 25 tokens/s | 320 à 360 tokens/s (8K) |
| 4x Radeon AI PRO R9700 | 74 à 77 tokens/s | 400 à 450 tokens/s (8K) |
[Mémoire, disque et détails](docs/DEEPSEEK41.md) · [Notes de version](https://github.com/Niko1221/Strata/releases)
</details>
<a id="the-other-sizes"></a>
Vous ne savez pas ce qui tient ? [Choisir un modèle pour votre ordinateur](docs/MODELS.md#pick-by-memory).
## Installation
Choisissez une option. Elles lancent toutes le même moteur et utilisent les mêmes modèles.
<a id="option-a-the-strata-app"></a>
### Option A : l'application Strata
### Application de bureau
Le plus simple. L'application contient le moteur : installez-la, choisissez un modèle, discutez. Téléchargez le fichier
correspondant à votre système et à votre carte graphique depuis la [page des versions](https://github.com/Niko1221/Strata/releases/latest).
Vous ne savez pas lequel prendre ? Lisez le paragraphe sous le tableau.
**[Télécharger Strata pour Windows, Linux ou Mac →](https://github.com/Niko1221/Strata/releases/latest)**
| Système | Votre carte graphique | Fichier à télécharger |
| --- | --- | --- |
| **Windows 10 / 11** | NVIDIA RTX, pilote 580 ou plus récent | [`Strata-0.2.3-windows-cuda13.exe`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-windows-cuda13.exe) |
| | NVIDIA, carte plus ancienne ou pilote antérieur à 580 | [`Strata-0.2.3-windows-cuda12.exe`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-windows-cuda12.exe) |
| | AMD Radeon série RX 9000 ou Radeon AI PRO R9700 | [`Strata-0.2.3-windows-hip.exe`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-windows-hip.exe) |
| | AMD Radeon série RX 7000 ou Ryzen AI Max (Strix Halo) | [`Strata-0.2.3-windows-hip-rdna3.exe`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-windows-hip-rdna3.exe) |
| **Linux** | NVIDIA RTX, pilote 580 ou plus récent | [`Strata-0.2.3-linux-nvidia-cuda13.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-nvidia-cuda13.AppImage) |
| | NVIDIA, carte plus ancienne ou pilote antérieur à 580 | [`Strata-0.2.3-linux-nvidia-cuda12.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-nvidia-cuda12.AppImage) |
| | AMD Radeon série RX 9000 ou Radeon AI PRO R9700 | [`Strata-0.2.3-linux-amd-rx9070.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-amd-rx9070.AppImage) |
| | AMD Radeon série RX 7000 | [`Strata-0.2.3-linux-amd-rx7900.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-amd-rx7900.AppImage) |
| | AMD Ryzen AI Max (Strix Halo) | [`Strata-0.2.3-linux-amd-strixhalo.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-amd-strixhalo.AppImage) |
| **Mac** (Apple Silicon, macOS 15 ou plus récent) | intégrée | [`Strata-0.2.3-mac.dmg`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-mac.dmg) |
Installez l'application, choisissez un modèle et commencez à discuter. Il y a un seul installateur par système. Au premier lancement, l'application télécharge le moteur de votre carte graphique. [Aide pour l'installateur](docs/INSTALL.md#the-app).
**Si vous hésitez :** pour NVIDIA, prenez `cuda13` si la version de votre pilote est 580 ou plus récente, sinon `cuda12`.
`cuda12` convient aux cartes plus anciennes et aux pilotes plus anciens. Les fichiers `hip` et `amd` sont pour les cartes
AMD. Les autres cartes AMD (RX 6000) n'ont pas encore d'application prête à l'emploi : voir
Le script analyse votre PC, télécharge le Terminal, vérifie son empreinte SHA-256 et le décompresse dans votre dossier
personnel. Il ne demande pas de droits d'administrateur. Sous Windows, ouvrez ensuite une nouvelle fenêtre de terminal.
Puis lancez Strata :
Sous Windows, ouvrez ensuite un nouveau terminal. Puis :
```sh
strata launch
```
Il pose une question à la fois : modèle, contexte, images, qui peut se connecter. Ensuite il télécharge le moteur pour votre
carte graphique (signé, et lancé à titre d'essai avant usage) ainsi que le modèle, puis commence à servir. Pour les scripts
et les agents d'IA, utilisez des options :
```sh
strata launch --model q2_0 --yes --no-tui
```
Pour lire le script d'installation avant de l'exécuter, téléchargez-le d'abord avec `curl -fsSL <address> -o install.sh`,
lisez-le, puis lancez `sh install.sh`. Tous les écrans et toutes les options : [docs/TERMINAL_LAUNCH.md](docs/TERMINAL_LAUNCH.md).
### Option C : laissez votre IA l'installer
Collez ceci dans Claude Code, Cursor, Codex, GitHub Copilot ou un autre assistant de programmation :
```text
Set up Strata on this PC for me: https://github.com/Niko1221/Strata - follow docs/AI_SETUP.md in that repository.
```
L'assistant vérifie votre carte graphique, votre RAM et votre disque, choisit le modèle adapté, l'installe, le démarre et
vous explique comment connecter vos applications. Les outils d'IA peuvent aussi installer, démarrer et arrêter Strata via son
[serveur MCP](docs/MCP_SERVER.md).
Vous préférez le compiler vous-même ? Voir [Compiler depuis les sources](#compiler-depuis-les-sources).
## Modèles
L'application et `strata launch` recommandent un modèle selon votre mémoire. C'est un seul modèle en plusieurs tailles,
plus ou moins compressé. Les petites tailles sont plus rapides. Les grandes sont un peu plus intelligentes.
### Peu de mémoire ? Essayez Gyro
[Gyro](https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF) (par agentionai) est la façon la plus légère de faire tourner ce modèle.
| Taille | Mémoire nécessaire (RAM + VRAM) | Téléchargement | Ce que c'est |
| --- | ---: | ---: | --- |
| **Gyro-S** | 27,6 Go | 58,5 Go | Pour les PC où rien d'autre ne tient. La qualité est inférieure à celle d'IQ2_XS. |
| **Gyro-M** | 35 Go | 92,0 Go | À peu près la qualité d'IQ3_XXS avec environ 20 % de mémoire GPU en moins. Il écrit les réponses 16 % plus lentement que Q2_0 sur une RTX 3060 de 12 Go et 21 % plus lentement sur une R9700. |
Gyro fonctionne sur les cartes NVIDIA et AMD (expérimental sur AMD) sur les cartes Intel Arc sous Linux (testé sur une Arc A750) et, depuis la 0.2.2, sur les Mac Apple Silicon (Gyro-S sur un Mac mini de 32 Go : 9 tokens/s, experts lus sur le SSD). Détails et mesures :
[docs/GYRO.md](docs/GYRO.md). Lancez-le avec `strata launch --model gyro-s`.
| **32 Go** | **Gyro-S**, ou le **[Coder](docs/MODELS.md#coder)** pour le code | Le Coder demande 32 Go de RAM et garde la moitié des experts, il est donc moins bon hors du code |
| **40 Go** | **[Swift 1.5](docs/MODELS.md#swift-15)** IQ2_XS (39,2 Go), ou **Gyro-M** (35 Go) | Swift réfléchit moins longtemps avant de répondre. Gyro-M approche la qualité d'IQ3_XXS, mais il écrit les réponses 16 à 21 % plus lentement que Q2_0 |
| **64 Go** | **IQ3_S** (54,8 Go), l'original ou Swift 1.5 | Le plus lent, le meilleur. Il égale le modèle complet dans les tests publiés |
| **96 Go ou plus** | **UD-IQ4_XS** d'Unsloth (environ 72 Go de RAM) jusqu'à **Q8_0** (environ 136 Go), 4 à 8 bits | Plus proche du modèle complet. Avec moins de RAM, Strata lit une partie depuis le SSD, ce qui est beaucoup plus lent |
Avoir moins de RAM que ce qu'indique une ligne peut encore fonctionner : Strata passe tout seul en
[mode peu de RAM](docs/MODELS.md#a-big-graphics-card-and-little-ram). Attendez-vous à plus de lenteur, et à bien plus avec
une petite carte graphique.
Autres choix :
- **[Swift 1.5](docs/MODELS.md#swift-15)** est une version affinée qui réfléchit moins longtemps avant de répondre, pour une qualité et un besoin en RAM à peu près identiques.
- **[NVFP4](docs/NVFP4.md)** est la version 4 bits de NVIDIA pour les cartes NVIDIA. Elle demande environ 80 Go de RAM pour garder tous les experts en RAM.
- **[Qwen3.6-35B-A3B](docs/QWEN36.md)** est un modèle plus petit et différent, qui lit les images et peut répondre à plusieurs requêtes à la fois.
- **[L'Uncensored IQ3_XXS d'OrcaRouter](docs/MODELS.md#orcarouter-uncensored-iq3_xxs)** se configure à la main.
Toutes les tailles, les téléchargements et ce qui tient où : [docs/MODELS.md](docs/MODELS.md).
## Quelle est sa vitesse ?
Mesuré avec une carte graphique de 12 Go (RTX 5070, Ryzen 5 7600, 64 Go de RAM). Voir les résultats d'autres cartes :
[vitesse de chaque taille](docs/MODELS.md#how-fast-is-each-size), [résultats de la communauté](docs/COMMUNITY_BENCHMARKS.md).
| Modèle | Les réponses s'affichent à | Lit votre prompt à |
Les barres utilisent la même échelle dans les deux tableaux. Une carte avec plus de VRAM est plus rapide. Une RTX 3090
(24 Go) devrait écrire environ 100 à 140 tokens par seconde. La version du moteur derrière chaque chiffre est indiquée dans
[DETAILS.md](docs/DETAILS.md#speed-measured).
## Utilisation
### Dans l'application
L'application comprend un **Chat**, un **Monitor** du modèle, de votre GPU, de votre CPU et de votre RAM, une page **Models** et les **Settings**. Le chat peut travailler dans un dossier de votre PC. En mode **Manual**, il demande
avant de modifier un fichier ou d'exécuter une commande. Tapez `/loop 10m <what to do>` pour répéter un message toutes les
10 minutes tant que le chat est ouvert.
Le chat peut chercher sur le web avec DuckDuckGo (ou Bing), sans compte. Vous pouvez ajouter une clé de recherche sous
**Settings, Tools**. La recherche est la seule fonction qui envoie quelque chose hors de votre PC (les mots recherchés).
Pour faire tourner deux modèles, activez **Run multiple models** (exécuter plusieurs modèles) et donnez à chacun sa propre carte de la même marque.
<p align="center"><img src="docs/media/app-chat.png" width="900" alt="The Strata app: a chat where the agent built a small voxel pagoda garden, with the page open in the preview on the right"><br>
<sub>Une discussion dans l'app Strata : l'agent a écrit un petit jardin de pagode en voxels dans un seul fichier HTML, et la page s'ouvre dans l'aperçu à côté de la conversation</sub></p>
### Dans le Terminal
<p align="center"><img src="docs/media/terminal-monitor.png" width="900" alt="The Strata Terminal monitor overview while the model writes a reply"><br>
<sub><code>strata monitor</code> pendant que le modèle répond : vitesse, GPU, mémoire et les dernières requêtes</sub></p>
strata update --check # is there a new engine? (strata update installs it)
strata doctor # check the installed engine files
```
`strata --help` liste tout. Démarrer et servir ne téléchargent rien. Seuls l'installation, `strata launch` (quand vous
choisissez un modèle) et `strata update` le font. Pour joindre Strata depuis votre téléphone ou un autre PC, utilisez
`strata launch --access lan`. La commande crée une clé d'API et refuse de démarrer sur un réseau sans elle. Aucun
navigateur ne s'ouvre, sauf si vous ajoutez `--open`. Guide : [docs/TERMINAL_LAUNCH.md](docs/TERMINAL_LAUNCH.md). Le client complet : [docs/TERMINAL.md](docs/TERMINAL.md).
### Avec des agents d'IA
Un agent installe Strata avec la commande en une ligne de l'option B et lance
`strata launch --model NAME --yes --no-tui`. Les étapes, les options et les codes de sortie sont dans [docs/AI_SETUP.md](docs/AI_SETUP.md),
et les règles du dépôt dans [AGENTS.md](AGENTS.md).
### Connecter vos autres applications
| Type d'application | URL de base |
| --- | --- |
| Compatible OpenAI (la plupart des applications et des agents) | `http://127.0.0.1:8080/v1` (n'importe quelle clé d'API, n'importe quel nom de modèle) |
| API Anthropic (Claude Code et d'autres) | `http://127.0.0.1:8080`, sans `/v1`. Ces applications ajoutent `/v1/messages` elles-mêmes. Claude Code : `ANTHROPIC_BASE_URL=http://127.0.0.1:8080` |
| API OpenAI Responses (Codex CLI) | `http://127.0.0.1:8080/v1/responses` ([configuration](docs/DETAILS.md#the-responses-api-and-codex-cli)) |
`strata connect` affiche les réglages exacts pour OpenCode, Codex CLI, Claude Code, oh-my-pi, pi, Hermes Agent et OpenClaw.
`strata connect opencode --write` écrit le fichier après confirmation et garde une sauvegarde.
### Images
Les images sont activées par défaut pour les modèles qui savent les lire. Joignez ou collez une image dans le chat, ou
envoyez-la depuis votre application. Le lecteur d'images tourne sur le processeur quand votre carte graphique a moins de
24 Go (quelques secondes par image, sans prendre de mémoire à la carte) et sur la carte graphique à partir de 24 Go. La
règle est la même pour NVIDIA et AMD. Vous pouvez désactiver les images par modèle, ou choisir le lecteur avec
`--vision off`, `--vision gpu` ou `--vision cpu`.
Les chiffres derrière la règle : sur une Radeon AI PRO R9700, le lecteur de la carte a pris 57 ms pour une image de
260 tokens et celui du processeur 3,1 s. Sur des cartes de 16 Go (Tesla P100 et RTX A4000, Linux), le lecteur de la carte
a coûté 3 à 5 % de vitesse sur le texte, parce qu'il occupe environ 1,4 Go que le modèle utiliserait sinon. Sur un Mac, le lecteur tourne sur le GPU quelle que soit la mémoire (0,45 s pour une image de 260 tokens sur un M4 de 32 Go, environ 1,2 Go de la
mémoire partagée) ; le processeur met environ 50 s par image, il ne lit donc qu'avec `--vision cpu`.
### Bon à savoir
- Choisissez le niveau de réflexion (**désactivé, faible, moyen ou élevé**) dans le menu du chat ou dans le réglage « reasoning effort » de votre application. Désactivé est le plus rapide.
- Strata répond à une requête à la fois par défaut. Pour en traiter plusieurs en même temps, utilisez `--parallel N` dans `strata launch` ou
`"parallel": 2` dans la configuration ([BATCHING.md](docs/BATCHING.md)). Sur une carte de 12 Go, chaque réponse devient alors plus lente.
- Strata lit en entier le premier message d'une discussion, soit environ 1 minute pour 30 000 tokens. Les messages suivants démarrent en quelques secondes.
- Où sont stockées vos discussions et l'API complète : [docs/INSTALL.md](docs/INSTALL.md#where-things-are-stored), [docs/DETAILS.md](docs/DETAILS.md#using-it).
## Compiler depuis les sources
<details>
<summary>Windows, Linux et Mac depuis le dépôt</summary>
[Téléchargez Strata](https://github.com/Niko1221/Strata/archive/refs/heads/main.zip) et décompressez-le, ou faites un `git clone`.
**Windows :** double-cliquez sur `START-HERE.bat`. **Linux :** lancez `./setup.sh` dans le dossier Strata.
**Mac (Apple Silicon, macOS 15 ou plus récent) :** installez une fois les Command Line Tools d'Apple avec `xcode-select --install`
(Xcode lui-même n'est pas nécessaire). Ouvrez le Terminal dans le dossier Strata et lancez `./setup.sh`. Le script compile
le moteur Mac sur votre Mac (environ 7 minutes), pose les mêmes questions, télécharge le modèle et le démarre. La fois
suivante, `./setup.sh` le démarre tout de suite. Quelle taille convient à votre Mac : [docs/MAC.md](docs/MAC.md).
L'installateur détecte votre carte et prépare le bon moteur. Il demande quel modèle et quelle taille, combien de contexte
(la quantité de texte que le modèle garde en tête) et s'il doit lire les images. Appuyez sur Entrée pour la réponse
recommandée. Puis il télécharge le modèle (environ 70 Go) et le démarre. Si le téléchargement s'arrête, relancez le script
et il reprend. Votre navigateur ouvre la page de Strata à l'adresse `http://127.0.0.1:8080`.
> Pendant le démarrage du modèle, votre PC peut être lent ou ne plus répondre pendant 1 à 3 minutes (le plus longtemps la
> première fois). Strata charge 35 à 55 Go dans votre RAM et en verrouille une partie pour la carte graphique. Patientez et
> ne fermez pas la fenêtre.
La fois suivante, relancez `START-HERE.bat` (ou `./setup.sh`). Il démarre tout de suite et ne télécharge rien deux fois.
Fermez sa fenêtre pour arrêter le modèle. `UPDATE.bat` (`./update.sh`) met Strata à jour sans le démarrer. Mise à jour,
Docker, plusieurs cartes, où vont les fichiers et toutes les options : [docs/INSTALL.md](docs/INSTALL.md).
Pour mesurer votre PC, lancez `BENCH.bat` (`./bench.sh`). Cela prend 5 à 10 minutes et produit un rapport que vous pouvez
**Facultatif : plus de vitesse sur une carte quand tout le modèle tient en RAM.** Deux réglages à activer soi-même, par
RubenTorney (#1904, #1905). Ils aident quand vous utilisez une seule carte graphique et que votre RAM contient tous les
experts, par exemple Q2_0 ou IQ3_XXS sur un PC de 64 Go. Ouvrez `strata-<model>.json` dans le dossier de données de Strata ([où](docs/INSTALL.md#where-things-are-stored)) et ajoutez :
-`"--adapt-async", "1"` à la liste `"args"`. Le cache des experts se met alors à jour sur un thread auxiliaire au lieu de
mettre la réponse en pause. Mesuré sur une RTX 5070 (Windows, Q2_0) : +4 à 5 % en décodage, +7 à 10 % sur le code et les
longs documents. Les réponses peuvent différer légèrement de celles du réglage par défaut, et restent identiques d'une exécution à l'autre.
-`"STRATA_PIN_BY_NEED": "1"` dans le bloc `"env"` (Windows). Windows ne peut verrouiller qu'environ la moitié de la RAM
pour la carte graphique, et ce réglage verrouille d'abord les experts dont la carte a le plus besoin. Mesuré sur la RTX 5070 : +4,4 % en lecture du prompt.
```json
"args":[...,"--adapt-async","1"],
"env":{"STRATA_PIN_BY_NEED":"1"}
```
Redémarrez Strata après la modification. Pour annuler, supprimez les deux entrées.
Suivez les invites pour télécharger et démarrer un modèle. [Guide du Terminal](docs/TERMINAL_LAUNCH.md).
</details>
## Comment ça marche ?
<p align="center">
<img src="docs/media/terminal-monitor.png" width="900" alt="Application Terminal de Strata affichant l'activité du modèle, la vitesse, la mémoire et les requêtes récentes">
<br><sub>Suivez votre modèle et votre matériel pendant qu'il tourne.</sub>
</p>
Les modèles de ce genre tournent d'habitude sur des serveurs qui disposent de centaines de gigaoctets de mémoire graphique.
Votre carte graphique en a 12 à 24. Strata fait tenir le modèle en répartissant le travail sur tout votre PC. Imaginez une
bibliothèque.
<a id="build-from-source"></a>
- Le modèle est une équipe de 24 576 petits spécialistes (les « experts »), comme 24 576 ouvrages de référence. Chaque mot n'en demande que 10.
- **Votre carte graphique est le bureau.** Il est petit et rapide. Les quelque 4 000 livres les plus utilisés y sont ouverts.
- **Votre RAM, ce sont les étagères de la salle de lecture.** Les 24 576 livres s'y trouvent. Quand un mot demande un livre
qui n'est pas sur le bureau, votre processeur va le chercher aussitôt sur l'étagère, pendant que la carte continue de travailler.
- **Votre SSD, ce sont les archives.** Il contient une table de consultation de 29 Go. Le modèle n'en lit que quelques lignes par
mot, elle n'a donc pas besoin de tenir en RAM.
[Compiler depuis les sources](docs/SOURCE_INSTALL.md) · [Docker](docs/DOCKER.md) · [Guide d'installation complet](docs/INSTALL.md)
<p align="center"><img src="docs/media/how-it-works.svg" width="860" alt="The model's 24,576 experts: the busiest on the graphics card, all of them in RAM, a lookup table on the SSD"></p>
<a id="what-you-need"></a>
Un petit assistant devine aussi les mots suivants. Le grand modèle les vérifie tous d'un coup et garde les bons. Vous
obtenez la même réponse, 1,6 à 1,8 fois plus tôt. Les textes longs sont lus par gros morceaux, jusqu'à 8 192 tokens à la
fois, à plus de 1 000 tokens par seconde.
## Matériel
<p align="center"><img src="docs/media/guess-and-check.svg" width="860" alt="A small helper guesses the next words; the big model checks them all at once and keeps the right ones"></p>
| **Apple Silicon · M1 ou plus récent** | [Installation sur Mac](docs/MAC.md) |
| **AMD Ryzen AI Max · Strix Halo** | [Installation Strix Halo](docs/STRIX_HALO.md) |
L'explication plus longue : [docs/HOW_IT_WORKS.md](docs/HOW_IT_WORKS.md). Chaque élément avec ses chiffres :
[les détails](docs/DETAILS.md#how-it-works) et l'[article](docs/paper/Strata-Paper.pdf).
<a id="using-it"></a>
## Un problème ?
## Outils
- **Mon PC s'est figé au premier démarrage de Strata.** Il charge le modèle en RAM. Patientez et ne fermez pas la fenêtre.
Toujours figé après 10 minutes ? Redémarrez le PC, fermez les autres programmes et réessayez, ou choisissez une taille plus petite.
- **Ça s'est arrêté pendant le téléchargement ou l'installation.** Relancez l'application ou `strata launch`, ou relancez
`START-HERE.bat` (ou `./setup.sh`). Cela reprend là où ça s'était arrêté.
- **C'est très lent, le voyant du disque clignote sans arrêt, ou il affiche « the engine stopped unexpectedly ».** Votre PC
n'a pas assez de RAM libre. Fermez d'autres programmes (les navigateurs en consomment beaucoup) ou choisissez une taille plus petite, comme Q2_0 ou IQ2_XS.
- **Il indique qu'il n'y a pas de moteur pour ce PC.** Dans l'application, ouvrez **Settings, Engine** et appuyez sur **Update engine** (mettre à jour le moteur). Dans le Terminal, lancez `strata update`. `strata doctor` montre quels fichiers du moteur sont installés.
- **Il indique que le port 8080 est déjà utilisé.** Strata tourne déjà. Cherchez sa fenêtre.
**Claude Code · Codex CLI · OpenCode · oh-my-pi · pi.dev · Hermes Agent · OpenClaw**, et d'autres via les API compatibles OpenAI et Anthropic.
Autres problèmes et solutions : [docs/TROUBLESHOOTING.md](docs/TROUBLESHOOTING.md). Toujours bloqué ? Ouvrez une
[issue](https://github.com/Niko1221/Strata/issues). Dans l'application, joignez le fichier obtenu par **Settings, Engine, Diagnostics, View logs, Save logs**. Dans le Terminal, collez la sortie de `strata doctor`. Si
vous avez lancé depuis les sources, joignez `strata-<model>.log` du dossier Strata. Pour un problème de sécurité,
signalez-le en privé : [SECURITY.md](SECURITY.md).
Lancez **`strata connect`** pour voir la configuration de votre outil. [Guide de connexion](docs/TERMINAL_LAUNCH.md#connect-your-apps-strata-connect).
## Remerciements et licence
<a id="guides"></a>
<a id="how-fast-is-it"></a>
<a id="how-does-it-work"></a>
<a id="something-went-wrong"></a>
Le modèle est [Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next), de l'équipe Qwen. Il a été compressé par
et Unsloth. Strata utilise des parties de [llama.cpp / ggml](https://github.com/ggml-org/llama.cpp). Tous les
remerciements : [docs/HOW_IT_WORKS.md](docs/HOW_IT_WORKS.md#credits). Strata est open source sous [licence MIT](LICENSE).
Quelques parties et tous les modèles ont leurs propres licences ([lesquelles](docs/HOW_IT_WORKS.md#license)).
## Aide
## Contribuer
Vous avez trouvé un bug, ou vous voulez proposer une modification ou un benchmark ? Lisez d'abord [CONTRIBUTING.md](CONTRIBUTING.md).
Il indique ce qu'il faut inclure pour que d'autres puissent vérifier votre travail.
[Tous les guides](docs) · [Dépannage](docs/TROUBLESHOOTING.md) · [Signaler un bogue](https://github.com/Niko1221/Strata/issues/new/choose) ·
[Benchmarks](docs/COMMUNITY_BENCHMARKS.md) · [Comment ça marche](docs/HOW_IT_WORKS.md) · [Contribuer](CONTRIBUTING.md)
## Soutenir Strata
Strata est gratuit et sous licence MIT. Je le construis et je le teste sur le matériel que je peux obtenir. S'il tourne bien sur votre PC, un café m'aide
à continuer de tester de nouvelles cartes et à corriger ce que vous signalez. Aucune obligation. L'utiliser et signaler ce que vous constatez aide tout autant.
**Aidez à garder l'IA locale ouverte à tous.** Un café aide à financer le développement et les tests sur davantage de matériel.
<a href="https://buymeacoffee.com/strataengine"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="50"></a>
<a href="https://buymeacoffee.com/strataengine"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me a Coffee : soutenir Strata" height="48"></a>
Vous souhaitez sponsoriser Strata ? Contactez-moi via [GitHub](https://github.com/Niko1221).
[Licence MIT](LICENSE) · [Crédits et licences des modèles](docs/HOW_IT_WORKS.md#credits) · [Confidentialité](docs/SAFETY.md)
<a id="option-c-let-your-ai-set-it-up"></a>
<sub>Pour les outils d'IA et les lecteurs automatiques : <a href="AGENTS.md">référence technique</a> · <a href="docs/AI_SETUP.md">instructions d'installation</a>.</sub>
<p align="center"><a href="https://github.com/Niko1221/Strata/releases/download/v0.1.10/Pagoda.mp4"><img src="docs/media/pagoda-preview.webp" width="720" alt="A voxel pagoda garden that Strata's model wrote, running in the browser"></a><br>
<a href="https://buymeacoffee.com/strataengine"><img src="https://img.shields.io/badge/Buy_me_a_coffee-FFDD00?logo=buymeacoffee&logoColor=000" alt="Buy Me a Coffee で Strata を支援する"></a>
<p align="center"><img src="docs/media/app-chat.png" width="900" alt="The Strata app: a chat where the agent built a small voxel pagoda garden, with the page open in the preview on the right"><br>
<sub>Strata アプリでのチャット。エージェントが小さなボクセルの塔の庭を 1 つの HTML ファイルとして書き、そのページが会話の横のプレビューで開きます</sub></p>
### ターミナルで
<p align="center"><img src="docs/media/terminal-monitor.png" width="900" alt="The Strata Terminal monitor overview while the model writes a reply"><br>
<p align="center"><img src="docs/media/how-it-works.svg" width="860" alt="The model's 24,576 experts: the busiest on the graphics card, all of them in RAM, a lookup table on the SSD"></p>
<p align="center"><img src="docs/media/guess-and-check.svg" width="860" alt="A small helper guesses the next words; the big model checks them all at once and keeps the right ones"></p>
<a href="https://buymeacoffee.com/strataengine"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me a Coffee: Strata を支援する" height="48"></a>
<a href="https://buymeacoffee.com/strataengine"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="50"></a>
<p align="center"><a href="https://github.com/Niko1221/Strata/releases/download/v0.1.10/Pagoda.mp4"><img src="docs/media/pagoda-preview.webp" width="720" alt="A voxel pagoda garden that Strata's model wrote, running in the browser"></a><br>
<sub>A voxel pagoda garden, 1 shot prompt running on an RTX 5070 with Strata (IQ3_S, 128K context) ·
<a href="https://github.com/Niko1221/Strata/releases/download/v0.1.10/Pagoda.mp4">full video (49 s)</a></sub></p>
<a href="https://buymeacoffee.com/strataengine"><img src="https://img.shields.io/badge/Buy_me_a_coffee-FFDD00?logo=buymeacoffee&logoColor=000" alt="Support Strata on Buy Me a Coffee"></a>
</p>
Strata runs **[Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)**, a 125-billion-parameter model that
normally needs a server, on a normal PC. It chats, writes code, reads pictures and works with your apps and coding
In the **desktop app** and the **Terminal**. What does not fit in memory is read from your SSD, so it needs a lot of free
disk space.
| Set (IST-DAS) | Download | Free disk needed |
| --- | ---: | ---: |
| 1.75 bpw | 239 GB | about 245 GB |
| 2.00 bpw | 257 GB | about 265 GB |
| 2.25 bpw | 274 GB | about 280 GB |
Runs on NVIDIA RTX 30, 40 and 50 series (Windows and Linux), AMD Radeon AI PRO R9700 (Linux), AMD Ryzen AI Max — Strix Halo
(Windows), and Apple Silicon Macs with 64 GB or more, text only ([Mac](docs/DEEPSEEK_MAC.md)). Setup downloads the files and the DeepSeek engine, then starts the model.
| | NVIDIA, older card or driver older than 580 | [`Strata-0.2.3-windows-cuda12.exe`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-windows-cuda12.exe) |
| | AMD Radeon RX 9000 series or Radeon AI PRO R9700 | [`Strata-0.2.3-windows-hip.exe`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-windows-hip.exe) |
| | AMD Radeon RX 7000 series or Ryzen AI Max (Strix Halo) | [`Strata-0.2.3-windows-hip-rdna3.exe`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-windows-hip-rdna3.exe) |
| | NVIDIA, older card or driver older than 580 | [`Strata-0.2.3-linux-nvidia-cuda12.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-nvidia-cuda12.AppImage) |
| | AMD Radeon RX 9000 series or Radeon AI PRO R9700 | [`Strata-0.2.3-linux-amd-rx9070.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-amd-rx9070.AppImage) |
| | AMD Radeon RX 7000 series | [`Strata-0.2.3-linux-amd-rx7900.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-amd-rx7900.AppImage) |
| | AMD Ryzen AI Max (Strix Halo) | [`Strata-0.2.3-linux-amd-strixhalo.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-amd-strixhalo.AppImage) |
| **Mac** (Apple Silicon, macOS 15 or newer) | built in | [`Strata-0.2.3-mac.dmg`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-mac.dmg) |
Install the app, choose a model, and start chatting. There is one installer for each system. On first start the app
downloads the engine for your graphics card. [Help with the installer](docs/INSTALL.md#the-app).
**If you are unsure:** for NVIDIA, take `cuda13` if your driver is version 580 or newer, otherwise `cuda12`. `cuda12` is
for older cards and older drivers. The `hip` and `amd` files are for AMD cards. RX 6000 cards have no installer of
their own: install `hip-rdna3` (Windows) or `amd-rx7900` (Linux), and on a PC with an RX 6000 card it fetches the RX 6000 engine by itself
(also `strata update`). That engine is built but not yet tested on a real RX 6000 card: see [AMD cards](docs/AMD_HIP.md).
The script checks your PC, downloads the Terminal, verifies its SHA-256 and unpacks it into your home folder. It needs no
administrator rights. On Windows, open a new terminal window afterwards. Then start Strata:
On Windows, open a new terminal afterwards. Then:
```sh
strata launch
```
It asks one question at a time: model, context, pictures, who may connect. Then it downloads the engine for your graphics
card (signed, and test-started before use) and the model, and starts serving. For scripts and AI agents, use flags:
```sh
strata launch --model q2_0 --yes --no-tui
```
To read the install script before it runs, download it first with `curl -fsSL <address> -o install.sh`, read it, then run
`sh install.sh`. All screens and flags: [docs/TERMINAL_LAUNCH.md](docs/TERMINAL_LAUNCH.md).
### Option C: let your AI set it up
Paste this into Claude Code, Cursor, Codex, GitHub Copilot or another coding assistant:
```text
Set up Strata on this PC for me: https://github.com/Niko1221/Strata - follow docs/AI_SETUP.md in that repository.
```
It checks your graphics card, RAM and disk, picks the model that fits, installs and starts it, and tells you how to
connect your apps. AI tools can also install, start and stop Strata through its [MCP server](docs/MCP_SERVER.md).
Prefer to build it yourself? See [Build from source](#build-from-source).
## Models
The app and `strata launch` recommend a model for your memory. It is one model in several sizes, compressed more or less.
Smaller sizes are faster. Larger sizes are a bit smarter.
### Low on memory? Try Gyro
[Gyro](https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF) (by agentionai) is the smallest way to run this model.
| Size | Memory it needs (RAM + VRAM) | Download | What it is |
| --- | ---: | ---: | --- |
| **Gyro-S** | 27.6 GB | 58.5 GB | For PCs that fit nothing else. Quality is lower than IQ2_XS. |
| **Gyro-M** | 35 GB | 92.0 GB | About IQ3_XXS quality with about 20% less GPU memory. It writes answers 16% (12 GB RTX 3060) to 21% (R9700) slower than Q2_0. |
Gyro runs on NVIDIA and AMD cards (AMD is experimental) on Intel Arc cards on Linux (tested on an Arc A750) and, from 0.2.2, on Apple Silicon Macs (Gyro-S on a 32 GB Mac mini: 9 tok/s, streaming from the SSD; [docs/MAC.md](docs/MAC.md#gyro-s-and-gyro-m-agentionais-trellis-experts)). Details and measurements:
[docs/GYRO.md](docs/GYRO.md). Start it with `strata launch --model gyro-s`.
### The other sizes
The tiers count your RAM plus your VRAM, less 5 GB that Windows and the system keep for themselves. Every other size stays a choice.
| Your memory (RAM + VRAM) | Take | Notes |
| --- | --- | --- |
| **32 GB** | **Gyro-S**, or the **[Coder](docs/MODELS.md#coder)** for code | The Coder needs 32 GB of RAM and keeps half of the experts, so it is weaker outside code |
| **40 GB** | **[Swift 1.5](docs/MODELS.md#swift-15)** IQ2_XS (39.2 GB), or **Gyro-M** (35 GB) | Swift thinks for a shorter time before it answers. Gyro-M is close to IQ3_XXS in quality, but it writes answers 16-21% slower than Q2_0 |
| **64 GB** | **IQ3_S** (54.8 GB), the original or Swift 1.5 | The slowest, the best: it matches the full model on the published tests |
| **96 GB or more** | Unsloth's **UD-IQ4_XS** (about 72 GB of RAM) to **Q8_0** (about 136 GB), 4 to 8 bit | Closer to the full model. With less RAM, Strata reads part of them from the SSD, which is much slower |
Less RAM than a row says can still work: Strata switches to its
[low-RAM mode](docs/MODELS.md#a-big-graphics-card-and-little-ram) by itself. Expect it to be slower, and much slower
with a small graphics card.
More choices:
- **[Swift 1.5](docs/MODELS.md#swift-15)** is a fine-tune that thinks for a shorter time before it answers, at about the same quality and RAM.
- **[NVFP4](docs/NVFP4.md)** is NVIDIA's 4-bit version for NVIDIA cards. It needs about 80 GB of RAM to keep every expert in RAM.
- **[Qwen3.6-35B-A3B](docs/QWEN36.md)** is a smaller, different model that reads pictures and can answer several requests at once.
- **[OrcaRouter's Uncensored IQ3_XXS](docs/MODELS.md#orcarouter-uncensored-iq3_xxs)** is set up by hand.
Every size, download and what fits where: [docs/MODELS.md](docs/MODELS.md).
## How fast is it?
Based on a 12 GB graphics card (RTX 5070, Ryzen 5 7600, 64 GB RAM). See results from other cards:
[speed of each model](docs/MODELS.md#how-fast-is-each-size), [community results](docs/COMMUNITY_BENCHMARKS.md).
| Model | Answers appear at | Reads your prompt at |
The bars use one scale for both tables. A card with more VRAM is faster: an RTX 3090 (24 GB) should write about 100-140
tokens per second. The engine version behind each number is in [DETAILS.md](docs/DETAILS.md#speed-measured).
## Using it
### In the app
The app has a **Chat**, a **Monitor** of the model and your GPU, CPU and RAM, a **Models** page and **Settings**. The chat
can work in a folder on your PC. In **Manual** mode it asks before it changes a file or runs a command. Type
`/loop 10m <what to do>` to repeat a message every 10 minutes while the chat is open.
The chat can search the web with DuckDuckGo (or Bing), with no account. You can add a search key under **Settings, Tools**.
Search is the one feature that sends something off your PC, the search words.
To run two models, turn on **Run multiple models** and give each model its own card of the same make.
<p align="center"><img src="docs/media/app-chat.png" width="900" alt="The Strata app: a chat where the agent built a small voxel pagoda garden, with the page open in the preview on the right"><br>
<sub>A chat in the Strata app: the agent wrote a small voxel pagoda garden as one HTML file, and the page opens in the preview next to the conversation</sub></p>
### In the Terminal
<p align="center"><img src="docs/media/terminal-monitor.png" width="900" alt="The Strata Terminal monitor overview while the model writes a reply"><br>
<sub><code>strata monitor</code> while the model writes a reply: speed, GPU, memory and the recent requests</sub></p>
strata update --check # is there a new engine? (strata update installs it)
strata doctor # check the installed engine files
```
`strata --help` lists everything. Starting and serving never download anything. Only the install, `strata launch` (when
you choose a model) and `strata update` do. To reach Strata from your phone or another PC, use
`strata launch --access lan`. It makes an API key (recommended; add `--no-api-key` to go without, with a warning). No browser opens unless
you add `--open`. Guide: [docs/TERMINAL_LAUNCH.md](docs/TERMINAL_LAUNCH.md). Whole client: [docs/TERMINAL.md](docs/TERMINAL.md).
### With AI agents
An agent installs Strata with the one-line command in Option B and runs
`strata launch --model NAME --yes --no-tui`. The steps, flags and exit codes are in [docs/AI_SETUP.md](docs/AI_SETUP.md),
and the repository rules are in [AGENTS.md](AGENTS.md).
### Connect your other apps
| App type | Base URL |
| --- | --- |
| OpenAI-compatible (most apps and agents) | `http://127.0.0.1:8080/v1` (any API key, any model name) |
| Anthropic API (Claude Code and others) | `http://127.0.0.1:8080`, without `/v1`. These apps add `/v1/messages` themselves. Claude Code: `ANTHROPIC_BASE_URL=http://127.0.0.1:8080` |
| OpenAI Responses API (Codex CLI) | `http://127.0.0.1:8080/v1/responses` ([setup](docs/DETAILS.md#the-responses-api-and-codex-cli)) |
`strata connect` prints the exact settings for OpenCode, Codex CLI, Claude Code, oh-my-pi, pi, Hermes Agent and OpenClaw.
`strata connect opencode --write` writes the file after asking and keeps a backup.
### Pictures
Pictures are on by default for models that can read them. Attach or paste a picture in the chat, or send it from your app.
The picture reader runs on the processor when your graphics card has less than 24 GB (a few seconds per picture, nothing
taken from the card's memory) and on the graphics card from 24 GB. The rule is the same for NVIDIA and AMD. Turn pictures
off per model, or choose the reader with `--vision off`, `--vision gpu` or `--vision cpu`.
The numbers behind the rule: on a Radeon AI PRO R9700 the card's reader took 57 ms for a 260-token picture and the
processor's took 3.1 s. On 16 GB cards (Tesla P100 and RTX A4000, Linux) the card's reader cost 3-5% of text speed,
because it holds about 1.4 GB that the model would use otherwise. On a Mac the reader runs on the Mac's GPU whatever its memory (0.45 s for a 260-token picture on an M4 with 32 GB, about 1.2 GB of the shared
memory); the processor reader needs about 50 s per picture there, so it runs only when you ask for it with `--vision cpu`.
### Good to know
- Choose the thinking level (**off, low, medium or high**) in the chat menu or in your app's "reasoning effort". Off is fastest.
- Strata answers one request at a time by default. For several at once, use `--parallel N` in `strata launch` or
`"parallel": 2` in the config ([BATCHING.md](docs/BATCHING.md)). On a 12 GB card this makes each answer slower.
- Strata reads the first message of a chat in full, about 1 minute per 30,000 tokens. Follow-up messages start in seconds.
- Where your chats are stored and the whole API: [docs/INSTALL.md](docs/INSTALL.md#where-things-are-stored), [docs/DETAILS.md](docs/DETAILS.md#using-it).
## Build from source
<details>
<summary>Windows, Linux and Mac from the repository</summary>
[Download Strata](https://github.com/Niko1221/Strata/archive/refs/heads/main.zip) and unzip it, or `git clone` it.
**Windows:** double-click `START-HERE.bat`. **Linux:** run `./setup.sh` in the Strata folder.
**Mac (Apple Silicon, macOS 15 or newer):** install Apple's Command Line Tools once with `xcode-select --install`
(Xcode itself is not needed). Open Terminal in the Strata folder and run `./setup.sh`. It builds the Mac engine on your
Mac (about 7 minutes), asks the same questions, downloads the model and starts it. Next time, `./setup.sh` starts it
right away. Which size fits your Mac: [docs/MAC.md](docs/MAC.md).
The installer finds your card and sets up the right engine. It asks which model and size, how much context (how much text
the model keeps in mind) and whether it should read pictures. Press Enter for the recommended answer. Then it downloads the
model (about 70 GB) and starts it. If the download stops, run it again and it continues. Your browser opens the Strata
page at `http://127.0.0.1:8080`.
> While the model starts, your PC can be slow or stop responding for 1-3 minutes (longest the first time). Strata loads
> 35-55 GB into your RAM and locks part of it for the graphics card. Wait, and don't close the window.
Next time, run `START-HERE.bat` (or `./setup.sh`) again. It starts right away and downloads nothing twice. Close its
window to stop the model. `UPDATE.bat` (`./update.sh`) updates Strata without starting it. Updating, Docker, several cards,
where the files go and every option: [docs/INSTALL.md](docs/INSTALL.md).
To measure your PC, run `BENCH.bat` (`./bench.sh`). It takes 5-10 minutes and writes a report you can share
**Optional: more speed on one card when the whole model fits in RAM.** Two opt-in settings by RubenTorney (#1904, #1905).
They help when you run one graphics card and your RAM holds every expert, for example Q2_0 or IQ3_XXS on a 64 GB PC. Open
`strata-<model>.json` in the Strata data folder ([where](docs/INSTALL.md#where-things-are-stored)) and add:
-`"--adapt-async", "1"` to the `"args"` list. The expert cache then updates on a helper thread instead of pausing the
answer. Measured on an RTX 5070 (Windows, Q2_0): +4-5% decode, +7-10% on code and long documents. Answers can differ
slightly from the default and stay the same from run to run.
-`"STRATA_PIN_BY_NEED": "1"` in the `"env"` block (Windows). Windows can lock only about half the RAM for the graphics
card, and this locks the experts the card needs first. Measured on the RTX 5070: +4.4% prompt reading.
```json
"args":[...,"--adapt-async","1"],
"env":{"STRATA_PIN_BY_NEED":"1"}
```
Restart Strata after the change. To undo, remove the two entries.
Follow the prompts to download and start a model. [Terminal guide](docs/TERMINAL_LAUNCH.md).
</details>
## How does it work?
<p align="center">
<img src="docs/media/terminal-monitor.png" width="900" alt="Strata terminal app showing model activity, speed, memory and recent requests">
<br><sub>Watch your model and hardware while it runs.</sub>
</p>
Models like this one usually run on servers with hundreds of gigabytes of graphics memory. Your graphics card has
12-24 GB. Strata makes the model fit by sharing the work across your whole PC. Picture a library.
<a id="build-from-source"></a>
- The model is a team of 24,576 small specialists ("experts"), like 24,576 reference books. Each word needs only 10 of them.
- **Your graphics card is the desk.** It is small and fast. The roughly 4,000 books used most often lie open on it.
- **Your RAM is the shelves in the reading room.** All 24,576 books are there. When a word needs a book that is not on the
desk, your processor reads it from the shelf at once, while the card keeps working.
- **Your SSD is the archive.** It holds a 29 GB lookup table. The model reads a few rows of it per word, so it does not
need to sit in RAM.
[Build from source](docs/SOURCE_INSTALL.md) · [Docker](docs/DOCKER.md) · [Full install guide](docs/INSTALL.md)
<p align="center"><img src="docs/media/how-it-works.svg" width="860" alt="The model's 24,576 experts: the busiest on the graphics card, all of them in RAM, a lookup table on the SSD"></p>
<a id="what-you-need"></a>
A small helper also guesses the next few words. The big model checks all of them at once and keeps the right ones. You get
the same answer, 1.6-1.8x sooner. Long texts are read in big pieces, up to 8,192 tokens at a time, at over 1,000 tokens
per second.
## Hardware
<p align="center"><img src="docs/media/guess-and-check.svg" width="860" alt="A small helper guesses the next words; the big model checks them all at once and keeps the right ones"></p>
and Unsloth. Strata uses parts of [llama.cpp / ggml](https://github.com/ggml-org/llama.cpp). All credits:
[docs/HOW_IT_WORKS.md](docs/HOW_IT_WORKS.md#credits). Strata is open source under the [MIT License](LICENSE). A few parts and
every model have their own licenses ([which ones](docs/HOW_IT_WORKS.md#license)).
## Help
## Contributing
Found a bug, or want to send a change or a benchmark? Read [CONTRIBUTING.md](CONTRIBUTING.md) first. It says what to
include so that others can check your work.
[All guides](docs) · [Troubleshooting](docs/TROUBLESHOOTING.md) · [Report a bug](https://github.com/Niko1221/Strata/issues/new/choose) ·
[Benchmarks](docs/COMMUNITY_BENCHMARKS.md) · [How it works](docs/HOW_IT_WORKS.md) · [Contribute](CONTRIBUTING.md)
## Support Strata
Strata is free and MIT-licensed. I build it and test it on whatever hardware I can get. If it runs well on your PC, a coffee helps
me keep testing new cards and fixing what you report. No pressure. Using it and reporting what you find helps just as much.
**Help keep local AI open to everyone.** A coffee helps fund development and testing on more hardware.
<a href="https://buymeacoffee.com/strataengine"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="50"></a>
<a href="https://buymeacoffee.com/strataengine"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me a Coffee: support Strata" height="48"></a>
[MIT license](LICENSE) · [Credits and model licenses](docs/HOW_IT_WORKS.md#credits) · [Privacy](docs/SAFETY.md)
<a id="option-c-let-your-ai-set-it-up"></a>
<sub>For AI tools and automated readers: <a href="AGENTS.md">technical reference</a> · <a href="docs/AI_SETUP.md">setup instructions</a>.</sub>
<a href="https://buymeacoffee.com/strataengine"><img src="https://img.shields.io/badge/Buy_me_a_coffee-FFDD00?logo=buymeacoffee&logoColor=000" alt="Wesprzyj Stratę na Buy Me a Coffee"></a>
**Bezpłatna, z otwartym kodem źródłowym i do uruchomienia u siebie.** Używaj Straty na swoim komputerze albo razem z ulubionymi narzędziami do programowania. Konto nie jest potrzebne.
## Modele
Najpierw polecamy **ISTA-DASLab**, potem **Unsloth** i innych dostawców — o ile zgodny model mieści się w pamięci.
| **Gyro · agentionai** | [Gyro-S, Gyro-M](https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF). Zużywa mniej pamięci, ale działa wolniej. |
W Windows otwórz potem nowe okno terminala. Następnie:
```sh
strata launch
```
Postępuj zgodnie z podpowiedziami, aby pobrać i uruchomić model. [Przewodnik po Terminalu](docs/TERMINAL_LAUNCH.md).
</details>
<p align="center">
<img src="docs/media/terminal-monitor.png" width="900" alt="Aplikacja Terminal Straty pokazująca aktywność modelu, szybkość, pamięć i ostatnie zapytania">
<br><sub>Obserwuj model i sprzęt w trakcie pracy.</sub>
</p>
<a id="build-from-source"></a>
[Budowanie ze źródeł](docs/SOURCE_INSTALL.md) · [Docker](docs/DOCKER.md) · [Pełny przewodnik instalacji](docs/INSTALL.md)
| **AMD Ryzen AI Max · Strix Halo** | [Konfiguracja Strix Halo](docs/STRIX_HALO.md) |
<a id="using-it"></a>
## Narzędzia
**Claude Code · Codex CLI · OpenCode · oh-my-pi · pi.dev · Hermes Agent · OpenClaw** i inne — przez API zgodne z OpenAI i Anthropic.
Uruchom **`strata connect`**, aby zobaczyć instrukcję konfiguracji dla swojego narzędzia. [Przewodnik po połączeniach](docs/TERMINAL_LAUNCH.md#connect-your-apps-strata-connect).
[Testy wydajności](docs/COMMUNITY_BENCHMARKS.md) · [Jak to działa](docs/HOW_IT_WORKS.md) · [Współtworzenie](CONTRIBUTING.md)
## Wesprzyj Stratę
**Pomóż zachować lokalną sztuczną inteligencję otwartą dla wszystkich.** Kawa pomaga finansować rozwój i testy na większej liczbie sprzętu.
<a href="https://buymeacoffee.com/strataengine"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me a Coffee: wesprzyj Stratę" height="48"></a>
[Licencja MIT](LICENSE) · [Podziękowania i licencje modeli](docs/HOW_IT_WORKS.md#credits) · [Prywatność](docs/SAFETY.md)
<a id="option-c-let-your-ai-set-it-up"></a>
<sub>Dla narzędzi AI i automatycznych czytników: <a href="AGENTS.md">dokumentacja techniczna</a> · <a href="docs/AI_SETUP.md">instrukcja instalacji</a>.</sub>
<p align="center"><a href="https://github.com/Niko1221/Strata/releases/download/v0.1.10/Pagoda.mp4"><img src="docs/media/pagoda-preview.webp" width="720" alt="A voxel pagoda garden that Strata's model wrote, running in the browser"></a><br>
<sub>Um jardim de pagodes em voxels, gerado com um único prompt, rodando numa RTX 5070 com o Strata (IQ3_S, contexto de 128K) ·
<a href="https://buymeacoffee.com/strataengine"><img src="https://img.shields.io/badge/Buy_me_a_coffee-FFDD00?logo=buymeacoffee&logoColor=000" alt="Apoie o Strata no Buy Me a Coffee"></a>
</p>
O Strata roda o **[Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)**, um modelo de 125 bilhões de parâmetros que
normalmente precisa de um servidor, num PC comum. Ele conversa, escreve código, lê imagens e funciona com os seus aplicativos e
agentes de programação. Suas conversas e arquivos ficam no seu PC.
**Primeira vez aqui?** Veja [o que você precisa](#o-que-você-precisa) e depois [instale o app](#opção-a-o-app-strata).
**Gratuito, de código aberto e seu para executar.** Use o Strata no desktop ou com suas ferramentas de programação favoritas. Não precisa de conta.
- **Roda num PC gamer.** 32 GB somando memória RAM e memória da placa de vídeo já bastam para começar.
- **Funciona com as suas ferramentas.** Ele fala as APIs da OpenAI e da Anthropic, então Claude Code, Codex CLI, OpenCode e outros agentes se conectam a ele.
- **Lê imagens.** Já vêm ativadas.
- **Gratuito.** Licença MIT, sem conta, sem nuvem.
- **Privado.** Sem telemetria. O que o Strata baixa, quando usa a internet e como confere os downloads está em [docs/SAFETY.md](docs/SAFETY.md).
## Modelos
Versão 0.2.3 · mecanismo 0.1.48
Recomendamos **primeiro a ISTA-DASLab**, depois a **Unsloth** e outros provedores, quando um modelo compatível cabe na memória.
No **aplicativo de desktop** e no **Terminal**. O que não cabe na memória é lido do seu SSD, por isso é preciso muito espaço livre em disco.
| Conjunto (IST-DAS) | Download | Espaço livre necessário |
| --- | ---: | ---: |
| 1.75 bpw | 239 GB | cerca de 245 GB |
| 2.00 bpw | 257 GB | cerca de 265 GB |
| 2.25 bpw | 274 GB | cerca de 280 GB |
Roda em NVIDIA RTX 30, 40 e 50 (Windows e Linux) e em AMD Radeon AI PRO R9700 (Linux). A instalação baixa os arquivos e o motor do DeepSeek e depois inicia o modelo.
Medido com o conjunto de 2.00 bpw no Strata 0.2.5. “Saída” é a velocidade da resposta; “Prompt” é a rapidez com que o Strata lê o seu texto.
| Hardware | Saída | Prompt |
| --- | ---: | ---: |
| RTX 5070 12 GB, 64 GB de RAM | 4 a 6 tokens/s | cerca de 110 tokens/s (4K) |
| RTX 3060 12 GB, 46 GB de RAM | 9 a 11 tokens/s | cerca de 155 tokens/s (4K) |
| 1x Radeon AI PRO R9700, 251 GB de RAM | 21 a 25 tokens/s | 320 a 360 tokens/s (8K) |
| 4x Radeon AI PRO R9700 | 74 a 77 tokens/s | 400 a 450 tokens/s (8K) |
[Memória, disco e detalhes](docs/DEEPSEEK41.md) · [Notas da versão](https://github.com/Niko1221/Strata/releases)
</details>
<a id="the-other-sizes"></a>
Não sabe o que cabe? [Escolha um modelo para o seu computador](docs/MODELS.md#pick-by-memory).
## Instalação
Escolha uma opção. Todas iniciam o mesmo mecanismo e usam os mesmos modelos.
<a id="option-a-the-strata-app"></a>
### Opção A: o app Strata
### Aplicativo de desktop
O jeito mais fácil. O app já traz o mecanismo: instale, escolha um modelo e converse. Baixe o arquivo para o seu sistema e a sua
placa de vídeo na [página de releases](https://github.com/Niko1221/Strata/releases/latest). Não sabe qual? Veja a explicação
logo abaixo da tabela.
**[Baixar o Strata para Windows, Linux ou Mac →](https://github.com/Niko1221/Strata/releases/latest)**
| Sistema | Sua placa de vídeo | Arquivo para baixar |
| --- | --- | --- |
| **Windows 10 / 11** | NVIDIA RTX, driver 580 ou mais novo | [`Strata-0.2.3-windows-cuda13.exe`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-windows-cuda13.exe) |
| | NVIDIA, placa mais antiga ou driver anterior ao 580 | [`Strata-0.2.3-windows-cuda12.exe`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-windows-cuda12.exe) |
| | AMD Radeon série RX 9000 ou Radeon AI PRO R9700 | [`Strata-0.2.3-windows-hip.exe`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-windows-hip.exe) |
| | AMD Radeon série RX 7000 ou Ryzen AI Max (Strix Halo) | [`Strata-0.2.3-windows-hip-rdna3.exe`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-windows-hip-rdna3.exe) |
| **Linux** | NVIDIA RTX, driver 580 ou mais novo | [`Strata-0.2.3-linux-nvidia-cuda13.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-nvidia-cuda13.AppImage) |
| | NVIDIA, placa mais antiga ou driver anterior ao 580 | [`Strata-0.2.3-linux-nvidia-cuda12.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-nvidia-cuda12.AppImage) |
| | AMD Radeon série RX 9000 ou Radeon AI PRO R9700 | [`Strata-0.2.3-linux-amd-rx9070.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-amd-rx9070.AppImage) |
| | AMD Radeon série RX 7000 | [`Strata-0.2.3-linux-amd-rx7900.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-amd-rx7900.AppImage) |
| | AMD Ryzen AI Max (Strix Halo) | [`Strata-0.2.3-linux-amd-strixhalo.AppImage`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-linux-amd-strixhalo.AppImage) |
| **Mac** (Apple Silicon, macOS 15 ou mais novo) | integrada | [`Strata-0.2.3-mac.dmg`](https://github.com/Niko1221/Strata/releases/latest/download/Strata-0.2.3-mac.dmg) |
Instale o aplicativo, escolha um modelo e comece a conversar. Há um instalador por sistema. Na primeira vez que abre, o aplicativo baixa o motor da sua placa de vídeo. [Ajuda com o instalador](docs/INSTALL.md#the-app).
**Se você estiver em dúvida:** para NVIDIA, use `cuda13` se o seu driver for a versão 580 ou mais nova, e `cuda12` nos outros casos. O `cuda12`
serve para placas e drivers mais antigos. Os arquivos `hip` e `amd` são para placas AMD. Outras placas AMD (RX 6000) ainda não têm app pronto: veja [Placas AMD](docs/AMD_HIP.md).
O script confere o seu PC, baixa o Terminal, verifica o SHA-256 e descompacta na sua pasta pessoal. Não precisa de direitos de
administrador. No Windows, abra uma nova janela do terminal depois. Então inicie o Strata:
No Windows, abra um novo terminal depois. Em seguida:
```sh
strata launch
```
Ele faz uma pergunta por vez: modelo, contexto, imagens, quem pode se conectar. Depois baixa o mecanismo para a sua placa de vídeo
(assinada e testada antes do uso) e o modelo, e começa a servir. Em scripts e agentes de IA, use flags:
```sh
strata launch --model q2_0 --yes --no-tui
```
Para ler o script de instalação antes de executá-lo, baixe-o primeiro com `curl -fsSL <address> -o install.sh`, leia e então rode
`sh install.sh`. Todas as telas e flags: [docs/TERMINAL_LAUNCH.md](docs/TERMINAL_LAUNCH.md).
### Opção C: deixe a sua IA configurar
Cole isto no Claude Code, Cursor, Codex, GitHub Copilot ou outro assistente de programação:
```text
Set up Strata on this PC for me: https://github.com/Niko1221/Strata - follow docs/AI_SETUP.md in that repository.
```
Ele confere a sua placa de vídeo, a RAM e o disco, escolhe o modelo que cabe, instala e inicia, e diz como conectar os seus
aplicativos. Ferramentas de IA também podem instalar, iniciar e parar o Strata pelo [servidor MCP](docs/MCP_SERVER.md) dele.
Prefere compilar por conta própria? Veja [Compilar a partir do código-fonte](#compilar-a-partir-do-código-fonte).
## Modelos
O app e o `strata launch` recomendam um modelo para a sua memória. É um único modelo em vários tamanhos, mais ou menos comprimidos.
Os tamanhos menores são mais rápidos. Os maiores são um pouco mais inteligentes.
### Pouca memória? Experimente o Gyro
O [Gyro](https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF) (da agentionai) é a menor forma de rodar este modelo.
| Tamanho | Memória necessária (RAM + VRAM) | Download | O que é |
| --- | ---: | ---: | --- |
| **Gyro-S** | 27,6 GB | 58,5 GB | Para PCs onde nenhum outro tamanho cabe. A qualidade é menor que a do IQ2_XS. |
| **Gyro-M** | 35 GB | 92,0 GB | Qualidade próxima à do IQ3_XXS, com cerca de 20% menos memória da GPU. Ele escreve as respostas 16% mais devagar que o Q2_0 numa RTX 3060 de 12 GB e 21% mais devagar numa R9700. |
O Gyro roda em placas NVIDIA e AMD (AMD é experimental) em placas Intel Arc no Linux (testado numa Arc A750) e, desde a 0.2.2, em Macs com Apple Silicon (Gyro-S num Mac mini de 32 GB: 9 tokens/s, com os especialistas lidos do SSD). Detalhes e medições:
[docs/GYRO.md](docs/GYRO.md). Inicie com `strata launch --model gyro-s`.
| **32 GB** | **Gyro-S**, ou o **[Coder](docs/MODELS.md#coder)** para código | O Coder precisa de 32 GB de RAM e mantém metade dos experts, então é mais fraco fora de código |
| **40 GB** | **[Swift 1.5](docs/MODELS.md#swift-15)** IQ2_XS (39,2 GB), ou **Gyro-M** (35 GB) | O Swift pensa por menos tempo antes de responder. O Gyro-M chega perto do IQ3_XXS em qualidade, mas escreve as respostas 16 a 21% mais devagar que o Q2_0 |
| **48 GB** | **IQ3_XXS** (47,0 GB) | Mais lento, melhor |
| **64 GB** | **IQ3_S** (54,8 GB), o original ou o Swift 1.5 | O mais lento, o melhor: empata com o modelo completo nos testes publicados |
| **96 GB ou mais** | **UD-IQ4_XS** da Unsloth (cerca de 72 GB de RAM) até **Q8_0** (cerca de 136 GB), de 4 a 8 bits | Mais perto do modelo completo. Com menos RAM, o Strata lê parte deles do SSD, o que é bem mais lento |
Com menos RAM do que a linha indica, ainda pode funcionar: o Strata muda sozinho para o
[modo de pouca RAM](docs/MODELS.md#a-big-graphics-card-and-little-ram). Espere mais lentidão, e muito mais com uma placa de vídeo pequena.
Mais opções:
- **[Swift 1.5](docs/MODELS.md#swift-15)** é um ajuste fino que pensa por menos tempo antes de responder, com qualidade e RAM parecidas.
- **[NVFP4](docs/NVFP4.md)** é a versão de 4 bits da NVIDIA para placas NVIDIA. Precisa de cerca de 80 GB de RAM para manter todos os experts na RAM.
- **[Qwen3.6-35B-A3B](docs/QWEN36.md)** é outro modelo, menor, que lê imagens e pode atender vários pedidos ao mesmo tempo.
- **[Uncensored IQ3_XXS da OrcaRouter](docs/MODELS.md#orcarouter-uncensored-iq3_xxs)** é configurado manualmente.
Todos os tamanhos, downloads e o que cabe em cada configuração: [docs/MODELS.md](docs/MODELS.md).
## Qual é a velocidade?
Com base numa placa de vídeo de 12 GB (RTX 5070, Ryzen 5 7600, 64 GB de RAM). Veja resultados de outras placas:
[velocidade de cada modelo](docs/MODELS.md#how-fast-is-each-size), [resultados da comunidade](docs/COMMUNITY_BENCHMARKS.md).
| Modelo | As respostas aparecem a | Lê o seu prompt a |
As barras usam a mesma escala nas duas tabelas. Uma placa com mais VRAM é mais rápida: uma RTX 3090 (24 GB) deve escrever cerca de 100 a 140
tokens por segundo. A versão do mecanismo por trás de cada número está em [DETAILS.md](docs/DETAILS.md#speed-measured).
## Como usar
### No app
O app tem um **Chat**, um **Monitor** do modelo e da sua GPU, CPU e RAM, uma página de **Modelos** e as **Configurações**. O chat
pode trabalhar numa pasta do seu PC. No modo **Manual**, ele pergunta antes de alterar um arquivo ou executar um comando. Digite
`/loop 10m <o que fazer>` para repetir uma mensagem a cada 10 minutos enquanto o chat estiver aberto.
O chat pode pesquisar na web com o DuckDuckGo (ou o Bing), sem conta. Você pode adicionar uma chave de busca em **Configurações, Ferramentas**.
A busca é o único recurso que envia algo para fora do seu PC: os termos da pesquisa.
Para rodar dois modelos, ative **Executar vários modelos** e dê a cada modelo a sua própria placa da mesma marca.
<p align="center"><img src="docs/media/app-chat.png" width="900" alt="The Strata app: a chat where the agent built a small voxel pagoda garden, with the page open in the preview on the right"><br>
<sub>Um chat no app Strata: o agente escreveu um pequeno jardim de pagode em voxels em um único arquivo HTML, e a página abre na pré-visualização ao lado da conversa</sub></p>
### No Terminal
<p align="center"><img src="docs/media/terminal-monitor.png" width="900" alt="The Strata Terminal monitor overview while the model writes a reply"><br>
<sub><code>strata monitor</code> enquanto o modelo responde: velocidade, GPU, memória e as solicitações recentes</sub></p>
```sh
strata launch # telas guiadas: modelo, contexto, imagens, acesso e depois servir
strata launch --model q2_0 --yes --no-tui # sem perguntas: serve, mostra o endereço e volta
strata launch --plan --model q2_0 # mostra o que aconteceria; nada muda
strata launch --status # o que está sendo servido agora
Um agente instala o Strata com o comando de uma linha da Opção B e executa
`strata launch --model NAME --yes --no-tui`. Os passos, as flags e os códigos de saída estão em [docs/AI_SETUP.md](docs/AI_SETUP.md),
e as regras do repositório estão em [AGENTS.md](AGENTS.md).
### Conecte os seus outros aplicativos
| Tipo de aplicativo | URL base |
| --- | --- |
| Compatível com OpenAI (a maioria dos apps e agentes) | `http://127.0.0.1:8080/v1` (qualquer chave de API, qualquer nome de modelo) |
| API da Anthropic (Claude Code e outros) | `http://127.0.0.1:8080`, sem `/v1`. Esses aplicativos adicionam `/v1/messages` sozinhos. Claude Code: `ANTHROPIC_BASE_URL=http://127.0.0.1:8080` |
| OpenAI Responses API (Codex CLI) | `http://127.0.0.1:8080/v1/responses` ([configuração](docs/DETAILS.md#the-responses-api-and-codex-cli)) |
`strata connect` mostra as configurações exatas para OpenCode, Codex CLI, Claude Code, oh-my-pi, pi, Hermes Agent e OpenClaw.
`strata connect opencode --write` grava o arquivo depois de perguntar e mantém um backup.
### Imagens
As imagens vêm ativadas nos modelos que sabem lê-las. Anexe ou cole uma imagem no chat, ou envie pelo seu aplicativo.
O leitor de imagens roda no processador quando a sua placa de vídeo tem menos de 24 GB (alguns segundos por imagem, sem tirar nada
da memória da placa) e roda na placa de vídeo a partir de 24 GB. A regra é a mesma para NVIDIA e AMD. Desative as imagens
por modelo, ou escolha o leitor com `--vision off`, `--vision gpu` ou `--vision cpu`.
Os números por trás da regra: numa Radeon AI PRO R9700, o leitor da placa levou 57 ms para uma imagem de 260 tokens e o do
processador levou 3,1 s. Em placas de 16 GB (Tesla P100 e RTX A4000, Linux), o leitor da placa custou de 3 a 5% da velocidade do texto,
porque ocupa cerca de 1,4 GB que o modelo usaria de outra forma. No Mac, a GPU faz a leitura, seja qual for a memória (0,45 s para uma imagem de 260 tokens em um M4 de 32 GB, cerca de 1,2 GB da memória compartilhada);
o processador leva cerca de 50 s por imagem, então só lê com `--vision cpu`.
### Bom saber
- Escolha o nível de raciocínio (**off, low, medium ou high**) no menu do chat ou no "reasoning effort" do seu aplicativo. Off é o mais rápido.
- Por padrão, o Strata responde a um pedido por vez. Para vários ao mesmo tempo, use `--parallel N` no `strata launch` ou
`"parallel": 2` na configuração ([BATCHING.md](docs/BATCHING.md)). Numa placa de 12 GB, isso deixa cada resposta mais lenta.
- O Strata lê a primeira mensagem de uma conversa por inteiro, cerca de 1 minuto a cada 30.000 tokens. As mensagens seguintes começam em segundos.
- Onde ficam as suas conversas e a API completa: [docs/INSTALL.md](docs/INSTALL.md#where-things-are-stored), [docs/DETAILS.md](docs/DETAILS.md#using-it).
## Compilar a partir do código-fonte
<details>
<summary>Windows, Linux e Mac a partir do repositório</summary>
[Baixe o Strata](https://github.com/Niko1221/Strata/archive/refs/heads/main.zip) e descompacte, ou use `git clone`.
**Windows:** dê dois cliques em `START-HERE.bat`. **Linux:** execute `./setup.sh` na pasta do Strata.
**Mac (Apple Silicon, macOS 15 ou mais novo):** instale uma vez as Command Line Tools da Apple com `xcode-select --install`
(o Xcode em si não é necessário). Abra o Terminal na pasta do Strata e execute `./setup.sh`. Ele compila o mecanismo do Mac no seu
Mac (cerca de 7 minutos), faz as mesmas perguntas, baixa o modelo e inicia. Da próxima vez, `./setup.sh` inicia
na hora. Qual tamanho cabe no seu Mac: [docs/MAC.md](docs/MAC.md).
O instalador encontra a sua placa e configura o mecanismo certo. Ele pergunta qual modelo e tamanho, quanto contexto (quanto texto
o modelo mantém em mente) e se deve ler imagens. Pressione Enter para aceitar a resposta recomendada. Depois ele baixa o
modelo (cerca de 70 GB) e inicia. Se o download parar, execute de novo e ele continua. O navegador abre a página do Strata em
`http://127.0.0.1:8080`.
> Enquanto o modelo inicia, o seu PC pode ficar lento ou parar de responder por 1 a 3 minutos (mais tempo na primeira vez). O Strata carrega
> de 35 a 55 GB na sua RAM e trava parte dela para a placa de vídeo. Espere e não feche a janela.
Da próxima vez, execute `START-HERE.bat` (ou `./setup.sh`) de novo. Ele inicia na hora e não baixa nada duas vezes. Feche a janela
dele para parar o modelo. `UPDATE.bat` (`./update.sh`) atualiza o Strata sem iniciá-lo. Atualização, Docker, várias placas,
onde ficam os arquivos e todas as opções: [docs/INSTALL.md](docs/INSTALL.md).
Para medir o seu PC, execute `BENCH.bat` (`./bench.sh`). Leva de 5 a 10 minutos e gera um relatório que você pode compartilhar
**Opcional: mais velocidade numa placa quando o modelo inteiro cabe na RAM.** Duas configurações opcionais do RubenTorney (#1904, #1905).
Elas ajudam quando você usa uma placa de vídeo e a sua RAM comporta todos os experts, por exemplo Q2_0 ou IQ3_XXS num PC de 64 GB. Abra
`strata-<model>.json` na pasta de dados do Strata ([onde fica](docs/INSTALL.md#where-things-are-stored)) e adicione:
-`"--adapt-async", "1"` à lista `"args"`. O cache de experts passa a ser atualizado numa thread auxiliar, em vez de pausar a
resposta. Medido numa RTX 5070 (Windows, Q2_0): +4 a 5% na geração, +7 a 10% em código e documentos longos. As respostas podem diferir
um pouco do padrão e continuam iguais de uma execução para outra.
-`"STRATA_PIN_BY_NEED": "1"` no bloco `"env"` (Windows). O Windows só consegue travar cerca de metade da RAM para a placa
de vídeo, e isso trava primeiro os experts de que a placa precisa. Medido na RTX 5070: +4,4% na leitura do prompt.
```json
"args":[...,"--adapt-async","1"],
"env":{"STRATA_PIN_BY_NEED":"1"}
```
Reinicie o Strata depois da alteração. Para desfazer, remova as duas entradas.
Siga as instruções para baixar e iniciar um modelo. [Guia do Terminal](docs/TERMINAL_LAUNCH.md).
</details>
## Como funciona?
<p align="center">
<img src="docs/media/terminal-monitor.png" width="900" alt="Aplicativo de Terminal do Strata mostrando a atividade do modelo, a velocidade, a memória e as solicitações recentes">
<br><sub>Acompanhe seu modelo e seu hardware enquanto ele roda.</sub>
</p>
Modelos como este costumam rodar em servidores com centenas de gigabytes de memória de vídeo. A sua placa de vídeo tem
12 a 24 GB. O Strata faz o modelo caber dividindo o trabalho entre o PC inteiro. Imagine uma biblioteca.
<a id="build-from-source"></a>
- O modelo é uma equipe de 24.576 pequenos especialistas ("experts"), como 24.576 livros de consulta. Cada palavra precisa de apenas 10 deles.
- **A sua placa de vídeo é a mesa.** Ela é pequena e rápida. Os cerca de 4.000 livros mais usados ficam abertos nela.
- **A sua RAM são as estantes da sala de leitura.** Os 24.576 livros estão lá. Quando uma palavra precisa de um livro que não está na
mesa, o seu processador lê da estante na hora, enquanto a placa continua trabalhando.
- **O seu SSD é o arquivo.** Ele guarda uma tabela de consulta de 29 GB. O modelo lê algumas linhas dela por palavra, então ela não
precisa ficar na RAM.
[Compilar a partir do código-fonte](docs/SOURCE_INSTALL.md) · [Docker](docs/DOCKER.md) · [Guia de instalação completo](docs/INSTALL.md)
<p align="center"><img src="docs/media/how-it-works.svg" width="860" alt="The model's 24,576 experts: the busiest on the graphics card, all of them in RAM, a lookup table on the SSD"></p>
<a id="what-you-need"></a>
Um pequeno ajudante também adivinha as próximas palavras. O modelo grande confere todas de uma vez e fica com as certas. Você recebe
a mesma resposta, de 1,6 a 1,8 vez mais rápido. Textos longos são lidos em blocos grandes, de até 8.192 tokens por vez, a mais de 1.000 tokens
por segundo.
## Hardware
<p align="center"><img src="docs/media/guess-and-check.svg" width="860" alt="A small helper guesses the next words; the big model checks them all at once and keeps the right ones"></p>
O Strata é gratuito e tem licença MIT. Eu o desenvolvo e testo em todo hardware que consigo arranjar. Se ele roda bem no seu PC, um café me ajuda
a continuar testando placas novas e corrigindo o que você relatar. Sem pressão. Usar e relatar o que encontrar ajuda tanto quanto.
**Ajude a manter a IA local aberta para todos.** Um café ajuda a financiar o desenvolvimento e os testes em mais hardware.
<a href="https://buymeacoffee.com/strataengine"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="50"></a>
<a href="https://buymeacoffee.com/strataengine"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me a Coffee: apoie o Strata" height="48"></a>
Quer patrocinar o Strata? Fale comigo pelo [GitHub](https://github.com/Niko1221).
[Licença MIT](LICENSE) · [Créditos e licenças dos modelos](docs/HOW_IT_WORKS.md#credits) · [Privacidade](docs/SAFETY.md)
<a id="option-c-let-your-ai-set-it-up"></a>
<sub>Para ferramentas de IA e leitores automáticos: <a href="AGENTS.md">referência técnica</a> · <a href="docs/AI_SETUP.md">instruções de instalação</a>.</sub>
<p align="center"><a href="https://github.com/Niko1221/Strata/releases/download/v0.1.10/Pagoda.mp4"><img src="docs/media/pagoda-preview.webp" width="720" alt="A voxel pagoda garden that Strata's model wrote, running in the browser"></a><br>
<a href="https://buymeacoffee.com/strataengine"><img src="https://img.shields.io/badge/Buy_me_a_coffee-FFDD00?logo=buymeacoffee&logoColor=000" alt="在 Buy Me a Coffee 上支持 Strata"></a>
<p align="center"><img src="docs/media/app-chat.png" width="900" alt="The Strata app: a chat where the agent built a small voxel pagoda garden, with the page open in the preview on the right"><br>
<sub>Strata 应用中的一次对话：智能体用一个 HTML 文件写出了一个小型体素宝塔花园，页面在对话旁的预览中打开</sub></p>
### 在终端里
<p align="center"><img src="docs/media/terminal-monitor.png" width="900" alt="The Strata Terminal monitor overview while the model writes a reply"><br>
<p align="center"><img src="docs/media/how-it-works.svg" width="860" alt="The model's 24,576 experts: the busiest on the graphics card, all of them in RAM, a lookup table on the SSD"></p>
<p align="center"><img src="docs/media/guess-and-check.svg" width="860" alt="A small helper guesses the next words; the big model checks them all at once and keeps the right ones"></p>
Strata 免费，采用 MIT 许可证。我在能拿到的各种硬件上开发和测试它。如果它在你的电脑上跑得不错，请我喝杯咖啡，能帮我继续测试新显卡、
修复你们反馈的问题。不用有压力。使用它并反馈你的发现，同样很有帮助。
**帮助本地 AI 保持开放。** 一杯咖啡能帮助我们在更多硬件上开发和测试。
<a href="https://buymeacoffee.com/strataengine"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="50"></a>
<a href="https://buymeacoffee.com/strataengine"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me a Coffee：支持 Strata" height="48"></a>
{"core":{"default_permission":{"identifier":"default","description":"Default core plugins set.","permissions":["core:path:default","core:event:default","core:window:default","core:webview:default","core:app:default","core:image:default","core:resources:default","core:menu:default","core:tray:default"]},"permissions":{},"permission_sets":{},"global_scope_schema":null},"core:app":{"default_permission":{"identifier":"default","description":"Default permissions for the plugin.","permissions":["allow-version","allow-name","allow-tauri-version","allow-identifier","allow-bundle-type","allow-register-listener","allow-remove-listener","allow-supports-multiple-windows"]},"permissions":{"allow-app-hide":{"identifier":"allow-app-hide","description":"Enables the app_hide command without any pre-configured scope.","commands":{"allow":["app_hide"],"deny":[]}},"allow-app-show":{"identifier":"allow-app-show","description":"Enables the app_show command without any pre-configured scope.","commands":{"allow":["app_show"],"deny":[]}},"allow-bundle-type":{"identifier":"allow-bundle-type","description":"Enables the bundle_type command without any pre-configured scope.","commands":{"allow":["bundle_type"],"deny":[]}},"allow-default-window-icon":{"identifier":"allow-default-window-icon","description":"Enables the default_window_icon command without any pre-configured scope.","commands":{"allow":["default_window_icon"],"deny":[]}},"allow-exit":{"identifier":"allow-exit","description":"Enables the exit command without any pre-configured scope.","commands":{"allow":["exit"],"deny":[]}},"allow-fetch-data-store-identifiers":{"identifier":"allow-fetch-data-store-identifiers","description":"Enables the fetch_data_store_identifiers command without any pre-configured scope.","commands":{"allow":["fetch_data_store_identifiers"],"deny":[]}},"allow-identifier":{"identifier":"allow-identifier","description":"Enables the identifier command without any pre-configured scope.","commands":{"allow":["identifier"],"deny":[]}},"allow-name":{"identifier":"allow-name","description":"Enables the name command without any pre-configured scope.","commands":{"allow":["name"],"deny":[]}},"allow-register-listener":{"identifier":"allow-register-listener","description":"Enables the register_listener command without any pre-configured scope.","commands":{"allow":["register_listener"],"deny":[]}},"allow-remove-data-store":{"identifier":"allow-remove-data-store","description":"Enables the remove_data_store command without any pre-configured scope.","commands":{"allow":["remove_data_store"],"deny":[]}},"allow-remove-listener":{"identifier":"allow-remove-listener","description":"Enables the remove_listener command without any pre-configured scope.","commands":{"allow":["remove_listener"],"deny":[]}},"allow-set-app-theme":{"identifier":"allow-set-app-theme","description":"Enables the set_app_theme command without any pre-configured scope.","commands":{"allow":["set_app_theme"],"deny":[]}},"allow-set-dock-visibility":{"identifier":"allow-set-dock-visibility","description":"Enables the set_dock_visibility command without any pre-configured scope.","commands":{"allow":["set_dock_visibility"],"deny":[]}},"allow-supports-multiple-windows":{"identifier":"allow-supports-multiple-windows","description":"Enables the supports_multiple_windows command without any pre-configured scope.","commands":{"allow":["supports_multiple_windows"],"deny":[]}},"allow-tauri-version":{"identifier":"allow-tauri-version","description":"Enables the tauri_version command without any pre-configured scope.","commands":{"allow":["tauri_version"],"deny":[]}},"allow-version":{"identifier":"allow-version","description":"Enables the version command without any pre-configured scope.","commands":{"allow":["version"],"deny":[]}},"deny-app-hide":{"identifier":"deny-app-hide","description":"Denies the app_hide command without any pre-configured scope.","commands":{"allow":[],"deny":["app_hide"]}},"deny-app-show":{"identifier":"deny-app-show","description":"Denies the app_show command without any pre-configured scope.","commands":{"allow":[],"deny":["app_show"]}},"deny-bundle-type":{"identifier":"deny-bundle-type","description":"Denies the bundle_type command without any pre-configured scope.","commands":{"allow":[],"deny":["bundle_type"]}},"deny-default-window-icon":{"identifier":"deny-default-window-icon","description":"Denies the default_window_icon command without any pre-configured scope.","commands":{"allow":[],"deny":["default_window_icon"]}},"deny-exit":{"identifier":"deny-exit","description":"Denies the exit command without any pre-configured scope.","commands":{"allow":[],"deny":["exit"]}},"deny-fetch-data-store-identifiers":{"identifier":"deny-fetch-data-store-identifiers","description":"Denies the fetch_data_store_identifiers command without any pre-configured scope.","commands":{"allow":[],"deny":["fetch_data_store_identifiers"]}},"deny-identifier":{"identifier":"deny-identifier","description":"DenLine truncated
{"core":{"default_permission":{"identifier":"default","description":"Default core plugins set.","permissions":["core:path:default","core:event:default","core:window:default","core:webview:default","core:app:default","core:image:default","core:resources:default","core:menu:default","core:tray:default"]},"permissions":{},"permission_sets":{},"global_scope_schema":null},"core:app":{"default_permission":{"identifier":"default","description":"Default permissions for the plugin.","permissions":["allow-version","allow-name","allow-tauri-version","allow-identifier","allow-bundle-type","allow-register-listener","allow-remove-listener","allow-supports-multiple-windows"]},"permissions":{"allow-app-hide":{"identifier":"allow-app-hide","description":"Enables the app_hide command without any pre-configured scope.","commands":{"allow":["app_hide"],"deny":[]}},"allow-app-show":{"identifier":"allow-app-show","description":"Enables the app_show command without any pre-configured scope.","commands":{"allow":["app_show"],"deny":[]}},"allow-bundle-type":{"identifier":"allow-bundle-type","description":"Enables the bundle_type command without any pre-configured scope.","commands":{"allow":["bundle_type"],"deny":[]}},"allow-default-window-icon":{"identifier":"allow-default-window-icon","description":"Enables the default_window_icon command without any pre-configured scope.","commands":{"allow":["default_window_icon"],"deny":[]}},"allow-exit":{"identifier":"allow-exit","description":"Enables the exit command without any pre-configured scope.","commands":{"allow":["exit"],"deny":[]}},"allow-fetch-data-store-identifiers":{"identifier":"allow-fetch-data-store-identifiers","description":"Enables the fetch_data_store_identifiers command without any pre-configured scope.","commands":{"allow":["fetch_data_store_identifiers"],"deny":[]}},"allow-identifier":{"identifier":"allow-identifier","description":"Enables the identifier command without any pre-configured scope.","commands":{"allow":["identifier"],"deny":[]}},"allow-name":{"identifier":"allow-name","description":"Enables the name command without any pre-configured scope.","commands":{"allow":["name"],"deny":[]}},"allow-register-listener":{"identifier":"allow-register-listener","description":"Enables the register_listener command without any pre-configured scope.","commands":{"allow":["register_listener"],"deny":[]}},"allow-remove-data-store":{"identifier":"allow-remove-data-store","description":"Enables the remove_data_store command without any pre-configured scope.","commands":{"allow":["remove_data_store"],"deny":[]}},"allow-remove-listener":{"identifier":"allow-remove-listener","description":"Enables the remove_listener command without any pre-configured scope.","commands":{"allow":["remove_listener"],"deny":[]}},"allow-set-app-theme":{"identifier":"allow-set-app-theme","description":"Enables the set_app_theme command without any pre-configured scope.","commands":{"allow":["set_app_theme"],"deny":[]}},"allow-set-dock-visibility":{"identifier":"allow-set-dock-visibility","description":"Enables the set_dock_visibility command without any pre-configured scope.","commands":{"allow":["set_dock_visibility"],"deny":[]}},"allow-supports-multiple-windows":{"identifier":"allow-supports-multiple-windows","description":"Enables the supports_multiple_windows command without any pre-configured scope.","commands":{"allow":["supports_multiple_windows"],"deny":[]}},"allow-tauri-version":{"identifier":"allow-tauri-version","description":"Enables the tauri_version command without any pre-configured scope.","commands":{"allow":["tauri_version"],"deny":[]}},"allow-version":{"identifier":"allow-version","description":"Enables the version command without any pre-configured scope.","commands":{"allow":["version"],"deny":[]}},"deny-app-hide":{"identifier":"deny-app-hide","description":"Denies the app_hide command without any pre-configured scope.","commands":{"allow":[],"deny":["app_hide"]}},"deny-app-show":{"identifier":"deny-app-show","description":"Denies the app_show command without any pre-configured scope.","commands":{"allow":[],"deny":["app_show"]}},"deny-bundle-type":{"identifier":"deny-bundle-type","description":"Denies the bundle_type command without any pre-configured scope.","commands":{"allow":[],"deny":["bundle_type"]}},"deny-default-window-icon":{"identifier":"deny-default-window-icon","description":"Denies the default_window_icon command without any pre-configured scope.","commands":{"allow":[],"deny":["default_window_icon"]}},"deny-exit":{"identifier":"deny-exit","description":"Denies the exit command without any pre-configured scope.","commands":{"allow":[],"deny":["exit"]}},"deny-fetch-data-store-identifiers":{"identifier":"deny-fetch-data-store-identifiers","description":"Denies the fetch_data_store_identifiers command without any pre-configured scope.","commands":{"allow":[],"deny":["fetch_data_store_identifiers"]}},"deny-identifier":{"identifier":"deny-identifier","description":"DenLine truncated
"markdownDescription":"Denies the unminimize command without any pre-configured scope."
},
{
"description":"This permission set configures which\nnotification features are by default exposed.\n\n#### Granted Permissions\n\nIt allows all notification related features.\n\n\n#### This default permission set includes:\n\n- `allow-is-permission-granted`\n- `allow-request-permission`\n- `allow-notify`\n- `allow-register-action-types`\n- `allow-register-listener`\n- `allow-remove-listener`\n- `allow-cancel`\n- `allow-get-pending`\n- `allow-remove-active`\n- `allow-get-active`\n- `allow-check-permissions`\n- `allow-show`\n- `allow-batch`\n- `allow-list-channels`\n- `allow-delete-channel`\n- `allow-create-channel`\n- `allow-permission-state`",
"type":"string",
"const":"notification:default",
"markdownDescription":"This permission set configures which\nnotification features are by default exposed.\n\n#### Granted Permissions\n\nIt allows all notification related features.\n\n\n#### This default permission set includes:\n\n- `allow-is-permission-granted`\n- `allow-request-permission`\n- `allow-notify`\n- `allow-register-action-types`\n- `allow-register-listener`\n- `allow-remove-listener`\n- `allow-cancel`\n- `allow-get-pending`\n- `allow-remove-active`\n- `allow-get-active`\n- `allow-check-permissions`\n- `allow-show`\n- `allow-batch`\n- `allow-list-channels`\n- `allow-delete-channel`\n- `allow-create-channel`\n- `allow-permission-state`"
},
{
"description":"Enables the batch command without any pre-configured scope.",
"type":"string",
"const":"notification:allow-batch",
"markdownDescription":"Enables the batch command without any pre-configured scope."
},
{
"description":"Enables the cancel command without any pre-configured scope.",
"type":"string",
"const":"notification:allow-cancel",
"markdownDescription":"Enables the cancel command without any pre-configured scope."
},
{
"description":"Enables the check_permissions command without any pre-configured scope.",
"type":"string",
"const":"notification:allow-check-permissions",
"markdownDescription":"Enables the check_permissions command without any pre-configured scope."
},
{
"description":"Enables the create_channel command without any pre-configured scope.",
"type":"string",
"const":"notification:allow-create-channel",
"markdownDescription":"Enables the create_channel command without any pre-configured scope."
},
{
"description":"Enables the delete_channel command without any pre-configured scope.",
"type":"string",
"const":"notification:allow-delete-channel",
"markdownDescription":"Enables the delete_channel command without any pre-configured scope."
},
{
"description":"Enables the get_active command without any pre-configured scope.",
"type":"string",
"const":"notification:allow-get-active",
"markdownDescription":"Enables the get_active command without any pre-configured scope."
},
{
"description":"Enables the get_pending command without any pre-configured scope.",
"type":"string",
"const":"notification:allow-get-pending",
"markdownDescription":"Enables the get_pending command without any pre-configured scope."
},
{
"description":"Enables the is_permission_granted command without any pre-configured scope.",
"markdownDescription":"Denies the register_action_types command without any pre-configured scope."
},
{
"description":"Denies the register_listener command without any pre-configured scope.",
"type":"string",
"const":"notification:deny-register-listener",
"markdownDescription":"Denies the register_listener command without any pre-configured scope."
},
{
"description":"Denies the remove_active command without any pre-configured scope.",
"type":"string",
"const":"notification:deny-remove-active",
"markdownDescription":"Denies the remove_active command without any pre-configured scope."
},
{
"description":"Denies the remove_listener command without any pre-configured scope.",
"type":"string",
"const":"notification:deny-remove-listener",
"markdownDescription":"Denies the remove_listener command without any pre-configured scope."
},
{
"description":"Denies the request_permission command without any pre-configured scope.",
"type":"string",
"const":"notification:deny-request-permission",
"markdownDescription":"Denies the request_permission command without any pre-configured scope."
},
{
"description":"Denies the show command without any pre-configured scope.",
"type":"string",
"const":"notification:deny-show",
"markdownDescription":"Denies the show command without any pre-configured scope."
},
{
"description":"This permission set configures which\nshell functionality is exposed by default.\n\n#### Granted Permissions\n\nIt allows to use the `open` functionality with a reasonable\nscope pre-configured. It will allow opening `http(s)://`,\n`tel:` and `mailto:` links.\n\n#### This default permission set includes:\n\n- `allow-open`",
"type":"string",
@@ -2491,6 +2701,60 @@
"type":"string",
"const":"shell:deny-stdin-write",
"markdownDescription":"Denies the stdin_write command without any pre-configured scope."
},
{
"description":"This permission set configures which kind of\nupdater functions are exposed to the frontend.\n\n#### Granted Permissions\n\nThe full workflow from checking for updates to installing them\nis enabled.\n\n\n#### This default permission set includes:\n\n- `allow-check`\n- `allow-download`\n- `allow-install`\n- `allow-download-and-install`",
"type":"string",
"const":"updater:default",
"markdownDescription":"This permission set configures which kind of\nupdater functions are exposed to the frontend.\n\n#### Granted Permissions\n\nThe full workflow from checking for updates to installing them\nis enabled.\n\n\n#### This default permission set includes:\n\n- `allow-check`\n- `allow-download`\n- `allow-install`\n- `allow-download-and-install`"
},
{
"description":"Enables the check command without any pre-configured scope.",
"type":"string",
"const":"updater:allow-check",
"markdownDescription":"Enables the check command without any pre-configured scope."
},
{
"description":"Enables the download command without any pre-configured scope.",
"type":"string",
"const":"updater:allow-download",
"markdownDescription":"Enables the download command without any pre-configured scope."
},
{
"description":"Enables the download_and_install command without any pre-configured scope.",
"type":"string",
"const":"updater:allow-download-and-install",
"markdownDescription":"Enables the download_and_install command without any pre-configured scope."
},
{
"description":"Enables the install command without any pre-configured scope.",
"type":"string",
"const":"updater:allow-install",
"markdownDescription":"Enables the install command without any pre-configured scope."
},
{
"description":"Denies the check command without any pre-configured scope.",
"type":"string",
"const":"updater:deny-check",
"markdownDescription":"Denies the check command without any pre-configured scope."
},
{
"description":"Denies the download command without any pre-configured scope.",
"type":"string",
"const":"updater:deny-download",
"markdownDescription":"Denies the download command without any pre-configured scope."
},
{
"description":"Denies the download_and_install command without any pre-configured scope.",
"type":"string",
"const":"updater:deny-download-and-install",
"markdownDescription":"Denies the download_and_install command without any pre-configured scope."
},
{
"description":"Denies the install command without any pre-configured scope.",
"type":"string",
"const":"updater:deny-install",
"markdownDescription":"Denies the install command without any pre-configured scope."
; Stop what Strata itself started from the install folder (the app, its host, the node runtime, the engine) before files are removed:
; a running copy keeps its files open, so the folder stayed behind after use. Only processes whose image path is inside INSTDIR are
; stopped; this uninstaller's own process and any program elsewhere are left alone.
!macroStrataStopOwnProcessesINSTALLDIR
System::Call'kernel32::SetEnvironmentVariableW(w "STRATA_INSTDIR", w "${INSTALLDIR}")'
System::Call'kernel32::GetCurrentProcessId()i.r9'
System::Call'kernel32::SetEnvironmentVariableW(w "STRATA_SELFPID", w "$9")'
; the installer is a 32-bit program, so plain powershell.exe is the 32-bit one, which cannot read the image path of a 64-bit process: use the 64-bit one
; The app writes into its own folders while it runs (Python bytecode caches under control\, for one), and the file-by-file delete below does not know them,
; so those folders stayed behind with a few dozen files. These four folders hold only what the installer put there and what the app made there;
; user data is never in them (it lives in the data folder), and the install folder itself is still removed with a plain RMDir.
The built UI is included. To rebuild after source edits, run `npm run build`. `preview:ui` uses isolated simulated data; `npm run dev` connects the normal host.
The previous project README is retained in [docs/README-historical.md](docs/README-historical.md). It contains older implementation status and is historical reference.
## Languages
The app speaks English, Chinese, Japanese, German, French, Spanish, Russian, Portuguese and Korean. The English text is the key: `src/locales/en.json`
(with `fix-en.json`, `tune-en.json` and `tools-en.json`) holds every string, and each language has the same four files (`fr.json`, `fix-fr.json`, ...).
- **A new string.** Write it with `tr("...")`, then add it to the English catalog (`node scripts/append-localization.mjs`, or the next `fx` number in
`scripts/make-fix-locales.py`). The language tests fail until every language has it.
- **French.** `python scripts/fr_sync.py missing todo.tsv` lists the English strings that have no French yet, one per line (`id`, a tab, the text).
Translate the second column ("vous", short plain words, one next step; a lone `=` keeps the English), then `python scripts/fr_sync.py apply todo.tsv` merges it
(it also sets the French spacing before `: ; ? !` and the typographic apostrophe) and `python scripts/fr_sync.py check` confirms nothing is missing.
- **Screenshots in a language.** `npm run build`, then `node scripts/language-shots.mjs fr <folder> models settings setup` (hermetic preview host, scratch profile, headless Chrome).
"""Writes src/locales/fix-<lang>.json: the interface strings that had no key in the locale files (#3479, #3825) and the strings of the 0.2.4.3 bug fixes.
Run from apps/host: python scripts/make-fix-locales.py
The keys are fx01.. in this order; fix-en.json is the English source. Placeholders ({0}, {1}) must be kept in every language."""
importjson
frompathlibimportPath
EN=[
"will stop running and be removed.",
"Downloaded, not loaded. Open Models to load it.",
"{0} lines added, {1} removed",
"Edited {0} files",
"Drop files to add them",
"Pictures and text files",
"Keep goal",
"Remove goal",
"The assistant stops working towards it. Your chat and files stay as they are.",
"This model can read pictures, but the engine installed here does not include the picture reader yet, so Read pictures is off. Update the engine in Settings → About to turn it on.",
"Showing the values of your tuning '{0}'.",
"Your tuning '{0}' sets",
"was {0}",
"· {0} RAM",
"The message was not cancelled",
"Extra settings for the program",
"Draft it with AI",
"Thinking time",
"Writing time",
"tokens/s",
"% of context",
"This page asks for files from the internet, which the preview does not load, so it may look incomplete. View code shows exactly what was written.",
"Keep A",
"Keep B",
"Took",
"Variation",
"Changed",
"Nothing",
"Both versions are ready. Compare A and B, then keep one.",
"The assistant is still answering",
"Send your message when it has finished.",
"The message was not queued",
"Scroll to the latest message",
"Remove task",
"Start a new conversation; keep this one",
"Have the assistant write a plan first; nothing runs until you approve it",
"Search the web, read the pages and answer with sources",
"Copied",
"Response",
"Message",
"PDF added",
"Select the text and copy it with Ctrl+C.",
"{0} is in this chat. Ask about it in your next message.",
"Heruntergeladen, aber nicht geladen. Öffne „Modelle“, um es zu laden.",
"{0} Zeilen hinzugefügt, {1} entfernt",
"{0} Dateien bearbeitet",
"Dateien hier ablegen, um sie hinzuzufügen",
"Bilder und Textdateien",
"Ziel behalten",
"Ziel entfernen",
"Der Assistent arbeitet nicht mehr darauf hin. Dein Chat und deine Dateien bleiben unverändert.",
"Dieses Modell kann Bilder lesen, aber die hier installierte Engine enthält den Bildleser noch nicht, deshalb ist „Bilder lesen“ aus. Aktualisiere die Engine unter Einstellungen → Info, um es einzuschalten.",
"Zeigt die Werte deiner Abstimmung „{0}“.",
"Deine Abstimmung „{0}“ setzt",
"war {0}",
"· {0} RAM",
"Die Nachricht wurde nicht abgebrochen",
"Zusätzliche Einstellungen für das Programm",
"Mit KI entwerfen",
"Denkzeit",
"Schreibzeit",
"Token/s",
"% des Kontexts",
"Diese Seite lädt Dateien aus dem Internet, was die Vorschau nicht tut. Sie kann deshalb unvollständig aussehen. „Code anzeigen“ zeigt genau, was geschrieben wurde.",
"A behalten",
"B behalten",
"Dauer",
"Schwankung",
"Geändert",
"Nichts",
"Beide Versionen sind fertig. Vergleiche A und B und behalte eine.",
"Der Assistent antwortet noch",
"Sende deine Nachricht, wenn er fertig ist.",
"Die Nachricht wurde nicht eingereiht",
"Zur neuesten Nachricht scrollen",
"Aufgabe entfernen",
"Neue Unterhaltung beginnen; diese bleibt erhalten",
"Der Assistent schreibt zuerst einen Plan; nichts läuft, bis du ihn freigibst",
"Im Web suchen, die Seiten lesen und mit Quellen antworten",
"Kopiert",
"Antwort",
"Nachricht",
"PDF hinzugefügt",
"Markiere den Text und kopiere ihn mit Strg+C.",
"{0} ist in diesem Chat. Frag in deiner nächsten Nachricht danach.",
],
"es":[
"dejará de ejecutarse y se eliminará.",
"Descargado, pero no cargado. Abre Modelos para cargarlo.",
"{0} líneas añadidas, {1} eliminadas",
"{0} archivos editados",
"Suelta los archivos para añadirlos",
"Imágenes y archivos de texto",
"Mantener el objetivo",
"Quitar el objetivo",
"El asistente deja de trabajar en ello. Tu chat y tus archivos se quedan como están.",
"Este modelo puede leer imágenes, pero el motor instalado aquí aún no incluye el lector de imágenes, así que Leer imágenes está desactivado. Actualiza el motor en Ajustes → Acerca de para activarlo.",
"Se muestran los valores de tu ajuste «{0}».",
"Tu ajuste «{0}» establece",
"antes {0}",
"· {0} de RAM",
"El mensaje no se canceló",
"Ajustes adicionales del programa",
"Redactar con IA",
"Tiempo de pensar",
"Tiempo de escribir",
"tokens/s",
"% del contexto",
"Esta página pide archivos de internet, que la vista previa no carga, así que puede verse incompleta. Ver código muestra exactamente lo que se escribió.",
"Mantener A",
"Mantener B",
"Tardó",
"Variación",
"Cambiado",
"Nada",
"Las dos versiones están listas. Compara A y B y quédate con una.",
"El asistente sigue respondiendo",
"Envía tu mensaje cuando termine.",
"El mensaje no se puso en cola",
"Ir al último mensaje",
"Quitar tarea",
"Empieza una conversación nueva; esta se conserva",
"El asistente escribe primero un plan; no se ejecuta nada hasta que lo apruebes",
"Busca en la web, lee las páginas y responde con fuentes",
"Copiado",
"Respuesta",
"Mensaje",
"PDF añadido",
"Selecciona el texto y cópialo con Ctrl+C.",
"{0} está en este chat. Pregunta por él en tu próximo mensaje.",
],
"ru":[
"перестанет работать и будет удалена.",
"Скачана, но не загружена. Откройте «Модели», чтобы загрузить её.",
"Добавлено строк: {0}, удалено: {1}",
"Изменено файлов: {0}",
"Перетащите файлы сюда, чтобы добавить",
"Картинки и текстовые файлы",
"Оставить цель",
"Удалить цель",
"Ассистент перестанет работать над ней. Ваш чат и файлы останутся как есть.",
"Эта модель умеет читать картинки, но установленный здесь движок пока не содержит модуль чтения картинок, поэтому «Читать картинки» выключено. Обновите движок в разделе «Настройки → О программе», чтобы включить.",
"Показаны значения вашей настройки «{0}».",
"Ваша настройка «{0}» задаёт",
"было {0}",
"· {0} ОЗУ",
"Сообщение не было отменено",
"Дополнительные настройки программы",
"Набросать с помощью ИИ",
"Время размышления",
"Время написания",
"токенов/с",
"% контекста",
"Эта страница запрашивает файлы из интернета, а предпросмотр их не загружает, поэтому она может выглядеть неполной. «Показать код» показывает, что именно было написано.",
"Оставить A",
"Оставить B",
"Заняло",
"Разброс",
"Изменено",
"Ничего",
"Обе версии готовы. Сравните A и B и оставьте одну.",
"Ассистент ещё отвечает",
"Отправьте сообщение, когда он закончит.",
"Сообщение не поставлено в очередь",
"К последнему сообщению",
"Удалить задачу",
"Начать новый разговор; этот сохраняется",
"Ассистент сначала напишет план; ничего не запустится, пока вы его не одобрите",
"Искать в интернете, читать страницы и отвечать со ссылками на источники",
"Скопировано",
"Ответ",
"Сообщение",
"PDF добавлен",
"Выделите текст и скопируйте его клавишами Ctrl+C.",
"{0} добавлен в этот чат. Спросите о нём в следующем сообщении.",
],
"pt":[
"vai parar de funcionar e será removido.",
"Baixado, mas não carregado. Abra Modelos para carregá-lo.",
"{0} linhas adicionadas, {1} removidas",
"{0} arquivos editados",
"Solte os arquivos para adicioná-los",
"Imagens e arquivos de texto",
"Manter a meta",
"Remover a meta",
"O assistente deixa de trabalhar nela. Seu chat e seus arquivos continuam como estão.",
"Este modelo consegue ler imagens, mas o motor instalado aqui ainda não inclui o leitor de imagens, então Ler imagens está desligado. Atualize o motor em Configurações → Sobre para ligar.",
"Mostrando os valores do seu ajuste \"{0}\".",
"Seu ajuste \"{0}\" define",
"era {0}",
"· {0} de RAM",
"A mensagem não foi cancelada",
"Configurações extras do programa",
"Rascunhar com IA",
"Tempo de raciocínio",
"Tempo de escrita",
"tokens/s",
"% do contexto",
"Esta página pede arquivos da internet, que a visualização não carrega, então pode parecer incompleta. Ver código mostra exatamente o que foi escrito.",
"Manter A",
"Manter B",
"Levou",
"Variação",
"Alterado",
"Nada",
"As duas versões estão prontas. Compare A e B e mantenha uma.",
"O assistente ainda está respondendo",
"Envie sua mensagem quando ele terminar.",
"A mensagem não entrou na fila",
"Ir para a última mensagem",
"Remover tarefa",
"Começar uma nova conversa; esta continua salva",
"O assistente escreve um plano primeiro; nada roda até você aprovar",
"Pesquisar na web, ler as páginas e responder com fontes",
"Copiado",
"Resposta",
"Mensagem",
"PDF adicionado",
"Selecione o texto e copie com Ctrl+C.",
"{0} está neste chat. Pergunte sobre ele na próxima mensagem.",
],
"ko":[
"실행이 중지되고 제거됩니다.",
"다운로드했지만 아직 불러오지 않았습니다. 모델 화면을 열어 불러오세요.",
"{0}줄 추가, {1}줄 삭제",
"파일 {0}개 편집함",
"파일을 놓아서 추가하세요",
"그림과 텍스트 파일",
"목표 유지",
"목표 제거",
"어시스턴트가 이 목표를 향한 작업을 멈춥니다. 대화와 파일은 그대로 남습니다.",
"이 모델은 그림을 읽을 수 있지만, 여기에 설치된 엔진에는 아직 그림 읽기 기능이 없어서 '그림 읽기'가 꺼져 있습니다. 설정 → 정보에서 엔진을 업데이트하면 켤 수 있습니다.",
"튜닝 '{0}'의 값을 보여 줍니다.",
"튜닝 '{0}'이(가) 설정하는 항목",
"이전 값 {0}",
"· RAM {0}",
"메시지가 취소되지 않았습니다",
"프로그램 추가 설정",
"AI로 초안 작성",
"생각한 시간",
"작성한 시간",
"토큰/초",
"컨텍스트 %",
"이 페이지는 인터넷의 파일을 불러와야 하는데 미리보기는 이를 불러오지 않으므로 불완전하게 보일 수 있습니다. '코드 보기'에서 작성된 내용을 그대로 볼 수 있습니다.",
"""Writes src/locales/gh-<lang>.json: the interface strings of the 0.2.7 GitHub items done in f-027-gh-app (the Monitor's Power card #3628,
approval notifications and the tray #3477, the finished-task sound #3586, the pinned port #3402, the interface size #3970).
Run from apps/host: python scripts/make-gh-locales.py
The keys are gh01.. in this order; gh-en.json is the English source. Placeholders ({0}, {1}) must be kept in every language.
A separate file keeps this work from colliding with the other catalogs (the same reason tune-, tools- and fix- exist)."""
importjson
importre
frompathlibimportPath
EN=[
"Energy since start",
"Cost since start",
"GPU fan",
"GPU clock",
"CPU temperature",
"SSD temperature",
"Electricity price",
"Price of one kilowatt-hour. The Monitor then shows what your graphics card has cost since Strata started. Leave it empty to hide the cost.",
"Price per kWh",
"Sound when work finishes",
"Plays a short sound when an answer or task that took a while is finished, or when Strata is in the background.",
"Play sound",
"Notify me when Strata needs an answer",
"In Manual mode, shows a desktop notification when the assistant waits for your approval and Strata is in the background.",
"Strata needs your answer",
"The assistant wants to use {0}. Open Strata to approve or deny.",
"a tool",
"Minimise to tray",
"Closing the window keeps Strata running in the tray, next to the clock. Choose Quit in the tray menu to stop it. Applies the next time Strata starts.",
"App port",
"The port Strata’s window and local address use. Leave it empty to let Strata choose (47831 when it is free). Applies the next time Strata starts.",
"Port (empty = automatic)",
"Using port {0} now.",
"Strata could not use its port",
"Port {0} is used by another program, so Strata started on port {1}. Chats and settings saved on port {0} are not shown until Strata can use it again. Close the other program and restart Strata.",
"Choose a port between 1024 and 65535, or leave it empty.",
"Makes all text, buttons and menus in Strata larger or smaller.",
"Preis einer Kilowattstunde. Der Monitor zeigt dann, was Ihre Grafikkarte seit dem Start von Strata gekostet hat. Leer lassen, um die Kosten auszublenden.",
"Preis pro kWh",
"Ton, wenn Arbeit fertig ist",
"Spielt einen kurzen Ton, wenn eine Antwort oder Aufgabe, die länger gedauert hat, fertig ist oder wenn Strata im Hintergrund läuft.",
"Ton abspielen",
"Benachrichtigen, wenn Strata eine Antwort braucht",
"Zeigt im manuellen Modus eine Desktop-Benachrichtigung, wenn der Assistent auf Ihre Freigabe wartet und Strata im Hintergrund läuft.",
"Strata braucht Ihre Antwort",
"Der Assistent möchte {0} verwenden. Öffnen Sie Strata, um zu erlauben oder abzulehnen.",
"ein Werkzeug",
"In den Infobereich minimieren",
"Beim Schließen des Fensters läuft Strata im Infobereich neben der Uhr weiter. Wählen Sie im Menü des Symbols „Beenden“, um es zu stoppen. Gilt ab dem nächsten Start von Strata.",
"App-Port",
"Der Port, den das Fenster und die lokale Adresse von Strata verwenden. Leer lassen, damit Strata ihn wählt (47831, wenn frei). Gilt ab dem nächsten Start von Strata.",
"Port (leer = automatisch)",
"Aktuell wird Port {0} verwendet.",
"Strata konnte seinen Port nicht verwenden",
"Port {0} wird von einem anderen Programm verwendet, daher wurde Strata auf Port {1} gestartet. Unter Port {0} gespeicherte Chats und Einstellungen werden nicht angezeigt, bis Strata ihn wieder verwenden kann. Schließen Sie das andere Programm und starten Sie Strata neu.",
"Wählen Sie einen Port zwischen 1024 und 65535 oder lassen Sie das Feld leer.",
"Macht alle Texte, Schaltflächen und Menüs in Strata größer oder kleiner.",
"Energie- und Temperaturwerte",
"pro kWh, in {0}",
"pro kWh",
"Währung ändern",
"Automatisch ({0})",
"Währung",
"weniger als {0}",
],
"es":[
"Energía desde el inicio",
"Coste desde el inicio",
"Ventilador de la GPU",
"Reloj de la GPU",
"Temperatura de la CPU",
"Temperatura del SSD",
"Precio de la electricidad",
"Precio de un kilovatio hora. El Monitor mostrará entonces lo que ha costado tu tarjeta gráfica desde que se inició Strata. Déjalo vacío para ocultar el coste.",
"Precio por kWh",
"Sonido al terminar el trabajo",
"Reproduce un sonido breve cuando termina una respuesta o tarea que tardó un rato, o cuando Strata está en segundo plano.",
"Reproducir sonido",
"Avisarme cuando Strata necesite una respuesta",
"En el modo manual, muestra una notificación de escritorio cuando el asistente espera tu aprobación y Strata está en segundo plano.",
"Strata necesita tu respuesta",
"El asistente quiere usar {0}. Abre Strata para aprobar o denegar.",
"una herramienta",
"Minimizar a la bandeja",
"Al cerrar la ventana, Strata sigue funcionando en la bandeja, junto al reloj. Elige Salir en el menú de la bandeja para detenerlo. Se aplica la próxima vez que se inicie Strata.",
"Puerto de la aplicación",
"El puerto que usan la ventana y la dirección local de Strata. Déjalo vacío para que Strata elija (47831 si está libre). Se aplica la próxima vez que se inicie Strata.",
"Puerto (vacío = automático)",
"Ahora se usa el puerto {0}.",
"Strata no pudo usar su puerto",
"El puerto {0} lo usa otro programa, así que Strata se inició en el puerto {1}. Los chats y ajustes guardados en el puerto {0} no se muestran hasta que Strata pueda volver a usarlo. Cierra el otro programa y reinicia Strata.",
"Elige un puerto entre 1024 y 65535, o déjalo vacío.",
"Hace más grandes o más pequeños todos los textos, botones y menús de Strata.",
"Lecturas de energía y temperatura",
"por kWh, en {0}",
"por kWh",
"Cambiar moneda",
"Automática ({0})",
"Moneda",
"menos de {0}",
],
"ru":[
"Энергия с запуска",
"Стоимость с запуска",
"Вентилятор GPU",
"Частота GPU",
"Температура CPU",
"Температура SSD",
"Цена электричества",
"Цена одного киловатт-часа. Тогда монитор покажет, во сколько видеокарта обошлась с запуска Strata. Оставьте поле пустым, чтобы скрыть стоимость.",
"Цена за кВт·ч",
"Звук по окончании работы",
"Проигрывает короткий звук, когда закончился долгий ответ или задача, а также когда Strata работает в фоне.",
"Проиграть звук",
"Уведомлять, когда Strata нужен ответ",
"В ручном режиме показывает уведомление на рабочем столе, когда помощник ждёт вашего подтверждения, а Strata работает в фоне.",
"Strata нужен ваш ответ",
"Помощник хочет использовать {0}. Откройте Strata, чтобы разрешить или отклонить.",
"инструмент",
"Сворачивать в трей",
"При закрытии окна Strata продолжает работать в трее рядом с часами. Чтобы остановить её, выберите «Выход» в меню значка. Действует со следующего запуска Strata.",
"Порт приложения",
"Порт, который используют окно и локальный адрес Strata. Оставьте поле пустым, чтобы Strata выбрала сама (47831, если свободен). Действует со следующего запуска Strata.",
"Порт (пусто = автоматически)",
"Сейчас используется порт {0}.",
"Strata не смогла использовать свой порт",
"Порт {0} занят другой программой, поэтому Strata запустилась на порту {1}. Чаты и настройки, сохранённые на порту {0}, не показываются, пока Strata не сможет снова его использовать. Закройте другую программу и перезапустите Strata.",
"Выберите порт от 1024 до 65535 или оставьте поле пустым.",
"Увеличивает или уменьшает весь текст, кнопки и меню в Strata.",
"Показания энергии и температуры",
"за кВт·ч, в валюте {0}",
"за кВт·ч",
"Сменить валюту",
"Автоматически ({0})",
"Валюта",
"меньше {0}",
],
"pt":[
"Energia desde o início",
"Custo desde o início",
"Ventoinha da GPU",
"Clock da GPU",
"Temperatura da CPU",
"Temperatura do SSD",
"Preço da eletricidade",
"Preço de um quilowatt-hora. O Monitor passa a mostrar quanto a sua placa de vídeo custou desde que o Strata iniciou. Deixe vazio para ocultar o custo.",
"Preço por kWh",
"Som ao terminar o trabalho",
"Toca um som curto quando uma resposta ou tarefa que demorou termina, ou quando o Strata está em segundo plano.",
"Tocar som",
"Avisar quando o Strata precisar de uma resposta",
"No modo manual, mostra uma notificação na área de trabalho quando o assistente espera sua aprovação e o Strata está em segundo plano.",
"O Strata precisa da sua resposta",
"O assistente quer usar {0}. Abra o Strata para aprovar ou negar.",
"uma ferramenta",
"Minimizar para a bandeja",
"Ao fechar a janela, o Strata continua em execução na bandeja, ao lado do relógio. Escolha Sair no menu da bandeja para encerrá-lo. Vale a partir da próxima vez que o Strata iniciar.",
"Porta do aplicativo",
"A porta usada pela janela e pelo endereço local do Strata. Deixe vazio para o Strata escolher (47831 quando estiver livre). Vale a partir da próxima vez que o Strata iniciar.",
"Porta (vazio = automático)",
"Usando a porta {0} agora.",
"O Strata não conseguiu usar a sua porta",
"A porta {0} está em uso por outro programa, então o Strata iniciou na porta {1}. Os chats e as configurações salvos na porta {0} não aparecem até o Strata poder usá-la de novo. Feche o outro programa e reinicie o Strata.",
"Escolha uma porta entre 1024 e 65535, ou deixe vazio.",
"Deixa todos os textos, botões e menus do Strata maiores ou menores.",
"Leituras de energia e temperatura",
"por kWh, em {0}",
"por kWh",
"Alterar moeda",
"Automática ({0})",
"Moeda",
"menos de {0}",
],
"ko":[
"시작 이후 에너지",
"시작 이후 비용",
"GPU 팬",
"GPU 클럭",
"CPU 온도",
"SSD 온도",
"전기 요금",
"1킬로와트시의 요금입니다. 모니터에 Strata를 시작한 뒤 그래픽 카드에 든 전기 요금이 표시됩니다. 비워 두면 비용을 표시하지 않습니다.",
"kWh당 요금",
"작업이 끝났을 때 소리",
"답변이나 오래 걸린 작업이 끝났을 때, 또는 Strata가 백그라운드에 있을 때 짧은 소리를 재생합니다.",
"소리 재생",
"Strata가 답을 필요로 할 때 알림",
"수동 모드에서 어시스턴트가 승인을 기다리고 Strata가 백그라운드에 있을 때 데스크톱 알림을 표시합니다.",
"Strata가 답을 기다리고 있습니다",
"어시스턴트가 {0}을(를) 사용하려고 합니다. Strata를 열어 승인하거나 거부하세요.",
"도구",
"트레이로 최소화",
"창을 닫아도 Strata는 시계 옆 트레이에서 계속 실행됩니다. 끝내려면 트레이 메뉴에서 '종료'를 선택하세요. 다음에 Strata를 시작할 때부터 적용됩니다.",
"앱 포트",
"Strata의 창과 로컬 주소가 사용하는 포트입니다. 비워 두면 Strata가 고릅니다(비어 있으면 47831). 다음에 Strata를 시작할 때부터 적용됩니다.",
"포트(비움 = 자동)",
"지금은 포트 {0}을(를) 사용 중입니다.",
"Strata가 지정한 포트를 사용할 수 없었습니다",
"포트 {0}을(를) 다른 프로그램이 사용 중이어서 Strata는 포트 {1}에서 시작했습니다. 포트 {0}에 저장된 채팅과 설정은 Strata가 다시 그 포트를 쓸 수 있을 때까지 표시되지 않습니다. 다른 프로그램을 닫고 Strata를 다시 시작하세요.",
"1024에서 65535 사이의 포트를 고르거나 비워 두세요.",
"Strata의 모든 글자, 버튼, 메뉴를 더 크게 또는 더 작게 만듭니다.",
"에너지 및 온도 측정값",
"kWh당, 통화 {0}",
"kWh당",
"통화 변경",
"자동({0})",
"통화",
"{0} 미만",
],
"fr":[
"Énergie depuis le démarrage",
"Coût depuis le démarrage",
"Ventilateur du GPU",
"Fréquence du GPU",
"Température du CPU",
"Température du SSD",
"Prix de l'électricité",
"Prix d'un kilowattheure. Le Moniteur affiche alors ce que votre carte graphique a coûté depuis le démarrage de Strata. Laissez vide pour masquer le coût.",
"Prix par kWh",
"Son à la fin du travail",
"Joue un court son quand une réponse ou une tâche qui a pris du temps est terminée, ou quand Strata est en arrière-plan.",
"Jouer le son",
"Me prévenir quand Strata a besoin d'une réponse",
"En mode manuel, affiche une notification sur le bureau quand l'assistant attend votre accord et que Strata est en arrière-plan.",
"Strata a besoin de votre réponse",
"L'assistant veut utiliser {0}. Ouvrez Strata pour accepter ou refuser.",
"un outil",
"Réduire dans la zone de notification",
"Fermer la fenêtre garde Strata actif dans la zone de notification, près de l'horloge. Choisissez Quitter dans le menu de l'icône pour l'arrêter. S'applique au prochain démarrage de Strata.",
"Port de l'application",
"Le port que la fenêtre et l'adresse locale de Strata utilisent. Laissez vide pour que Strata choisisse (47831 s'il est libre). S'applique au prochain démarrage de Strata.",
"Port (vide = automatique)",
"Port {0} utilisé en ce moment.",
"Strata n'a pas pu utiliser son port",
"Le port {0} est utilisé par un autre programme, donc Strata a démarré sur le port {1}. Les conversations et réglages enregistrés sur le port {0} ne s'affichent pas tant que Strata ne peut pas le réutiliser. Fermez l'autre programme et relancez Strata.",
"Choisissez un port entre 1024 et 65535, ou laissez vide.",
"Agrandit ou réduit tous les textes, boutons et menus de Strata.",
# Agent bridge (Strata chats as DeepSeek Harness sessions)
This page is for developers who work on the app host's agent code. It lists the files, the frames and routes, and the rules the
bridge enforces. Tests are listed at the end.
The Harness is a separate, pinned component (`DSH 0.2.1-alpha.3`, commit `57190202`, MIT). The host finds it through
`STRATA_HARNESS_DIR` (or `<STRATA_COMPONENTS_DIR>/harness`), starts one `dsh --profile web` process on a loopback port,
and talks to its Gateway. Nothing here downloads or builds anything.
@@ -14,12 +17,13 @@ and talks to its Gateway. Nothing here downloads or builds anything.
| `children.mjs` | pure projection of one helper agent's events into its row and its approval card |
| `preset.mjs` | the agent preset (which tools the chat has) |
| `plugins/strata-gate.mjs` | the Manual-mode approval gate (a Harness plugin, loaded by path) |
| `plugins/strata-search.mjs`, `plugins/strata-models.mjs` | the app's web search registered with the Harness; helper agents on a named model |
| `plugins/strata-context.mjs`, `context-budget.mjs`, `context-fit.mjs` | measured context pressure, summaries that fit the window, the check before a message goes out |
| `agent-bridge.mjs` | the chat-to-session map, turns, approvals, Stop, status |
## Frames sent on `/api/chat/stream` for an agent turn
`queued`, `start` (`agent:true`), `reasoning`, `delta`, `agent-tool` (new: upserts a tool card by call id with
`queued`, `start` (`agent:true`), `reasoning`, `delta`, `agent-tool` (upserts a tool card by call id with
(why a chat is not using the agent). The chat UI renders them with its existing components.
@@ -57,7 +61,7 @@ The chat's Mode menu is real Harness behaviour. Each turn the page sends the cha
| Bypass | `strata-bypass` | `danger-full-access` | `never` | Full access without prompts, as before. |
- Manual relies on the gate, so a Manual turn does not start unless the bridge has confirmed the gate is loaded (`/strata-gate` command): without it Manual would be an open sandbox that nobody asks about.
- The gate listens to `tools/pre-execute` (the seam the Harness's own Auto-review layer uses) and returns an `ask` decision; the Harness's approval service then asks, runs the tool only after `allowed-once`, and records the answer. Read-only tools and a short allowlist of look-only shell commands (Get-ChildItem, git status, ...) run unasked; MCP and unknown tools ask.
- The gate listens to `tools/pre-execute` (the seam the Harness's own Auto-review layer uses) and returns an `ask` decision; the Harness's approval service then asks, runs the tool only after `allowed-once`, and records the answer. Read-only tools and a short allowlist of look-only shell commands (Get-ChildItem, git status, ...) run unasked, each only with the options listed for it in `strata-gate.mjs` (an unknown option, or one that writes, runs a program, sets something or never ends, asks); MCP and unknown tools ask.
- Web search and web pages (`web_search`, `web_fetch`) ask once per chat in Manual, not once per call. The gate asks through the approval service itself (`webDecision`: "The assistant wants to search the web for: ... Search words go to <service>." / "... open the web page ..."), remembers a yes for the chat (a helper agent shares its chat's answer, and two searches started together share one card), and on a no gives the model `WEB_REFUSAL` ("The person did not allow web access ... Go on without it and say so."). Plan and Bypass are unchanged. Settings -> Tools has a switch, "Let the assistant search the web without asking" (off by default, `webWithoutAsking` in `tools-settings.mjs`, passed to the gate as `config.web.skipAsk`), that skips the question in Manual. A scheduled task has nobody at the card, so its web call is answered no unless that switch is on. Tests: `tests/web-ask.test.mjs`.
- The mode is read once per turn; an unknown or missing value is Manual.
@@ -70,7 +74,7 @@ The chat's Mode menu is real Harness behaviour. Each turn the page sends the cha
What the person chose is saved by `tools-settings.mjs` (inside the host's settings file) and shown by `src/components/SettingsTools.tsx` (`GET/PUT /api/agent/tools`, `POST/PATCH/DELETE /api/agent/tools/custom[/id]`). A change reaches the agent from the next message: the profile carries a `toolsSignature`, and the bridge restarts the Harness (after the running turns end) when it differs.
- Tool groups (`tools-catalog.mjs`): each is one switch in plain words and is one or more preset plugins (`withoutDisabled`); `fetch` and `search` are the two halves of the Harness's `tool-web` (both off drops it). Goals off: `/goal` answers that goals are turned off.
- Web search needs no account. The default is DuckDuckGo's HTML result page (`../web-search-ddg.mjs`), with Bing's plain result page (`../web-search-bing.mjs`) as the second source. Manners, because DuckDuckGo refuses a program that asks fast (a real box: two searches in 1.2 s answered, the third got HTTP 202 and every DuckDuckGo page refused for many minutes): a browser User-Agent and the page's own request shape (Referer, Origin, Sec-Fetch-*), one request at a time per source, 2.5 to 4 s apart (random), at most 8 a minute, a 30 minute session cache per question, a cooldown after a refusal that doubles (1, 2, 4 ... 10 minutes, `Retry-After` honoured) during which that source gets no request, no lite-page follow-up after a refusal, one retry for a timeout or a 5xx, a 15 s limit. When every source refuses, the tool answers one plain sentence (`LIMITED_SENTENCE`: "Web search is limited right now. Try again in a minute, or add a search key (Exa, Perplexity or DeepSeek) in Settings -> Tools -> Web search.") plus an instruction to the model not to retry; the turn guard in `agent-bridge.mjs` (see Turn guard) stops a model that retries anyway. **The results come from DuckDuckGo or Bing**; they are not ranked or filtered here. The person may switch to Exa, Perplexity or DeepSeek with their own key (`../web-search-keyed.mjs`; DeepSeek uses the Harness's own provider): the key is saved in the settings file, reaches the agent only as the environment variable `STRATA_SEARCH_KEY`, never the profile file and never the screen (it only learns `keySet`). A keyed service without its key falls back to DuckDuckGo. The profile pins `web.searchProvider` to the chosen id so two usable backends cannot make the choice ambiguous.
- Web search needs no account. The default is DuckDuckGo's HTML result page (`../web-search-ddg.mjs`), with Bing's plain result page (`../web-search-bing.mjs`) as the second source. Manners, because DuckDuckGo refuses a program that asks fast: a browser User-Agent and the page's own request shape (Referer, Origin, Sec-Fetch-*), one request at a time per source, 2.5 to 4 s apart (random), at most 8 a minute, a 30 minute session cache per question, a cooldown after a refusal that doubles (1, 2, 4 ... 10 minutes, `Retry-After` honoured) during which that source gets no request, no lite-page follow-up after a refusal, one retry for a timeout or a 5xx, a 15 s limit. When every source refuses, the tool answers one plain sentence (`LIMITED_SENTENCE`: "Web search is limited right now. Try again in a minute, or add a search key (Exa, Perplexity or DeepSeek) in Settings -> Tools -> Web search.") plus an instruction to the model not to retry; the turn guard in `agent-bridge.mjs` (see Turn guard) stops a model that retries anyway. **The results come from DuckDuckGo or Bing**; they are not ranked or filtered here. The person may switch to Exa, Perplexity or DeepSeek with their own key (`../web-search-keyed.mjs`; DeepSeek uses the Harness's own provider): the key is saved in the settings file, reaches the agent only as the environment variable `STRATA_SEARCH_KEY`, never the profile file and never the screen (it only learns `keySet`). A keyed service without its key falls back to DuckDuckGo. The profile pins `web.searchProvider` to the chosen id so two usable backends cannot make the choice ambiguous.
- The browser: Playwright's MCP server (`@playwright/mcp`, staged next to the Harness by `apps/desktop/scripts/stage-harness.mjs`) on the person's own Chrome, Edge or Chromium (`browser-probe.mjs`), `--isolated` so it never sees their profile. Tools appear as `mcp__browser__browser_*`. In Manual, opening a page, clicking, typing, uploading, running a script in the page and closing tabs ask (plain reasons: "Open the web page ...", "Click on the page: ..."); looking (snapshot, screenshot, console, network, waiting, listing tabs) does not. Plan refuses every connected tool. The screen says plainly when no browser, or no tool server, is found.
- The person's own tools: `instructions` (the tool returns their text; look-only, never asks), `command` (runs their command with the assistant's text in the `STRATA_INPUT` environment variable, never spliced into the command line; asks in Manual, refused in Plan) and `server` (a web address or a program, added as `mcp__custom-<name>__*`; asks in Manual). Instruction and command tools are served by `custom-tools-mcp.mjs` (stdio MCP) from a JSON file the settings module rewrites on every change, so a switch in Settings is seen without a restart.
-`strata-gate.mjs` implements the Manual/Plan rules above (`planDecision`, `browserDecision`, the look-only instruction server).
@@ -81,7 +85,7 @@ What the person chose is saved by `tools-settings.mjs` (inside the host's settin
- A helper's request for approval is the same card (state `proposed`, id `<helper session>:<call>`), answered with the same `/api/chat/approval`; a chat can only answer its own helpers.
- The run (the chat's one NDJSON response) stays open until the parent is idle, every helper has finished and reported back, and no armed goal is waiting; `done` is sent then.
- Stop: pauses an active goal, interrupts every helper (`subagents/interruptByParent`; `session/cancel` on a helper session is refused by the Harness), cancels the parent, waits; a report that would wake the parent afterwards is cancelled too.
- The number of helpers at once equals the engine's request slots (`maxActiveSubagents` in the profile); the Harness refuses more with a message the model reads.
- The number of helpers at once is the engine's request slots, with a floor of 4 for a local model and 8 for an API model (`maxActiveSubagents` in the profile). The Harness refuses more with a message the model reads.
## /goal
@@ -89,8 +93,15 @@ What the person chose is saved by `tools-settings.mjs` (inside the host's settin
## Thinking and the output budget
The Harness staging script (`apps/desktop/scripts/stage-harness.mjs`) applies Strata's guarded pi-ai 1.1.0 patch to its staged copy
(`harness-budget.mjs`, #3684): below 64K the context safety room is at most `min(4096, floor(contextWindow / 16))` and never more
than half the estimated remaining room, so an 8K window keeps output room after a tool result instead of ending with a one-token
cap. The caller's output cap and the one-token floor remain; windows of 64K or more keep the previous calculation. `BUNDLE.json.patches`
records the patch identity, the dependency version and the before/after SHA-256. Unknown dependency bytes stop staging for review.
Source and already-installed Harness copies are not changed.
- A local route is reasoning-capable (`reasoning: true`) with `compat.thinkingFormat: 'chat-template'`: the request carries `chat_template_kwargs {enable_thinking, reasoning_effort}`, which the Strata server reads (a top-level `enable_thinking` is ignored by it, which is why pi-ai's `qwen` format left the model thinking). The chat's Reasoning menu travels as `thinking` in the stream body; the bridge sends it with `session/selectModel` (`reasoningEffort`) whenever the model or the level changed; a level a route does not have is dropped (never an error). API routes keep the provider default (the menu is disabled for them in the UI).
-`localOutputBudget(context)`= context minus 4096 (it was context/4 capped at 8192, which cut big answers off at exactly 8192 tokens); pi-ai sends the window minus the prompt. App-loaded models get `fit_max_tokens: true` so the engine clamps instead of answering 400. Sampling is not set here: the Harness sends none, so the server's own defaults apply.
-`localOutputBudget(context)`is 30 percent of the window, at least 4096 and at most 32768 (9,830 tokens at 32K, 19,660 at 64K). The Harness's automatic compaction reads it as the room a reply reserves. pi-ai sends the window minus the prompt. App-loaded models get `fit_max_tokens: true` so the engine clamps instead of answering 400. Sampling is not set here: the Harness sends none, so the server's own defaults apply.
## Tasks
@@ -108,7 +119,7 @@ Failures the person sees are plain sentences with a next step (`AgentUnavailable
## Helper agents on a named model (`plugins/strata-models.mjs`, `model-routes.mjs`)
The Harness's `subagent` tool takes an exact `provider` route key and model id, and the app had its model choice switched off, so every helper ran on the chat's model. The profile now enables the Harness's `subagent-model-selection-settings` with every route the app can reach (local instances and saved API models), and the `strata-models` plugin wraps the two tool definitions the Harness registers for each chat (`subagent`, `list_subagent_models`) in place:
The Harness's `subagent` tool takes an exact `provider` route key and model id. The profile enables the Harness's `subagent-model-selection-settings` with every route the app can reach (local instances and saved API models), and the `strata-models` plugin wraps the two tool definitions the Harness registers for each chat (`subagent`, `list_subagent_models`) in place:
- the model is found from its id, `provider/id`, the name in the model picker, `connection name`, or "local" (any letters). A close guess is not started: the tool answers "Did you mean ...?" and nothing runs until the call is made again with that model. Several fits are listed; nothing fitting lists the models that can be used and says they can be added in Models.
- the tool's own text says when to use it ("ask a subagent", "ask model X"), not to look for the model with shell commands, and not to wait with sleep. The system prompt (`preset.mjs`) says the same, so it reaches every model, local or hosted.
@@ -126,7 +137,7 @@ Pictures are on by default for every model that can read them: an API model read
## Turn guard (`turn-guard.mjs`)
A turn that keeps failing the same way is stopped by the app, never left to the model. Seen on a search-heavy `/plan`: every web search was refused, the model searched again and rewrote its plan after each one until the context reached 100%, then the turn ended with no answer and no message.
A turn that keeps failing the same way is stopped by the app, never left to the model.
- **Counted, from what the turn shows:** web searches that failed in a row (3), searches in one turn (10 in a plan turn, 25 otherwise), a plan turn that starts its numbered list from `1.` again (4 times), and a context of 93% or more (the Harness's own compaction did not keep up).
- **Stopped the way Stop stops** (`session/cancel`, helpers first) but ended as a normal turn (`done`, not cancelled, no error card), so the plan the model had written is still proposed. The plan card carries the reason (`plan.caution`); any other turn ends with one `delta` frame that has `own: true` and `key` (the English sentence, translated by the page; `GUARD_NOTES` are in the interface catalog in all eight languages). Each sentence says what happened and gives one next step.
@@ -134,9 +145,7 @@ A turn that keeps failing the same way is stopped by the app, never left to the
- **Never silent:** a turn that ends (`completed` or `max-tokens`) with no answer text gets "I ran out of room in this chat..." when the last request used 85% of the window or more, otherwise "I stopped without writing an answer..." (not when tool cards were shown, not after Stop, not after an error, which says why by itself). Goal runs have their own ending lines and are not guarded.
- Tests: `tests/agent-turn-guard.test.mjs` (counting, always; end to end with the real Harness, the scripted model's `searchloop` and `bloat` scripts and `searchMock: 'limited'` on the bridge, which runs the real search code on a pretend web that refuses, no network).
## Context: measured pressure, summaries that fit, the check before a message goes out (2026-10-22)
Found by the owner's hands-on test of the installed 0.2.1 (Gyro-S, 64K): a 118 KB log paste was 66,500 tokens to the engine and about 29,500 to the Harness (4 characters a token; logs and code are about 2); the engine refused it with a 400, the refusal stayed in the history, and then neither the next message nor `/compact` could go through, because the summary request carried the same history and was refused for its size too. `/compact` itself was never a command of the app: it went to the model as a question and the model answered "nothing to compact".
## Context: measured pressure, summaries that fit, the check before a message goes out
- **`plugins/strata-context.mjs`** (inserted by the profile, like the gate) wraps `tokenMeter.measure` and `llm.stream` in place, once each. Pressure = the last prompt size the model server reported for an answer, plus what changed since priced at the density that measurement showed (`context-budget.mjs`: `pressureNow`, `densityOf`); never lower than the Harness's own price. Node prices are scaled the same way, so the retained tail is in real tokens. A summary request (`purpose: 'compaction'`) that would not fit the window is asked for in pieces: a message bigger than a piece is cut for the request only, each piece builds on the summary before it, a piece the server refuses is halved and tried again. `/strata-context` (a Harness command, like `/strata-gate`) answers `{tokens, window}` for a session so the host can tell the person true figures.
- **`context-fit.mjs`**: before a message is sent the model server counts it (`POST /v1/messages/count_tokens`, its own tokenizer). `decideFit`: send; summarise first and then send (the message fits an otherwise empty chat); or never fits. The last one is not sent: the error frame has `code: 'CONTEXT_MESSAGE_TOO_LARGE'` and the page offers "Attach as a file" (`oversize.mjs`, `/api/chat/oversize`: the text is saved in the chat's folder, the assistant reads it in parts) or "Send a shortened preview". An engine that cannot count lets the message through; the Harness's own compact-and-retry is the net.
@@ -144,3 +153,15 @@ Found by the owner's hands-on test of the installed 0.2.1 (Gyro-S, 64K): a 118 K
- **Visible:** `compaction/start` shows "Summarizing earlier messages..." in the status line; `compaction/end` adds one plain line to the reply. The projector also sends a `context` frame after every model request, so the meter follows the running turn.
- **A refusal that is left** (`CONTEXT_WINDOW_EXCEEDED`) is a plain sentence with the engine's text under Details and the buttons "Summarize and continue" and "Start a new chat". "Try sending the message again" is for other errors only.
- Tests: `tests/agent-context.test.mjs` (arithmetic; the real Harness against the scripted worker with `window`, `charsPerToken` and `noCount`), `tests/oversize.test.mjs`, `tests/context-ui.test.mjs`. The run against a real engine is `scripts/context-e2e.mjs`.
## History / older measurements
- **Search refusals (2026-10):** on a real box two DuckDuckGo searches in 1.2 s answered, the third got HTTP 202, and every DuckDuckGo page refused for many minutes.
The pacing, cooldown and cache rules above come from that test.
- **Search-heavy `/plan` (2026-10):** every web search was refused, the model searched again and rewrote its plan after each one until the context reached 100%.
The turn ended with no answer and no message. The turn guard comes from that run.
- **Large log paste (2026-10-22, installed 0.2.1, Gyro-S, 64K):** a 118 KB log was 66,500 tokens to the engine and about 29,500 to the Harness
(4 characters a token; logs and code are about 2). The engine refused it with a 400. The refusal stayed in the history, so neither the next message nor `/compact` could
go through, because the summary request carried the same history. `/compact` was also not a command of the app yet: it went to the model as a question. This
produced `strata-context.mjs`, `context-fit.mjs` and the host's own `/compact`.
- **Output budget:** an earlier local budget of a quarter of the window, capped at 8192, cut big answers off at exactly 8192 tokens.
/** What `ls`, `dir`, `gci` and `Get-ChildItem` take. */
constLS={f:'aAbBcdDFgGhHiklLmnNopqQrRsStuUvxX1CZ',v:'ITw',ps:'path literalpath filter include exclude recurse depth name force file directory hidden readonly system attributes',
l:'all almost-all human-readable recursive directory classify reverse size inode dereference color group-directories-first sort time time-style block-size indicator-style si numeric-uid-gid no-group literal hide ignore quote-name escape author file-type full-time width tabsize'};
/** What `cat`, `type`, `gc` and `Get-Content` take. (-Wait never ends, so it is not here.) */
constCAT={f:'AbeEnstTuv',ps:'path literalpath totalcount tail head raw encoding readcount force filter include exclude delimiter stream asbytestream',
l:'show-all number-nonblank show-ends number squeeze-blank show-tabs show-nonprinting'};
constMEASURE={ps:'property sum average maximum minimum line word character ignorewhitespace inputobject standarddeviation allstats'};
constSELECT={ps:'first last skip skiplast property expandproperty unique index excludeproperty inputobject'};
constWHERE={ps:'property value inputobject eq ne gt lt ge le like notlike match notmatch contains notcontains in notin is isnot ceq cne cgt clt cge cle clike cnotlike cmatch cnotmatch ccontains cnotcontains cin cnotin'};
// POSIX. `sort` and `date` take no PowerShell names here: `sort -o FILE` writes, `date -s` sets the clock, and a PowerShell-looking `-Second` or `-Top`
// must not be read by a real sort or date as one of those. (Sort-Object and Get-Date above are the PowerShell forms.)
add('stat',{f:'LtxlnqrsFf',v:'c',l:'dereference file-system terse format printf'});
add('file',{f:'bhikLnNrvI',l:'brief dereference no-dereference mime mime-type mime-encoding keep-going no-pad no-buffer raw'});// no -C, -m, -f, -P (they write or read other files), no -z/-s/-p (they run decompressors, read devices, reset times)
add('du',{f:'abchHklmsxLPSD0',v:'tB',vn:'d',l:'all apparent-size bytes total dereference-args human-readable si summarize one-file-system dereference no-dereference separate-dirs count-links inodes null max-depth threshold block-size time time-style exclude'});
add('df',{f:'ahHiklmPTv',v:'txB',l:'all human-readable si inodes local portability print-type total sync no-sync output block-size type exclude-type'});
add('tree',{f:'adfilpsugDFqNQrtvxCnASUchJX',v:'LPIHT',vn:'',l:'help version filelimit charset sort timefmt noreport dirsfirst filesfirst prune du si gitignore matchdirs ignore-case inodes device nolinks'});// no -o, no -R
// `less` and `more` are not here: they wait for keys, and `less` can run commands.
/** The names that run without a card (for the tests). */
// `git config KEY` reads and `git config KEY VALUE` writes, so a key is only taken with --get; no --edit, --add, --unset, --file.
config:spec({f:'lz',l:'list get get-all get-regexp show-origin show-scope local global system worktree bool int bool-or-int path expiry-date type null name-only includes no-includes',lv:'default',
/** The PCI vendor ids of the display adapters Windows reports as present right now (PowerShell, Win32_VideoController); empty when that cannot be read. */
// Files inside a signed macOS app must stay intact after installation.
if(process.platform==='darwin'&&resolve(bundleDir).split(sep).some(part=>part.toLowerCase().endsWith('.app')))returnkeepBundle(log,'macOS app bundles must remain unchanged after signing');
log(`engine bundle: the copy that came with the installer (${packages.map(p=>p.name).join(', ')}) was removed; the same engine is complete in the engine store`);
return{removed:true,packages:packages.length};
}catch(error){
returnkeepBundle(log,`could not be checked (${error?.message??error})`);
}
}
functionkeepBundle(log,why){
log(`engine bundle: kept in the install folder (${why})`);
return{removed:false,reason:why};
}
/**
* What the host does once its engine step has settled: for an engine that is (or was) there, look at the install folder's leftover bundle.
* The lines go to the host's output (console.log), which the desktop shell copies to its log file (`[host] engine bundle: ...`).
// DeepSeek-V4.1-Flash's own context limit, read from its pack (the app offers sizes up to it: src/lib/deepseek-context.ts).
// Project Maya v1.1.28 (#442, @needmorevram, MIT) did the same in its dashboard. strata_control/dsv41_context.py reads the same file for setup and the Terminal.
import{readFileSync}from'node:fs';
import{join}from'node:path';
/** The pack's own limit: `max_position_embeddings` in <pack>/ds4.txt. Null when there is no pack, no file or no such line (the choices then stay as before). */
/* The app's version has one source: the desktop shell passes it (STRATA_APP_VERSION, from apps/desktop/src-tauri/Cargo.toml). Outside the app it is read from the same file in a checkout, else 'dev'. */
/* The engine that ships inside the installer is put in the components folder when this PC has none (or an older one); no network, hashes checked. The model list waits for it. */
/* An update from a per-GPU installer leaves its engine-bundle behind; it goes only when the same engine is complete in the store (sizes and hashes). */
/* A build made for another make of graphics card is not unpacked; Control's answer then carries the plain sentence (with the right download) instead of its own. */
entry=serving.open({client:'Strata chat',endpoint:connection.protocol==='anthropic'?'/messages':connection.protocol==='openai-responses'?'/responses':'/chat/completions',model:`${connection.name} · ${model.name}`,cancel:()=>controller.abort(newDOMException('Cancelled by the user','AbortError'))});
serving.mark(entry.id,{state:'generating'});
constresult=awaitmodelProviders.stream({selection,message,history,images,signal:requestSignal,systemPrompt:['You are Strata, a helpful assistant. Answer clearly. Do not claim to have read files or performed actions unless their results are included in the conversation.',skillInstructions(appliedSkills)].filter(Boolean).join('\n\n'),onFrame:frame=>{
constresult=awaitmodelProviders.stream({selection,message,history,images,thinking,signal:requestSignal,systemPrompt:['You are Strata, a helpful assistant. Answer clearly. Do not claim to have read files or performed actions unless their results are included in the conversation.',skillInstructions(appliedSkills)].filter(Boolean).join('\n\n'),onFrame:frame=>{
try{letat=componentsDirectory();while(at&&!existsSync(at)&&dirname(at)!==at)at=dirname(at);constf=statfsSync(at);info.freeBytes=Number(f.bavail)*Number(f.bsize);}catch{/* unknown: no warning */}
returnjson(res,200,info);
}
if(path==='/api/engine/job'&&method==='GET'){
// The newest engine download / file install, read without waiting for the native snapshot to connect (the first-run card needs it before that).
if(!/\.(tar\.gz|tgz|tar\.xz|zip)$/i.test(name))returnerror(res,400,'This is not a Strata engine package. Choose the .tar.gz or .zip file from the release page.');
constsafeId=value=>typeofvalue==='string'&&/^[a-z0-9](?:[a-z0-9_-]|\.(?!\.)){0,79}$/i.test(value);/* a dot inside the name: deepseek-v4.1-flash-ist-das-2.00bpw */
letsnapshot={available:false,repo:preferences.repo,summary:'Connect a local Strata installation.',catalog:null,installed:null,hardware:null,error:null};
constsavedJobs=readJson(join(directory,'native-jobs.json'),[]);for(constjof(Array.isArray(savedJobs)?savedJobs:[]).slice(-12))jobs.set(j.id,{...j,...(j.state==='running'?{state:'error',error:'The host restarted before completion. Check diagnostics, then retry; completed files are kept.',summary:'Interrupted operation'}:{})});
constsavedJobs=readJson(join(directory,'native-jobs.json'),[]);/* An app update ends the app on purpose: the installer replaces it and starts it again. The job that asked for it was saved as running together with the app version it started from (appUpdateFrom); a start that finds a different version knows the update finished, instead of calling it an interrupted operation. */
/* A record an OLDER host wrote (no appUpdateFrom): the app's own program file was replaced after the job began, so the installer finished. <install>/host/server is ROOT's parent chain: ROOT = <install>/host. */
constappProgramNewerThan=started=>{try{constt=Date.parse(started);if(!Number.isFinite(t))returnfalse;constdir=dirname(ROOT),exe=join(dir,process.platform==='win32'?'strata-desktop.exe':'strata-desktop'),times=[];for(constfof[exe,join(dir,'uninstall.exe')])if(existsSync(f)){constst=statSync(f);times.push(st.mtimeMs,st.ctimeMs);}/* the installer keeps the build time as the program's modified time but writes uninstall.exe and changes the program file's metadata when it installs */returntimes.some(x=>x>t);}catch{returnfalse;}};
for(constjof(Array.isArray(savedJobs)?savedJobs:[]).slice(-12))jobs.set(j.id,{...j,...(settledByAppUpdate(j)?{state:'done',error:null,summary:`Strata is updated to ${thisAppVersion()}.`,finishedAt:newDate().toISOString()}:neverFinishedAppUpdate(j)?{state:'error',error:`The update to ${j.appUpdateTo??'the new version'} did not finish: the app closed for the installer, but the old version ${j.appUpdateFrom} started again. Try Update again, or download the installer from the Strata release page (https://github.com/Niko1221/Strata/releases/latest).`,summary:'The app update did not finish',finishedAt:newDate().toISOString()}:j.state==='running'?{state:'error',error:'The host restarted before completion. Check diagnostics, then retry; completed files are kept.',summary:'Interrupted operation'}:{})});
/* The component updates and the rollback go through Control, not through the native snapshot, so they work in the first seconds after the app starts and on a PC that has no engine yet. */
constupdateLog=line=>{try{console.log(`update · ${line}`);}catch{/* no console */}};
constNO_INSTALL='Installing, updating, tuning with calibration or rolling back model files and engines happens from the Terminal (strata install), never while the app runs.';
if(JSON.stringify(next)!==JSON.stringify(config)){try{writeModelConfig(path,next);}catch{/* read-only folder: the load reports it */}}}}
functioninstalledInfo(info){if(!info)returninfo;return{...info,models:(info.models??[]).map(entry=>{constcfg=/^strata-[a-z0-9_-]+\.json$/i.test(entry.config??'')?readJson(entry.configPath??join(repo(),entry.config)):{};constat=cfg.args?.indexOf('--native')??-1,nativeFile=at>=0?cfg.args[at+1]:null;constdataFolder=typeofnativeFile==='string'&&basename(dirname(dirname(nativeFile)))==='models'?dirname(dirname(dirname(nativeFile))):info.data_folder;/* A model the person pointed at their own GGUF folder (not <data>/models/<tag>): remembered, so a later load runs setup with that folder again instead of downloading the files (#3199). */constownGgufDir=typeofnativeFile==='string'&&basename(dirname(dirname(nativeFile)))!=='models'&&existsSync(nativeFile)?dirname(nativeFile):undefined;return{...entry,ownGgufDir,diskGb:entry.diskGb??entry.disk_gb,diskBytes:entry.diskBytes??entry.disk_bytes,sharedBytes:entry.sharedBytes??entry.shared_bytes,removableGb:entry.removableGb??entry.removable_gb,dataFolder,visionPrepared:!!cfg.vision,/* The picture files are prepared when the run config has a vision section: the Models screen then loads straight away instead of running setup again (it used to read 'off' for every model, which sent every load of a picture-capable model through "Apply settings and load", and setup writes the arguments, the draft layer and the reader back each time). */images:cfg.vision?'yes':'off',draftLayer:cfg.mtp!==false,drafting:cfg.drafting??'auto',/* Picture reading is ON unless the person turned it off: no saved choice and not loaded is undefined, not false (owner rule: pictures on by default). */visionEnabled:cfg.ui_vision_enabled??(()=>{construnning=[...instances.values()].find(i=>i.model===entry.model&&i.status==='ready');returnrunning?!!running.vision:undefined;})(),visionDevice:cfg.ui_vision_device,resourceProfile:cfg.ui_resource_profile,lowRamMode:cfg.args?.includes('--resident-experts')?'resident':cfg.args?.includes('--mmap-experts')?'mmap':'auto'};})};}
functioninstalledInfo(info){if(!info)returninfo;return{...info,models:(info.models??[]).map(entry=>{constcfg=/^strata-[a-z0-9_.-]+\.json$/i.test(entry.config??'')?readJson(entry.configPath??join(repo(),entry.config)):{};constat=cfg.args?.indexOf('--native')??-1,nativeFile=at>=0?cfg.args[at+1]:null;constdataFolder=typeofnativeFile==='string'&&basename(dirname(dirname(nativeFile)))==='models'?dirname(dirname(dirname(nativeFile))):info.data_folder;/* A model the person pointed at their own GGUF folder (not <data>/models/<tag>): remembered, so a later load runs setup with that folder again instead of downloading the files (#3199). */constownGgufDir=typeofnativeFile==='string'&&basename(dirname(dirname(nativeFile)))!=='models'&&existsSync(nativeFile)?dirname(nativeFile):undefined;return{...entry,ownGgufDir,contextMax:configContextLimit(cfg)??undefined,diskGb:entry.diskGb??entry.disk_gb,diskBytes:entry.diskBytes??entry.disk_bytes,sharedBytes:entry.sharedBytes??entry.shared_bytes,removableGb:entry.removableGb??entry.removable_gb,dataFolder,visionPrepared:!!cfg.vision,/* The picture files are prepared when the run config has a vision section: the Models screen then loads straight away instead of running setup again (it used to read 'off' for every model, which sent every load of a picture-capable model through "Apply settings and load", and setup writes the arguments, the draft layer and the reader back each time). */images:cfg.vision?'yes':'off',draftLayer:cfg.mtp!==false,drafting:cfg.drafting??'auto',/* Picture reading is ON unless the person turned it off: no saved choice and not loaded is undefined, not false (owner rule: pictures on by default). */visionEnabled:cfg.ui_vision_enabled??(()=>{construnning=[...instances.values()].find(i=>i.model===entry.model&&i.status==='ready');returnrunning?!!running.vision:undefined;})(),visionDevice:cfg.ui_vision_device,resourceProfile:cfg.ui_resource_profile,lowRamMode:cfg.args?.includes('--resident-experts')?'resident':cfg.args?.includes('--mmap-experts')?'mmap':'auto'};})};}
/* The settings file Strata Control and the installer read (control-runtime.mjs sets STRATA_SETTINGS_FILE to it): the app's data folder is written there. */
if(refreshing)return!force&&snapshot.available?state():force?refreshing.then(()=>refresh(true)):refreshing;/* a forced read that meets one already running (started before an engine update finished) waits for it and reads again, so it never settles on the old answer */if(!force&&Date.now()-lastRead<45000)returnstate();
functionconfigure({repository}){if(packaged){if(instances.size||[...jobs.values()].some(j=>j.state==='running'))thrownewError('Unload models and finish operations before changing the installation.');constnext=realpathSync(String(repository));if(!statSync(next).isDirectory())thrownewError('Choose the folder that holds your Strata model configurations.');preferences={...preferences,installDir:next};packagedInstall=null;disconnect();writeFileSync(settingsPath,JSON.stringify(preferences,null,2),{mode:0o600});lastRead=0;returnrefresh(true);}if(instances.size||[...jobs.values()].some(j=>j.state==='running'))thrownewError('Unload models and finish operations before changing the installation.');constnext=realpathSync(String(repository));constold=preferences;preferences={...preferences,repo:next};try{repo();}catch(e){preferences=old;throwe;}disconnect();writeFileSync(settingsPath,JSON.stringify(preferences,null,2),{mode:0o600});lastRead=0;returnrefresh(true);}
functionsetDownloadDirectory(input,{insideJob=false}={}){if(!insideJob&&[...jobs.values()].some(j=>j.state==='running'))thrownewError('Finish the current operation before changing the folder.');if(typeofinput!=='string'||!input.trim()||/[\r\n\0]/.test(input))thrownewError('Choose a model folder.');letpath;try{path=realpathSync(input.trim());}catch{thrownewError('That folder does not exist. Check the path, or that the drive is connected.');}if(!statSync(path).isDirectory())thrownewError('Choose a folder, not a file.');/* A whole drive (D:\) is not filled with models: D:\Strata is made and used. */if(isDriveRoot(path)){constown=dataFolderFromPick(path).path;makeFolder(own);path=realpathSync(own);}try{accessSync(path,constants.R_OK|constants.W_OK);}catch{thrownewError('Strata cannot write in that folder. Choose a folder you can save files in.');}preferences={...preferences,downloadDirectory:path};/* the same choice for the installer and the Terminal: the Control settings file */if(packaged)rememberDataFolder(controlSettingsFile(),path,Date.now()/1000,true);writeFileSync(settingsPath,JSON.stringify(preferences,null,2),{mode:0o600});shareDataFolder();emit();returnstate();}
functionrecordFile(model){if(!safeId(model))thrownewError('Choose an installed model.');constentry=snapshot.installed?.models?.find(i=>i.model===model);if(!entry)thrownewError('This model is not installed. Download and prepare it first.');constpath=realpathSync(entry.configPath??join(repo(),entry.config));if(dirname(path)!==repo()||!/^strata-[a-z0-9_-]+\.json$/i.test(basename(path)))thrownewError('The model configuration is outside the Strata installation.');return{entry,path};}
functionrecordFile(model){if(!safeId(model))thrownewError('Choose an installed model.');constentry=snapshot.installed?.models?.find(i=>i.model===model);if(!entry)thrownewError('This model is not installed. Download and prepare it first.');constpath=realpathSync(entry.configPath??join(repo(),entry.config));if(dirname(path)!==repo()||!/^strata-[a-z0-9_.-]+\.json$/i.test(basename(path)))thrownewError('The model configuration is outside the Strata installation.');return{entry,path};}
/* The text file a power user edits by hand: where it is and what is wrong in it (run-config-check.mjs). */
functionsettingsFile(model){const{path}=recordFile(model);return{model,path,name:basename(path),...checkRunConfig(readFileSync(path,'utf8')),note:'Changes you make in the file apply the next time the model starts.'};}
/* A mistake in the file holds the start back with the line, instead of an engine that dies on it. Warnings (an unknown key) never do. */
functionassertSettingsFile(path){constcheck=checkRunConfig(readFileSync(path,'utf8'));if(!check.ok)thrownewError(`The settings file for this model has a mistake, so the model was not started. ${describeIssues(check.issues,'error').join(' ')}`);}
functionsave(model,{values,mcp}){if([...jobs.values()].some(j=>j.state==='running'))thrownewError('Finish the current operation before saving native settings.');const{path,config}=record(model);if(busy(requests().filter(r=>instances.get(r.instanceId)?.model===model)))thrownewError('Finish this model’s requests before saving its settings.');constpatch=validateNativeSettings(values??{}),next={...config};for(constfieldofnativeFields())if(field.supported===false&&patch[field.id]!=null&&JSON.stringify(patch[field.id])!==JSON.stringify(config[field.id]))thrownewError(`${field.label} needs a newer native Strata installation. Update it in Settings → About.`);for(const[key,value]ofObject.entries(patch)){if(key==='env')next.env=restoreNativeEnv(value,config.env??{});elseif(value===null)deletenext[key];elsenext[key]=value;}if('vram_reserve_mib'inpatch){constbase=captureUserArgs(next);constmine=setUserFlag(base,'--vram-reserve-mib',patch.vram_reserve_mib);if(mine.length)next[USER_KEY]=mine;elsedeletenext[USER_KEY];next.args=mine.length&&patch.vram_reserve_mib!==null?mergeUserArgs(next.args??[],mine):setUserFlag(next.args??[],'--vram-reserve-mib',null);deletenext.ui_reserve_applied;if(Array.isArray(next.args_written))next.args_written=[...next.args];}if(mcp!==undefined){next.mcp_servers=validateMcpServers(mcp,config.mcp_servers??config.mcpServers??{});deletenext.mcpServers;}copyFileSync(path,path+'.bak');chmodSync(path+'.bak',0o600);consttemp=path+'.ui-tmp';writeFileSync(temp,JSON.stringify(next,null,2)+'\n',{mode:0o600});renameSync(temp,path);returnsettings(model);}
functioncheckedSettingsConfig(text){constcheck=checkAndParseRunConfig(text);if(!check.ok)thrownewError(`The settings file for this model has a mistake, so the model was not started. ${describeIssues(check.issues,'error').join(' ')}`);returncheck.config;}
functionsave(model,{values,mcp}){if([...jobs.values()].some(j=>j.state==='running'))thrownewError('Finish the current operation before saving native settings.');const{path,config}=record(model);if(busy(requests().filter(r=>instances.get(r.instanceId)?.model===model)))thrownewError('Finish this model’s requests before saving its settings.');constpatch=validateNativeSettings(values??{}),next={...config};if(patch.ds_experts!=null&&!dsExpertsSupported(config))thrownewError('The Speed choice needs the newer DeepSeek engine. Update the engine in Settings, then choose it.');for(constfieldofnativeFields())if(field.supported===false&&patch[field.id]!=null&&JSON.stringify(patch[field.id])!==JSON.stringify(config[field.id]))thrownewError(`${field.label} needs a newer native Strata installation. Update it in Settings → About.`);for(const[key,value]ofObject.entries(patch)){if(key==='env')next.env=restoreNativeEnv(value,config.env??{});elseif(value===null)deletenext[key];elsenext[key]=value;}if('vram_reserve_mib'inpatch){constbase=captureUserArgs(next);constmine=setUserFlag(base,'--vram-reserve-mib',patch.vram_reserve_mib);if(mine.length)next[USER_KEY]=mine;elsedeletenext[USER_KEY];next.args=mine.length&&patch.vram_reserve_mib!==null?mergeUserArgs(next.args??[],mine):setUserFlag(next.args??[],'--vram-reserve-mib',null);deletenext.ui_reserve_applied;if(Array.isArray(next.args_written))next.args_written=[...next.args];}if(mcp!==undefined){next.mcp_servers=validateMcpServers(mcp,config.mcp_servers??config.mcpServers??{});deletenext.mcpServers;}copyFileSync(path,path+'.bak');chmodSync(path+'.bak',0o600);consttemp=path+'.ui-tmp';writeFileSync(temp,JSON.stringify(next,null,2)+'\n',{mode:0o600});renameSync(temp,path);returnsettings(model);}
/* A chosen drafter that is not on this computer yet is part of the download: its size is listed beside the model's (Control decides; nothing is fetched here). */
@@ -207,6 +218,55 @@ export function createNativeControl({directory,requests=()=>[],changed=()=>{},sp
if(still.length)thrownewError(`Some files could not be removed: ${still.slice(0,3).map(p=>basename(p)).join('; ')}`);
/* The configuration a load starts the engine with (before its port and key): the saved model config, the user's choices and the resource profile. The tuner measures exactly this, so a saved tuning belongs to the same choices a load makes. */
/* DeepSeek-V4.1-Flash runs on its own engine (strata-dsv41) that takes none of the Flash-Next options the app adds on its own (request slots, the RAM budget, the output clamp, the resource profile): a flag it does not know ends the start with "unknown option". */
/* The DeepSeek engine's version, read from the folder the run config names (engine-dsv41/<version>-<platform>-...). --experts exists from engine 0.1.50: an older engine ends the start with "unknown option", so the choice is offered and passed only for 0.1.50 or newer. */
/* The app's suggested cards only fill a gap: a model whose settings file already names its GPUs ("gpu": [1, 0], with "gpu_order": "as_given") keeps them (#3256). A card the person picked in the screen always wins. */
/* The model's own choices win over what setup wrote into the arguments: "Draft layer (MTP)" off leaves --mtp out of every start, and a picture reading that is off leaves --vision out (setup writes both again whenever it runs; the saved settings are applied here, on every load). */
/* The Qwen engine's --vision stands alone, so it goes (with a picture word after it, if an older file has one). The DeepSeek engine's is a pair, `--vision on|off`, that setup wrote and the engine reads: it stays whole (taking out only the name left `off` behind: `--eos 1 off`, "unknown option off", #4358 #4359 #4505 #4507). */
for(constargofcfg.args??[])if(arg.endsWith('.gguf')&&!existsSync(resolve(repo(),arg)))thrownewError('An engine argument names a missing GGUF. Restore the saved arguments or repair the files.');
/* The app's own choice of request slots (4, fewer where the card has no room) is part of the configuration, so a tune measures the engine a load will start: slots take VRAM from the expert cache. */
/* The agent asks for long answers (a thinking model can use most of the window): let the engine clamp the output to what is left of the context instead of refusing the request. */
/* DeepSeek's Speed choice: the engine decides by itself (--experts auto: 5 when the cards total 16 GB or less, else 6), so nothing is passed unless the person chose 5 or 6. */
/* Only the options strata-dsv41 takes go to it (#4504): --vram-reserve-mib is its --reserve-mib, and --pool-workers and the other Qwen-engine names are left out, in the arguments and in the person's saved options. */
/* Which drafter runs, when it is not the one that was asked for (the engine's own words, "Drafting: MTP (DFlash2 didn't fit)"): under Details of the load. */
asyncfunctionnoteDrafting(job,instance,key){letnote='';try{constr=awaitfetcher(`http://127.0.0.1:${instance.port}/metrics`,{headers:{authorization:`Bearer ${key}`},signal:AbortSignal.timeout(2000)});if(r.ok)note=String((awaitr.json())?.engine?.drafting_note??'');}catch{/* a convenience */}constlines=[job.drafterNote,note].filter(Boolean);if(lines.length)job.details=[job.details,...lines].filter(Boolean).join('\n');}
if(externalRunning().length)thrownewError('A Strata server is already running outside this UI. Stop it before loading here, or connect its API in Settings → Models.');
if(args.multiple&&instances.size>=8)thrownewError('Unload an instance before adding another. Up to eight native instances can run here.');
/* Another model on the same graphics card: the engine takes the card's free memory, so a second one usually finds none. Loading goes on (recommend, never force); if it fails, the reason is said. */
constsharedSentence=sharedWith?`GPU ${instance.gpus.find(g=>sharedWith.gpus.includes(g))+1} is already used by ${sharedWith.label}, so this model did not fit there. Load it on another GPU, or unload ${sharedWith.label} first.`:null;
try{awaitstartWith(retryConfig);append(id,'Strata Control reports the engine is ready (second try).');}
catch(second){constsaid2=humanStartError(second);instance.status='error';instance.error=said2;/* the first stop and the second try stay in job.details; the pop-up shows the last stop once */emit();thrownewError(instance.error);}}}
proc.on('error',error=>{instance.status='error';instance.error=error.message;emit();});proc.on('exit',code=>{if(instances.has(id)&&instance.status!=='unloading'){instance.status='error';instance.error=`Engine stopped${code===null?'':` (exit ${code})`}. Open diagnostics or reload.`;emit();}});emit();
asyncfunctionbenchmark(job,args){constinstance=resolveInstance(args.instanceId);job.instanceId=instance.id;if(busy(requests().filter(r=>r.instanceId===instance.id)))thrownewError('Wait for this model to finish before benchmarking.');conststarted=Date.now();job.abort=newAbortController();constresponse=awaitfetcher(`http://127.0.0.1:${instance.port}/v1/chat/completions`,{method:'POST',headers:{'content-type':'application/json',authorization:`Bearer ${instance.key}`},signal:AbortSignal.any([job.abort.signal,AbortSignal.timeout(180000)]),body:JSON.stringify({model:instance.apiModelId,messages:[{role:'user',content:'Explain how a rainbow forms in clear, simple language.'}],temperature:0,reasoning_effort:'none',max_tokens:200,stream:false})});constbody=awaitresponse.json();if(!response.ok)thrownewError(body.error?.message??'Benchmark failed.');if(job.state!=='running')return;constresult={summary:'Measured benchmark complete.',at:newDate().toISOString(),model:instance.apiModelId,durationMs:Date.now()-started,tokensIn:body.usage?.prompt_tokens??null,tokensOut:body.usage?.completion_tokens??null,outputTps:body.timings?.predicted_per_second??null,promptTps:body.timings?.prompt_per_second??null,source:'engine',settings:'Greedy · thinking off · up to 200 output tokens'};writeFileSync(join(directory,'native-benchmark.json'),JSON.stringify(result,null,2),{mode:0o600});returnresult;}
functionresolveInstance(id){constaliases=[...instances.values()].filter(i=>i.aliases?.includes(id));if(aliases.length>1&&!instances.has(id)&&![...instances.values()].some(i=>i.apiModelId===id))thrownewError('This alias matches more than one instance. Choose its exact model ID.');consti=instances.get(id)??[...instances.values()].find(i=>i.apiModelId===id)??aliases[0]??(!id&&instances.size===1?[...instances.values()][0]:undefined);if(!i)thrownewError('Choose a loaded model ID.');if(i.status!=='ready')thrownewError('This model is not ready.');returni;}
asyncfunctionprogram(job,args){if(packaged)returncomponentProgram(job);if(externalRunning().length)thrownewError('Stop the native server running outside this UI before changing engine files.');if(instances.size)thrownewError('Unload native models before changing engine files.');constfolder=repo();let{command,args:prefix}=python(),argv;
asyncfunctionprogram(job,args){if(packaged)returncomponentProgram(job,args);if(externalRunning().length)thrownewError('Stop the native server running outside this UI before changing engine files.');if(instances.size)thrownewError('Unload native models before changing engine files.');constfolder=repo();let{command,args:prefix}=python(),argv;
if(job.kind==='calibrate'||job.kind==='tune'){if(!state().capabilities.calibration)thrownewError('This installation calibrates NVIDIA GPUs only. Use the short benchmark on other devices.');constmodel=args.model??snapshot.installed?.models?.[0]?.model;constchosen=record(model);if(!chosen.entry.ready)thrownewError('Prepare this model before calibrating.');command=venvPython(folder);prefix=[];if(!existsSync(command))thrownewError('Prepare the Python environment first.');argv=['-c','import sys; from pathlib import Path; import setup; sys.exit(0 if setup.calibrate_config(Path(sys.argv[1])) else 1)',chosen.path];}
elseif(job.kind==='rollback'){if(!previousEngine())thrownewError('No saved previous engine is available.');argv=[join(folder,'setup.py'),'--rollback-engine'];}
elseif(job.kind==='update-engine'){if(!snapshot.installed?.models?.length)thrownewError('Prepare a model first. Its installer downloads the matching engine.');argv=[join(folder,'setup.py'),'--update'];}
elseif(event.event==='waiting')job.summary='Waiting for the running model to stop…';
emit();
};
updateLog(`${job.kind} started (app ${thisAppVersion()??'unknown'})`);
if(job.kind==='install-engine-file'){if(typeofargs.path!=='string'||!args.path)thrownewError('Choose the engine file first.');letdone;try{done=awaitcontrol.installEngineFile(args.path,{onEvent,signal:job.abort.signal});}finally{try{constown=resolve(directory,'incoming-engine')+sep;if(resolve(args.path).startsWith(own))rmSync(args.path,{force:true});}catch{/* the copy in the host's own folder; a leftover is harmless */}}return{summary:updateSummary(done)==='Everything is already up to date.'?(done.results?.[0]?.note??'Nothing was changed.'):updateSummary(done),result:done};}
if(job.kind==='rollback'){constdone=awaitcontrol.rollbackEngine();return{summary:'The previous engine is back. Your models and settings are unchanged.',result:done};}
if(verdict.available){awaitappUpdater.install({signal:job.abort.signal,onProgress:p=>{if(p.total){job.summary=`Downloading the app update: ${Math.round(100*p.downloaded/p.total)}%`;job.result={install:{step:job.summary,download:{file:'Strata app',percent:Math.round(100*p.downloaded/p.total)}}};emit();}}});notes.push('The app update is installed. Strata will restart.');}
updateLog(`app check: ${verdict.available?`version ${verdict.version} is available`:'no newer app version'}`);
if(verdict.available){job.appUpdateFrom=thisAppVersion();job.appUpdateTo=verdict.version??null;emit();updateLog(`app update to ${verdict.version??'?'} starting; the installer will end this app`);awaitappUpdater.install({signal:job.abort.signal,onProgress:p=>{if(p.total){job.summary=`Downloading the app update: ${Math.round(100*p.downloaded/p.total)}%`;job.result={install:{step:job.summary,download:{file:'Strata app',percent:Math.round(100*p.downloaded/p.total)}}};emit();}}});notes.push('The app update is installed. Strata will restart.');}
/* The resource preference writes the Qwen engine's options (--pool-workers, --vram-reserve-mib); the DeepSeek engine takes neither (#4504), so its settings file is left as it is. */
if(args.calibrate!==false&&state().capabilities.calibration){job.summary='Measuring the best engine settings for this computer…';emit();awaitprogram(job,args);}
functionoperation(kind,args={}){if(!['install','load','benchmark','tune','autotune','recheck','calibrate','update-engine','update-strata','rollback','remove'].includes(kind))thrownewError('Choose a supported native operation.');if(!snapshot.available)thrownewError(snapshot.error??'Connect the native Strata installation first.');if(kind==='install'||kind==='load'){constrefusal=engineRefusal(snapshot.catalog,args.model,snapshot.engineInfo);if(refusal)thrownewError(refusal);}returnbegin(kind,job=>kind==='install'?install(job,args):kind==='load'?loadEngine(job,args):kind==='benchmark'?benchmark(job,args):kind==='tune'?tune(job,args):kind==='autotune'||kind==='recheck'?autotune(job,args):kind==='remove'?removeModel(job,args):program(job,args),{guided:args.guided===true,model:args.model,...(args.family?{family:args.family}:{})});}
functionoperation(kind,args={}){if(!['install','load','benchmark','tune','autotune','recheck','calibrate','update-engine','update-strata','rollback','install-engine-file','remove'].includes(kind))thrownewError('Choose a supported native operation.');if(!snapshot.available&&!(packaged&&COMPONENT_KINDS.includes(kind)))thrownewError(snapshot.error??'Connect the native Strata installation first.');if(kind==='install'||kind==='load'){constrefusal=engineRefusal(snapshot.catalog,args.model,snapshot.engineInfo);if(refusal)thrownewError(refusal);}returnbegin(kind,job=>kind==='install'?install(job,args):kind==='load'?loadEngine(job,args):kind==='benchmark'?benchmark(job,args):kind==='tune'?tune(job,args):kind==='autotune'||kind==='recheck'?autotune(job,args):kind==='remove'?removeModel(job,args):program(job,args),{guided:args.guided===true,model:args.model,...(args.family?{family:args.family}:{})});}
/* ---- The models folder card (0.2.2): what a picked folder means, and the three choices. The copying is Control's (strata models move). ---- */
if(renamed.length)parts.push(`${renamed.map(r=>`${r.from} is passed as ${r.to}`).join(', ')}`);
if(dropped.length)parts.push(`left out because the DeepSeek engine does not take them: ${dropped.map(d=>[d.flag,d.value].filter(v=>v!==null&&v!==undefined).join(' ')).join(', ')}`);
{id:'mtp',label:'Draft layer (MTP)',group:'Serving',type:'boolean',hint:'The draft layer guesses several words ahead, which makes answers faster, and costs about 1 GB of graphics memory and some RAM. Turn it off if the model stops while it starts, or the card is small: the expert cache gets that memory back and answers are slower. Default is on.'},
{id:'drafting',label:'Drafting',group:'Serving',type:'enum',choices:['auto','mtp','dflash2','off'],hint:'Which drafter guesses words ahead to make answers faster. Auto (recommended) lets Strata pick from what it measured on cards like yours; today that is the MTP. DFlash2 is for Qwen3.8-Flash-Next only and downloads a 1.2 GB drafter once. If DFlash2 cannot run, the MTP is used and Details says why. Off turns the draft layer off.'},
{id:'batchMtp',label:'MTP per request',group:'Serving',type:'boolean',hint:'Each concurrent request uses the draft layer. Left alone, Strata turns it on for one graphics card with 24 GB or more and several requests at once, and off otherwise; choose on or off here to decide yourself.'},
{id:'ds_experts',label:'Speed (DeepSeek)',group:'Serving',type:'enum',choices:['5','6'],forModel:'deepseek',hint:'Faster uses 5 of 6 experts per word: a small quality change, usually not noticeable. Full quality uses all 6. Left alone, the engine picks Faster on a PC whose graphics cards total 16 GB or less.'},
{id:'parallel',label:'Concurrent requests',group:'Serving',type:'number',min:1,max:8,hint:'More slots need more memory and can slow each response.'},
{id:'mcp_max_result_chars',label:'Maximum tool result characters',group:'Tools',type:'number',min:100,max:1000000},
];
/** Engine environment names the app shows and edits: STRATA_* and the DeepSeek engine's DS4_*. Any other name in a model's file is kept untouched and not shown. */
{id:'faster_long_prompts',label:'Faster long prompts',group:'Speed',
hint:'Reads very long prompts faster by approximating part of the work. Off by default: answers about long documents can miss or mix up details.',
default:'off',env:{STRATA_KVA:'1'}},
{id:'faster_12gb_nvidia',label:'Faster answers on 12 GB NVIDIA cards',group:'Speed',
hint:'Prefers the parts of the model already on the graphics card, which makes answers faster on 12 GB cards. Off by default: in long agent sessions it can repeat itself and give weaker results. Turn on if you mostly chat and want the speed.',
default:'off',env:{STRATA_ROUTE_PRIOR:'1.5'}},
{id:'faster_radeon_9000',label:'Faster answers on Radeon RX 9000 / AI PRO R9700',group:'Speed',
hint:'Shares the work between the graphics card and the processor in a way that makes answers faster on these cards. Turn off to go back to the previous setting, which is slower.',
hint:'Reads long prompts faster and loads the parts of the model it needs faster. Turn off if DeepSeek stops working or its answers look wrong; it then runs as in version 0.2.6.',
*experts)andtheenginedeclinesitonanyother,so"Faster long prompts"isshownforamodelwithoutan"arch"only.
*/
/** A discrete NVIDIA card of 13 GiB or less: the engine's rule for residency-biased routing (not an integrated chip, not a card that shares system memory). */
if(id==='faster_12gb_nvidia')return!arch&&backend==='cuda'&&smallNvidia(gpus);// the engine's routing needs the Flash-Next shape (512 experts, top-10) and a CUDA build
/** The switch is outside its default when every entry it writes is in the environment with its value; anything less or other (a hand-edited file, STRATA_KVA=0) is the default state and is left alone. */
/** What the settings answer carries: each switch that applies, with its words, its default, the entries it writes outside the default, and whether the feature is on. */
/** Settings the app itself applies when it starts a model (they are not read from the engine's source, so the engine's text never says it supports them). */
/** The engine arguments without the draft layer: --mtp <folder or count>, --spec-min-p <p> (its confidence threshold) and --batch-mtp. --spec stays: prompt lookup still uses its windows. Same rule as strata_control/draft_layer.py. */
exportfunctionwithoutDraftLayer(args){
constout=[];constlist=(args??[]).map(String);
@@ -39,7 +99,7 @@ export function validateNativeSettings(values){
for(const[key,value]ofObject.entries(values)){
if(key==='args'){if(!Array.isArray(value)||value.length>200||value.some(v=>typeofv!=='string'||v.length>4000||/[\n\r\0]/.test(v))||value.some(v=>['--host','--port','--api-key'].includes(v)))thrownewError('Engine arguments must be text values. Network access stays managed by Strata.');result.args=[...value];continue;}
if(key==='env'){if(!value||typeofvalue!=='object'||Array.isArray(value))thrownewError('Engine environment must be an object.');constenv={};for(const[name,text]ofObject.entries(value)){if(!/^STRATA_[A-Z0-9_]+$/.test(name)||typeoftext!=='string'||text.length>4000||text.includes('\0'))thrownewError('Use STRATA_ environment flags with text values.');env[name]=text;}result.env=env;continue;}
if(key==='env'){if(!value||typeofvalue!=='object'||Array.isArray(value))thrownewError('Engine environment must be an object.');constenv={};for(const[name,text]ofObject.entries(value)){if(!ENV_FLAG.test(name)||typeoftext!=='string'||text.length>4000||text.includes('\0'))thrownewError('Use STRATA_ environment flags with text values.');env[name]=text;}result.env=env;continue;}
if(key==='sampling'){if(!value||typeofvalue!=='object'||Array.isArray(value))thrownewError('Sampling must be an object.');constlimits={temperature:[0,10],top_p:[Number.MIN_VALUE,1],top_k:[1,64],min_p:[0,1],repetition_penalty:[Number.MIN_VALUE,10],presence_penalty:[0,10],frequency_penalty:[0,10],penalty_last_n:[0,1048576],seed:[1,2147483647]};for(const[name,n]ofObject.entries(value)){if(name==='experimental_speed_projection'){if(typeofn!=='boolean')thrownewError('Speed projection must be on or off.');continue;}if(!limits[name]||!Number.isFinite(n)||(['top_k','seed','penalty_last_n'].includes(name)&&!Number.isInteger(n))||n<limits[name][0]||n>limits[name][1])thrownewError(`Choose a valid sampling ${name}.`);}result.sampling={...value};continue;}
exportfunctionrestoreNativeEnv(env,previous={}){constresult=Object.fromEntries(Object.entries(previous).filter(([name])=>!/^STRATA_[A-Z0-9_]+$/.test(name)));for(const[name,value]ofObject.entries(env)){if(value===MASK){if(previous[name]===undefined)thrownewError(`Enter a new value for ${name}.`);result[name]=previous[name];}elseresult[name]=value;}returnresult;}
exportfunctionrestoreNativeEnv(env,previous={}){constresult=Object.fromEntries(Object.entries(previous).filter(([name])=>!ENV_FLAG.test(name)));for(const[name,value]ofObject.entries(env)){if(value===MASK){if(previous[name]===undefined)thrownewError(`Enter a new value for ${name}.`);result[name]=previous[name];}elseresult[name]=value;}returnresult;}
if(appNew||controlNew)Object.assign(strata,{status:'available',latestVersion:appNew??controlNew,message:`A new version of Strata is available.${appNotes?`${appNotes}`:''}${appNew&&appCannot?`${appCannot}`:''}`});
constappCheckLine="Strata couldn't check for a new version of the app itself. Try again later.";
if(appNew||controlNew)Object.assign(strata,{status:'available',latestVersion:appNew??controlNew,message:`A new version of Strata is available.${appNotes?`${appNotes}`:''}${appNew&&appCannot?`${appCannot}`:''}${appCheckFailed?`${appCheckLine}`:''}`});
/** The C# that asks D3DKMT (KMTQAITYPE_ADAPTERPERFDATA = 62, 64-byte D3DKMT_ADAPTER_PERFDATA, Temperature at byte 56) for one LUID; -1 = no answer. */
exportconstGPU_TEMP_CSHARP='using System;using System.Runtime.InteropServices;public static class StrataGpuTemp{[StructLayout(LayoutKind.Sequential)]struct Q{public uint L;public int H;public uint A;}[StructLayout(LayoutKind.Sequential)]struct Info{public uint A;public int T;public IntPtr P;public uint S;}[DllImport("gdi32.dll")]static extern int D3DKMTOpenAdapterFromLuid(ref Q q);[DllImport("gdi32.dll")]static extern int D3DKMTCloseAdapter(ref uint a);[DllImport("gdi32.dll")]static extern int D3DKMTQueryAdapterInfo(ref Info i);public static int Read(int high,uint low){Q q=new Q();q.L=low;q.H=high;if(D3DKMTOpenAdapterFromLuid(ref q)!=0)return -1;IntPtr buf=Marshal.AllocHGlobal(64);try{for(int k=0;k<64;k++)Marshal.WriteByte(buf,k,0);Info i=new Info();i.A=q.A;i.T=62;i.P=buf;i.S=64;if(D3DKMTQueryAdapterInfo(ref i)!=0)return -1;return Marshal.ReadInt32(buf,56);}finally{Marshal.FreeHGlobal(buf);uint a=q.A;D3DKMTCloseAdapter(ref a);}}}';
`$u=@{};foreach($r in $e){if($r.Name -match '(luid_0x[0-9A-Fa-f]+_0x[0-9A-Fa-f]+)_phys_\\d+_eng_\\d+_engtype_(\\w+)'){$k=$matches[1].ToLower()+'|'+$matches[2];$u[$k]=[double]$u[$k]+[double]$r.UtilizationPercentage}};`+
`$b=@{};foreach($k in $u.Keys){$l=$k.Split('|')[0];if(-not $b.ContainsKey($l) -or $u[$k] -gt $b[$l]){$b[$l]=$u[$k]}};`+
`foreach($r in $m){if($r.Name -match '(luid_0x[0-9A-Fa-f]+_0x[0-9A-Fa-f]+)'){$l=$matches[1].ToLower();$d=-1;if($t){try{$p=$l.Split('_');$d=[StrataGpuTemp]::Read([Convert]::ToInt32($p[1].Substring(2),16),[Convert]::ToUInt32($p[2].Substring(2),16))}catch{}};[Console]::Out.WriteLine('G '+$l+' '+[math]::Round([double]$b[$l],0)+' '+$r.DedicatedUsage+' '+$d)}};`+
{tr(!groups.length?<pclassName="model-picker-empty">{tr("No local models downloaded yet.")}</p>:null)}
{tr(installedNative.map(m=><buttonkey={m.config||m.model}className="menu__item model-picker-option model-picker-api"role="menuitem"disabled={streaming}onClick={()=>{setOpen(false);go('models');}}><span><strong>{tr(m.model_name||m.model)}</strong><small>{tr("Downloaded, not loaded. Open Models to load it.")}</small></span></button>))}
{tr(!groups.length&&!installedNative.length?<pclassName="model-picker-empty">{tr("No local models downloaded yet.")}</p>:null)}
<Dialogopen={toolsOpen}compacttitle={tr("Enable connected tools?")}onClose={()=>setToolsOpen(false)}actions={<><BtnonClick={()=>setToolsOpen(false)}>{tr("Keep tools off")}</Btn><Btnvariant="primary"onClick={()=>{setChatPreferences(session?.id,{useTools: true,mode:'bypass'});setToolsOpen(false);}}>{tr("Enable tools")}</Btn></>}><p>{tr("Connected servers can act on the resources you configured. Calls run automatically when the model chooses them, and their results appear in chat.")}</p><pclassName="small muted">{tr("This switches this conversation to Bypass permissions. Servers: ")}{tr((nativeModel?.metrics?.mcp?.servers??[]).map((s: any)=>s.name).join(', '))}.</p></Dialog>
<buttonref={trigger}className="composer-add"aria-label={tr("Add or open")}title={tr("Add or open")}disabled={reading||streaming}aria-haspopup="menu"aria-expanded={menu}onClick={()=>{setPreferences(null);setMenu(v=>!v);}}><Iconname="plus"size={22}/></button>
/** A number field that lets the person type a decimal point: it keeps the text while it is typed and tells the page the number once it is one (#3736). */
<inputclassName="st-input"inputMode="decimal"size={7}aria-label={tr("Price per kWh")}placeholder="0.30"value={text}onChange={e=>setText(e.target.value)}onBlur={commit}onKeyDown={e=>{if(e.key==='Enter')commit();}}/>
<spanclassName="small muted">{money.currency?tr("per kWh, in {0}",[currencyLabel(money.currency,money.locale)]):tr("per kWh")}</span>
body=<><divclassName="onboard-welcome-copy"><h1tabIndex={-1}data-initial-focus>{tr("The engine is ready.")}</h1><p>{tr("Next: pick a model.")}</p></div>
body=<><divclassName="onboard-welcome-copy"><h1tabIndex={-1}data-initial-focus>{tr("Downloading the engine")}</h1>
<p>{bytes&&job?.result?.install?.download?`${doneMb} of ${mb(bytes)} MB${eta?' · '+eta:''}`:tr(job?.summary??'Starting…')}</p></div>
<divclassName="onboard-error"style={{maxWidth: 520,margin:'0 auto',width:'100%'}}><Barvalue={percent}label="Downloading the engine"/><pclassName="small muted"style={{textAlign:'center'}}>{tr("You can close Strata; it continues where it stopped.")}</p></div></>;
}
elseif(!info){
body=<divclassName="onboard-welcome-copy"><h1tabIndex={-1}data-initial-focus>{tr("Welcome to Strata")}</h1><prole="status">{tr("Looking at your graphics card…")}</p></div>;
}
else{
constcard=gpuLine(info);
constfound=card?<p>{tr("Strata found your graphics card:")}<br/><strong>{card}</strong>{cards>1?<><br/><spanclassName="small muted">{tr("{0} cards found. Strata uses the first one for the engine; change it later in Settings.",[cards])}</span></>:null}</p>
:<p>{tr("Strata did not recognise a supported graphics card.")}</p>;
letmain,actions;
if(info.status==='offer'&&info.offer){
main=<>{found}<p>{tr("To run models on it, Strata needs one download:")}<br/><strong>{tr("The engine for this card")}:{mb(info.offer.bytes)}MB</strong></p>
{info.freeBytes!==undefined&&info.freeBytes<info.offer.bytes*2.5?<pclassName="small inline-error">{tr("The engine needs about {0} GB of free space on this drive and only {1} GB is free. Free some space, then continue.",[Math.ceil(info.offer.bytes*2.5/1e9),(info.freeBytes/1e9).toFixed(1)])}</p>:null}
<pclassName="small muted">{tr("It comes from Strata’s own update server and is checked before it is used. Your models and chats are never touched by it.")}</p></>;
actions=<><Btnvariant="primary"disabled={busy}onClick={()=>voidstart()}>{tr(failed?'Continue':'Download the engine')}</Btn>{pick("I already have the file…")}</>;
}
elseif(info.status==='offline'){
main=<>{found}<p>{tr("Strata could not reach github.com. Check your connection and try again, or download the engine file on another PC and choose it here.")}</p></>;
actions=<><Btnvariant="primary"disabled={busy}onClick={look}>{tr("Try again")}</Btn>{pick("Choose a downloaded file…")}</>;
}
elseif(info.status==='wrong-make'){
main=<>{found}<p>{info.message??tr("The engine installed here is for another make of graphics card.")}</p></>;
actions=<><Btnvariant="primary"disabled={busy}onClick={look}>{tr("Try again")}</Btn>{pick("Choose a downloaded file…")}</>;
}
else{
main=<>{found}<p>{tr("Strata can’t pick an engine for this PC from its update list.")}</p>{info.message?<pclassName="small muted">{info.message}</p>:null}{(info.reasons??[]).length?<ulclassName="small muted"style={{textAlign:'left',display:'inline-block'}}>{(info.reasons??[]).map(r=><likey={r}>{r}</li>)}</ul>:null}</>;
actions=<><Btnvariant="primary"disabled={busy}onClick={look}>{tr("Try again")}</Btn>{pick("Choose a downloaded file…")}<buttontype="button"className="onboard-continue-anyway"onClick={onDone}>{tr("Continue without an engine")}</button></>;
}
body=<><divclassName="onboard-welcome-copy"><h1tabIndex={-1}data-initial-focus>{tr("Welcome to Strata")}</h1>{main}</div>
{failed?<divclassName="onboard-error"style={{maxWidth: 560,margin:'0 auto'}}><prole="alert"className="small inline-error">{job?.error??job?.summary}</p><pclassName="small muted">{tr("The download stopped. Continue picks it up where it stopped.")}</p></div>:null}
conststeps=['Welcome','Your model','Prepare','Make it yours'];
/** What a person is told before choosing DeepSeek: the size, where it runs, how fast, what it needs. Every line is a warning or a fact; nothing here stops the download. */
<p>{tr("This is a very large model. The download is {0} GB and it needs about as much free disk space. It runs on NVIDIA RTX 30, 40 and 50 cards, on a Radeon AI PRO R9700 (Linux) or on a Ryzen AI Max / Strix Halo PC (Windows); how fast it is depends on the memory and card (the line below says what to expect on this computer). What does not fit in memory is read from your SSD, so an NVMe drive matters.",[Math.round(version.download_gb)])}</p>
<pclassName="small muted">{tr("You can stop the download and continue later; it picks up where it stopped. Strata warns you here but never blocks you.")}</p>
@@ -201,9 +211,10 @@ export function GuidedSetup() {
<divclassName="onboard-stage-heading"><h1ref={titleRef}tabIndex={-1}data-initial-focus>{tr(title)}</h1>{tr(step===2?<p>{tr("You can always change your models later.")}</p>:step===4&&tuneJob?.state==='running'?<p>{tr("We\u2019re measuring, applying your preference, and getting ready to chat.")}</p>:null)}</div>
<OnboardingMemoryhardware={hardware}native={native?.hardware}/>{tr(!version||version.needsNewerEngine?<pclassName="small muted">{tr("This model needs a newer engine. Update your engine or import a local GGUF for this model.")}</p>:null)}
<divclassName="onboard-version-heading"><strong>{tr("Choose a version")}</strong><span>{tr(version?`${catalogVersionLabel(version)} selected`:'')}</span></div>
<divclassName="onboard-folder"><Iconname="folder"size={20}/><div><strong>{tr(directory?'Your model folder':'Already downloaded the model?')}</strong><small>{tr(directory||"Choose the folder and we'll start your models with Strata.")}</small></div><Btnsize="sm"onClick={()=>setFolderOpen(v=>!v)}>{tr(folderOpen?'Close':directory?'Change':'Choose folder')}</Btn></div>
constnoFit=selectedSize?.on_this_pc==='does not fit';
// Graphics cards (strata_control/gpu_fit.py: the one rule of the app, the Terminal and setup): the recommended option is pre-selected, the person may pick another
// A card a loaded model already uses (a second model on it usually does not fit): said on its row, and not the pre-selected one while another is free.
{tr(noFit?<pclassName="small muted native-spill-note"><Iconname="disk"size={16}/>{tr("May run slower \u00B7 Model data can be read from SSD.")}</p>:null)}
<divclassName="native-choice__col"><spanclassName="native-choice__label">{tr("Computer use")}</span><divclassName="native-resource-choice"><buttonaria-pressed={resourceProfile==='work'}data-selected={resourceProfile==='work'}onClick={()=>setResourceProfile('work')}><Iconname="computer"size={20}/><span><strong>{tr("Leave 20% for my work")}</strong><small>{tr("Headroom for your other apps")}</small></span>{tr(resourceProfile==='work'?<Iconname="check"size={15}/>:null)}</button><buttonaria-pressed={resourceProfile==='performance'}data-selected={resourceProfile==='performance'}onClick={()=>setResourceProfile('performance')}><Iconname="bolt"size={20}/><span><strong>{tr("Full performance")}</strong><small>{tr("Give Strata priority")}</small></span>{tr(resourceProfile==='performance'?<Iconname="check"size={15}/>:null)}</button></div></div>
{tr(expert?<sectionid="native-expert-settings"ref={expertSection}className="st-card native-execution"><h2className="h4">{tr("Expert settings")}</h2>{tr(tunedNow?<pclassName="small muted native-tuned-note">{tr("Showing the values of your tuning '{0}'.",[tunedNow.name])}</p>:null)}<labelclassName="native-toggle"><span><strong>{tr("Run multiple models")}</strong><small>{tr("Loading another model keeps the others loaded.")}</small></span><buttonclassName="st-toggle"role="switch"aria-label={tr("Run multiple models")}aria-checked={multiple}onClick={()=>setMultiple(v=>!v)}/></label>{tr(cardsGpus.length>1?<labelclassName="input-field"><span>{tr("Layer split")}</span><inputclassName="st-input"value={split}placeholder={tr("auto or 16,32")}onChange={e=>setSplit(e.target.value)}/></label>:null)}<divclassName="native-field-grid"><labelclassName="input-field"><span>{tr("Context memory")}</span><SelectariaLabel={tr("Native KV format")}value={kv}options={[{value:'int8',label:'INT8'},{value:'q4_0',label:'Q4'},{value:'k8v4',label:'K8 / V4'}]}onChange={setKv}/></label><labelclassName="input-field"><span>{tr("Low-RAM mode")}</span><SelectariaLabel={tr("Native low RAM mode")}value={lowRam}options={['auto','resident','mmap'].map(v=>({value: v,label: v==='auto'?'Automatic':v==='resident'?'Keep experts in RAM':'Read experts from disk'}))}onChange={setLowRam}/></label></div>{tr(installed?<><NativeSettingsFilemodel={installed.model}/><NativeSettingsmodel={installed.model}/></>:<pclassName="small muted">{tr("More engine parameters become available after preparing this version.")}</p>)}</section>:null)}
</div>
<asideclassName="stack stack--lg"><PcFitCardgpuName={pcGpu?.name}ramGb={pcRam}vramGb={pcVram}inUse={keptModels.length?{ramGb: Math.round(keptRam),vramGb: Math.round(keptVramOn(chosenCards))}:null}fit={selectedSize?fitOn(selectedSize,fitHardware):null}/><sectionclassName="st-card"><h2className="h4">{tr("Model files")}</h2><pclassName="small muted">{tr("Strata prepares all shards, the draft layer and shared files. Completed downloads are reused.")}</p><codeclassName="native-path">{installed?.dataFolder??native?.downloadDirectory??native?.installed?.data_folder??'Choose an installation'}</code><Btnsize="sm"variant="ghost"onClick={()=>{setSettingsSection('Storage');go('settings');}}>{tr("Storage settings")}</Btn></section><Btnsize="sm"variant="ghost"onClick={onOtherModels}>{tr("Other local models\u2026")}</Btn></aside>
<asideclassName="stack stack--lg"><PcFitCardgpuName={pcGpu?.name}ramGb={pcRam}vramGb={pcVram}diskGb={selectedSize?.multi_file?selectedSize.disk_gb??selectedSize.download_gb : null}inUse={keptModels.length?{ramGb: Math.round(keptRam),vramGb: Math.round(keptVramOn(chosenCards))}:null}fit={selectedSize?fitOn(selectedSize,fitHardware,true,{context,kv}):null}/><sectionclassName="st-card"><h2className="h4">{tr("Model files")}</h2><pclassName="small muted">{tr("Strata prepares all shards, the draft layer and shared files. Completed downloads are reused.")}</p><codeclassName="native-path">{installed?.dataFolder??native?.downloadDirectory??native?.installed?.data_folder??'Choose an installation'}</code><Btnsize="sm"variant="ghost"onClick={()=>{setSettingsSection('Storage');go('settings');}}>{tr("Storage settings")}</Btn></section><Btnsize="sm"variant="ghost"onClick={onOtherModels}>{tr("Other local models\u2026")}</Btn></aside>
</div>
<Dialogopen={!!plan}title={tr(installed?.ready?"Apply settings and load":"Download and prepare")}compactonClose={()=>!pending&&setPlan(null)}actions={<><Btndisabled={pending}onClick={()=>setPlan(null)}>{tr("Cancel")}</Btn><Btnvariant="primary"disabled={pending||!!plan?.plan?.disk_short}onClick={()=>voiddownload()}>{tr(installed?.ready?"Apply & load":"Download & load")}</Btn></>}><p><strong>{tr(plan?.plan?.title??selectedFamily?.title)}</strong>{tr(installed?.ready?null:<>{tr(" \u00B7 approximately ")}{tr(plan?.plan?.download_gb??selectedSize?.download_gb)}{tr(" GB")}</>)}</p>{tr(plan?.drafting?.needed?<pclassName="small muted">{tr("Choosing DFlash2 also downloads its drafter: about {0} GB, once. If it does not fit in graphics memory, the model uses the MTP instead.",[plan.drafting.gb])}</p>:null)}{tr(installed?.ready?<pclassName="small muted">{tr("The model files are already on this computer. Nothing will be downloaded; only the settings you chose are applied.")}</p>:null)}{tr(plan?.plan?.disk_short?<ErrorNoticemessage={plan.plan.disk_short}/>:null)}{tr(installed?.ready?<pclassName="small muted">{tr("Any currently loaded native models will be unloaded first.")}</p>:<><pclassName="small muted">{tr("Existing files are checked and reused. Any currently loaded native models will be unloaded before preparing files.")}</p><pclassName="small muted">{tr("Your computer may feel slower while weights load. You can stop the download and resume later.")}</p></>)}{tr(selectedSize?.images?<pclassName="small muted">{tr("Vision encoder files are included. Picture reading loads only when enabled.")}</p>:null)}</Dialog>
/** One of the speed features: what it is called, what it does, whether it is on by default, and the engine environment it writes outside its default (server/native-settings.mjs). */
interfaceSpeedSwitch{
id: string;
label: string;
hint: string;
default:'on'|'off';
env: Record<string,string>;
/** What an earlier test build wrote for the same state: it reads as set, and goes with it. */
@@ -90,16 +114,10 @@ export function NativeSettings({ model }: {
<summary>{tr('Native engine settings')}</summary>
{error?<><ErrorNoticemessage={tr('Couldn’t load model settings.')}next={tr('Check that the engine is connected and this model is installed, then try again.')}onRetry={()=>setRevision(value=>value+1)}/><detailsclassName="native-settings-diagnostic"><summaryclassName="small">{tr('Details')}</summary><pclassName="small muted">{error}</p></details></>:<pclassName="small muted"role="status">{tr('Reading saved settings…')}</p>}
}}/></label>))}</section></>:tab==='json'?<><pclassName="small muted">{tr("Engine arguments and STRATA environment flags are editable here. Program paths and network access remain managed by the host.")}</p><textareaclassName="st-input native-json"aria-label={tr("Native configuration JSON")}spellCheck={false}value={text}onChange={e=>{setText(e.target.value);setNotice('');}}/></>:<><pclassName="small muted">{tr("Add local programs or HTTP MCP servers. Connected tools can read or change the resources you give them access to. Enable tools in chat when you want the model to use them.")}</p><NativeMcpEditortext={mcp}onChange={value=>{setMcp(value);setNotice('');}}/><details><summaryclassName="small">{tr("Configuration example")}</summary><preclassName="native-example">{'{\n "files": {\n "command": "npx",\n "args": ["-y", "@modelcontextprotocol/server-filesystem", "/your/folder"],\n "disabled": false\n }\n}'}</pre><pclassName="xs muted">{tr("Saved credentials are masked. Keep the mask to preserve an existing value; type a new value to replace it.")}</p></details></>)}
{tr(tab==='fields'?<><inputclassName="st-input"type="search"aria-label={tr("Search native parameters")}placeholder={tr("Search parameters…")}value={query}onChange={e=>setQuery(e.target.value)}/>{tr(['Responses','GPU & memory','Serving','Image reading','Tools','Recovery'].map(group=>rows.some(f=>f.group===group)?<sectionclassName="native-settings-group"key={group}><h3className="h4">{tr(group)}</h3>{tr(rows.filter(f=>f.group===group).map(f=><labelclassName="native-setting-row"key={f.id}><span><strong>{tr(f.label)}</strong>{tr(f.supported===false?<small>{tr("Requires a newer Strata installation.")}</small>:f.hint?<small>{tr(f.hint)}</small>:null)}</span><NativeFieldInputfield={f}value={values[f.id]}onChange={v=>change(f.id,v)}/></label>))}</section>:null))}{tr(switchRows.length?<sectionclassName="native-settings-group"><h3className="h4">{tr("Speed")}</h3>{tr(switchRows.map(sw=><labelclassName="native-setting-row"key={sw.id}><span><strong>{tr(sw.label)}</strong><small>{tr(sw.hint)}</small></span><SelectariaLabel={tr(sw.label)}value={switchIsOn(sw,values.env)?'on':'off'}options={sw.default==='on'?[{value:'on',label:'On (recommended)'},{value:'off',label:'Off'}]:[{value:'off',label:'Off (recommended)'},{value:'on',label:'On'}]}onChange={v=>changeSwitch(sw,v==='on')}/></label>))}</section>:null)}<sectionclassName="native-settings-group"><h3className="h4">{tr("Sampling")}</h3>{tr([['temperature','Temperature'],['top_p','Top-p'],['top_k','Top-k'],['min_p','Min-p'],['repetition_penalty','Repetition penalty'],['presence_penalty','Presence penalty'],['frequency_penalty','Frequency penalty'],['penalty_last_n','Penalty window'],['seed','Seed']].map(([id,label])=><labelclassName="native-setting-row"key={id}><span><strong>{tr(label)}</strong></span><DecimalInputariaLabel={tr("Native {0}",[label])}value={values.sampling?.[id]??null}onChange={n=>{constsampling={...values.sampling};if(n===null)deletesampling[id];elsesampling[id]=n;change('sampling',sampling);}}/></label>))}</section></>:tab==='json'?<><pclassName="small muted">{tr("Engine arguments and STRATA environment flags are editable here. Program paths and network access remain managed by the host.")}</p><textareaclassName="st-input native-json"aria-label={tr("Native configuration JSON")}spellCheck={false}value={text}onChange={e=>{setText(e.target.value);setNotice('');}}/></>:<><pclassName="small muted">{tr("Add local programs or HTTP MCP servers. Connected tools can read or change the resources you give them access to. Enable tools in chat when you want the model to use them.")}</p><NativeMcpEditortext={mcp}onChange={value=>{setMcp(value);setNotice('');}}/><details><summaryclassName="small">{tr("Configuration example")}</summary><preclassName="native-example">{'{\n "files": {\n "command": "npx",\n "args": ["-y", "@modelcontextprotocol/server-filesystem", "/your/folder"],\n "disabled": false\n }\n}'}</pre><pclassName="xs muted">{tr("Saved credentials are masked. Keep the mask to preserve an existing value; type a new value to replace it.")}</p></details></>)}
{tr(inUse?<pclassName="xs muted">{tr("Models that stay loaded use about {0} GB of RAM and {1} GB of this card’s memory.",[inUse.ramGb,inUse.vramGb])}</p>:null)}
{tr(fit?.known?<pclassName="pc-note">{tr(fit.spillGb>=1?'About {0} GB will be read from the SSD, so answers can be a little slower.':'Fits fully in memory.',[Math.round(fit.spillGb)])}</p>:null)}
{tr(diskGb?<pclassName="pc-note">{tr("Needs about {0} GB of free disk space. What does not fit in memory is read from your SSD.",[Math.round(diskGb)])}</p>:null)}
{tr(fit?.known&&(diskGb?fit.spillGb>=1 : true)?<pclassName="pc-note">{tr(fit.spillGb>=1?'About {0} GB will be read from the SSD, so answers can be a little slower.':'Fits fully in memory.',[Math.round(fit.spillGb)])}</p>:null)}
// ONE list and ONE total: the optimized models (sizes from their real files, shared files counted once) and the other local models together.
constnativeModels=native?.installed?.models??[];
constnativeModels=(native?.installed?.models??[]).filter(m=>m.ready||(m.diskGb??0)>0);// a model whose files are all gone (an old config) is not a 0.0 GB download
{tr(!checking&&update?.status==='available'?(native?.available?<Btnsize="sm"variant="primary"disabled={!!active||(name==='Engine'&&!native.installed?.models.length&&!native.capabilities.componentUpdates)}onClick={()=>setConfirm(name==='Strata'?'update-strata':'update-engine')}>{tr("Update available")}</Btn>:update.releaseUrl?<aclassName="st-btn st-btn--secondary st-btn--sm"href={update.releaseUrl}target="_blank"rel="noreferrer"title={tr("Open the release page to update this installation")}>{tr("Update available")}<Iconname="download"size={15}/></a>:<spanclassName="settings-version__label">{tr(labels[update.status])}</span>):<spanclassName="settings-version__label">{tr(checking?'Checking…':update?labels[update.status]:version==='Not installed'?'Not installed':'Installed')}</span>)}
<divclassName="settings-update-actions"><Btnsize="sm"icon="refresh"disabled={checking}className={checking?'settings-update-checking':undefined}onClick={()=>voidcheck()}>{tr(checking?'Checking…':error||result?.strata.status==='error'?'Try again':result?'Check again':'Check for updates')}</Btn><Btnsize="sm"variant="ghost"onClick={()=>setRollback(true)}>{tr("Roll back\u2026")}</Btn></div>
<divclassName="settings-update-inline"role="status"aria-live="polite">{tr(checking?'Checking for new versions…':error?<spanclassName="inline-error">{tr(error)}</span>:result?<><spanclassName="sr-only">{tr(result.strata.message)}{tr(result.engine.message)}</span>{tr("Checked at ")}{tr(checkedLabel(result.checkedAt))}</>:null)}</div>
</div>
@@ -95,11 +96,11 @@ export function SettingsUpdates() {
</>;
}
/** "Strata 0.2.3 · Engine 0.1.48": the app's own version first, the engine's beside it (About). */
/** "Strata 0.2.8 · Engine 0.1.54": the app's own version first, the engine's beside it (About). */
return<buttonclassName="st-toggle"role="switch"aria-checked={shell.minimizeToTray}aria-label={tr("Minimise to tray")}onClick={()=>voidsave({tray:!shell.minimizeToTray})}/>;
}
/** The port the app's window and local address use (#3402). Empty lets Strata choose; the choice applies the next time Strata starts. */
<strong>{selected&&!apart.has(selected.id)?versionName(selected,true):tr(moreLabel??"Choose a version")}</strong>{tr(selected&&!apart.has(selected.id)&&downloaded?.has(selected.id)?<spanclassName="version-row__have">{tr("Downloaded")}</span>:null)}
<spanclassName="version-trigger__need muted">{tr(fit&&selected?.multi_file&&selected.disk_gb?tr("{0} RAM + VRAM, plus about {1} GB of disk",[gb(fit.memoryGb),Math.round(selected.disk_gb)]):fit?`${gb(fit.memoryGb)} RAM + VRAM`:'')}</span>
if(started&&typeof(startedasPromise<void>).catch==='function')(startedasPromise<void>).catch(()=>{/* the browser asked for a click first: no sound this time */});
returntrue;
}catch{
returnfalse;
}
}
/** Called when an answer or a task is finished. */
/** A file that is not a picture or plain text (a PDF) is saved by the host in the chat's folder and named in the next message, so the assistant can open it (#3838). */
/** A number typed by a person: "0.95", ".95", "0,95" (a comma is a decimal point in many languages). null while it is not a whole number yet ("", ".", "0.", "-"). */
/** The largest size not flagged as untested; also how far the list goes while the pack does not say more (a pack without ds4.txt behaves as before). */
exportconstTESTED_CONTEXT=262144;
/** The model's own length: the most any pack can say. */
exportconstNATIVE_CONTEXT=1048576;
exportconstUNTESTED_NOTE='Not tested yet. A context this long takes graphics memory from the model and reads long prompts slowly.';
/** The sizes to offer: up to the pack's limit (`contextMax`, from its ds4.txt) when the installed model has one, else up to 256K. */
/** Decimal megabytes (1 MB = 1,000,000 bytes, at least 1): the same figure as the Terminal and the release page. The only size helper in MB; no binary units. */
/* The Tune model dialog's strings live in tune-<lang>.json (apps/host/scripts/make-tune-locales.py) so other interface work does not collide with them. */
/* The tool-server, picture and agent-result strings of the tools and images fixes live in tools-<lang>.json for the same reason. */
/** Optimized models that are on this PC but not loaded yet: the chat picker must list them, or it says "none downloaded" while Storage shows them (#3814). */
Loaded 100 of 929 files, more files were not shown because too many files have changed in this diff.
Show more
Reference in new issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Niko1221