Coder: expert-pruned 256-expert Coder, M5 Max, 5 interleaved pairs, identical greedy tokens in all 40 runs per class, over --spec 2: story -2.5%
(0/5), code +7.1%, JSON +9.3%, agent turn +12.5% (5/5 each). It has no Mac setup entry yet; the name is what a config for it would pass.
fp16 decode (PR 4807): setup writes STRATA_METAL_F16=1 into the config's env for a resident IQ3_S or Q2_0 pack (STRATA_METAL_F16=0 there turns
it off; streaming and the other sets are untouched). The owner's rule was an option unless it passes the 61-task quality suite: off / on / off
again on IQ3_S 54/61 each, on Q2_0 54/61 each, the same 7 tasks failing in all arms, no family drop, no new hallucination flag, 0 pass/fail
changes between two runs of the same arm; decode +14% (IQ3_S) and +15% (Q2_0) in the suite runs, top-1 96-97% and mean KL 0.009-0.016 against fp32
(phaseA-tests/mac-q/RESULTS.md sections 20, 22, 30).
The engine gives 5 experts per word on two or more RDNA4 cards that together hold at least half of the experts (4 x R9700), but strata launch --plan, the Speed screen and the app's Speed label called Full quality the default there.
dsv41_fit.rdna4_faster() applies the engine's rule with the plan's own numbers: one card the run uses (on Windows one AMD card): yes; several: yes when the plan puts at least half of the set's experts on them (3 or more R9700 for every official set; two hold 37-47%, so they keep 6). strata launch uses it with the chosen model; the model catalog carries rdna4_faster per DeepSeek size for the app (NativeSize, Speed label); --speed help and docs/MODELS.md say it. Other makes and older Radeons are unchanged.
One R9700 holds under a fifth of a 2.25 bpw set, so the CPU lane and the SSD set the speed, as on the 16 GiB cards and the multi-card RDNA4 rule that already take 5.
1 x R9700 on p620 (125 GB RAM), 2.25 bpw, frozen expert profile, 5 interleaved pairs, 5 against 6 experts: decode 15.24 against 13.08 tok/s (+16%, 5 of 5 pairs), a scripted agent loop -24% wall time (5 of 5), the long-read loop -12% (5 of 5). KL against the 6-expert run 0.030 (6 against 6: 0.008); against the original model 0.1106 against 0.1067-0.1071. quality027 quick: 18 of 19 with 5 experts and with 6, the same task flagged in both.
ds4_experts_rule.hpp: experts_per_word() takes single_rdna4 (the engine passes rdna4() with one device); DS4_DEFAULTS=026 gives the model's 6 back. The Terminal's Speed words (strata launch --plan and screens), the app's Speed label (arch added to RunCard) and docs/MODELS.md say the same. Several RDNA4 cards keep the existing rule (they must hold half of the experts); the screens still do not mirror that one.
Port of pull request 4737 by InB4DevOps (commit c14ee15d9 on v0.2.8). The mixer's input-only QSA/GDN work runs on graph branches; scheduling only, no arithmetic change.
STRATA_SM121_PERF=0, STRATA_DEFAULTS=026 or STRATA_SIDE_MIXER=0 turn it off. Measured here on a Spark, IQ3_S, greedy: ids identical 48 of 48; alone +1.0..+2.9% per class, with mode 8 +3.6% pooled.
Not yet measured at parallel 4.
Mode 2 of STRATA_SELECT_SIMT was added for HIP in the previous commit as an opt-in. For the RX 6000 cards its contributor measured it
on (gfx103x) it now joins the one community switch: on by default with STRATA_AMD_COMMUNITY_TUNING on, any STRATA_SELECT_SIMT value
the user sets wins, and no other card changes. Values change in the last bits (another summation order); the contributor's numbers are
labelled as theirs in docs/AMD_HIP.md and have not been reproduced (no gfx103x card here). amd_community_test checks the default and
that the switch turns it off.
Unset, the output is unchanged (the build with the change, unset, gives the same greedy ids as the build without it in 12 of 12 runs, gemv_parity / gr_parity / expert_parity pass). With STRATA_METAL_F16=1 the decode windows of a resident model use half-precision GEMVs (Q2_0, Q3_K-Q6_K, IQ4_XS) and the Q2_0 routed experts.
M5 Max, spec 3 + min-p 0.5, 3 interleaved pairs, decode tok/s unset -> on: Q2_0 story 105.0 -> 114.9 (+9.4%), code 125.6 -> 143.7 (+15.5%), json 130.8 -> 135.1 (+3.3%, one noisy pair), agent 116.1 -> 131.6 (+13.4%); IQ3_S (dense GEMVs only) story 93.1 -> 104.3 (+8.8%), code 103.6 -> 118.5 (+12.5%), json 122.5 -> 135.4 (+10.6%), agent 103.8 -> 117.8 (+12.3%).
It changes arithmetic: teacher-forced over 4,995 positions against the same build unset, Q2_0 top-1 97.12%, mean KL 0.0094 (p99 0.119), perplexity 4.456 against 4.442; IQ3_S top-1 96.16%, mean KL 0.0160 (p99 0.211), perplexity 4.119 against 4.135. The half range is not guarded (see the PR). Off by default.
Ports the two RDNA4 pull requests onto this tree by hand (three-way merge of the rejected hunks). Where the contributor made a
default depend on a device query, the default is the gfx12 table or the env variable instead, so STRATA_GFX12_DEFAULTS=0 still
gives the old behaviour; the LDS-only barriers and the S2 and head changes that were only run on gfx1201 apply to gfx12 only,
every other target keeps what it ran before.
- iq_kernels.cu (3259): LDS-only barriers in the grid / activation staging, first-pass barrier skipped, vectorised ed_stage_entry,
expdec_gu_kernel launch bounds in waves per EU on HIP, EXPDEC down read at first use; expdec_bench builds on HIP.
- s2_expert_grouped.cu (3264): single-entry gu/down specialisation for the draft layer, fused swiglu + q8_0 on the hit paths,
LDS-only barrier on gfx12.
- verify_kernels.cu, quantize_act.cu, elementwise.cu, shared_expert.cu, fused_gr.cu, mtp.cpp, verify.cpp (3264): multi-block argmax
and top-prob on gfx12, warp-per-block quantize_q8_0_scaled, mapped copies in one launch on HIP, shared expert reuses the q8_1
image, counter-free q8_mixed in the hyper-connection read, the draft layer and the head write the q8_1 image themselves.
- parity tests of both pull requests (verify_parity on HIP, s2/shared/gr parity cases).
Ports three community changes that we have no card to run (no gfx1010, no gfx1030 here), on by default under the one
setting STRATA_AMD_COMMUNITY_TUNING (=0 turns the tables off; cmake -DSTRATA_AMD_COMMUNITY_TUNING=OFF compiles the portable
dot for gfx1010 again). Each is limited to the card or exact library build it was made for.
- gfx1010 (RX 5700 / 5700 XT) signed dot4 takes the SDWA byte-select sequence gfx1012 already used: PR 4855, author Keoian.
Same arithmetic, four signed byte products accumulated modulo 2^32.
- rocBLAS table gfx1030 / 5.4.0.ad84daa81c (TheRock gfx103X-all 7.13 wheels): PR 3343, author elightcap.
- rocBLAS table gfx1030 / 5.2.0.dabb6df2b98-dirty (Arch Linux ROCm 7.2.4): PR 3924, author schmittcn.
include/strata/core/amd_community.hpp holds the switch, the list of community tables and the engine log line (printed only
on gfx1010 and gfx1030); gemm.cu refuses a community table with the switch off, the existing loader still refuses any table
whose architecture or full rocBLAS version string differs. amd_community_test checks that the switch turns each one off.
The contributors' numbers are labelled as theirs in docs/AMD_HIP.md and have not been reproduced.
Ported from glorysdj's draft pull request #4185 (commits 961f62963 and 0726ef5ad of that branch; the documentation commit is not taken).
A compile option STRATA_SYCL_IQ_DECODE_FAST (default OFF: packed Q2 expansion, IQ4 bit-select lookup, fast IQ3_S sign arithmetic) and fourteen runtime controls, all default off:
STRATA_IQ3S_GU_SLM, STRATA_IQ4NL_DOWN_SLM, STRATA_RESIDENT_PLAN_COMPACT, STRATA_IQ2S_GU_ROW_REUSE, STRATA_IQ3XXS_GU_ROW_REUSE, STRATA_GR_CACHED_NORM, STRATA_GR_LO_REUSE, STRATA_MMVQ_SHAPE_RPW, STRATA_GDN_QK_REG, STRATA_EXPERT_ROW_E3, STRATA_Q6_SHAPE_GRF, STRATA_IQ3S_GU_ROW_REUSE, STRATA_GR_REDUCE_ILP, STRATA_ATTN_MERGE_ACTIVE.
Conflicts resolved against this tree: the kprof launch wrappers kept (new launches in the conflicted hunks use strata::kprof::kp), the resident-plan launch keeps the kprof wrapper with the compact thread count, native_expert_bench.cpp keeps this tree's fixed bench (the PR's older version of the same hunks is dropped), qsa_prompt_attn_parity.cpp keeps the per-cell check and adds the PR's FP32HASH line and QA_FP32_GRAPH timing.
Also in the diff: docs/INTEL.md section from the PR.
Port of pull request 4283 by bitaccesscomau (two commits 3956b94a7 and fcb0e399c on v0.2.8), applied without conflict on the 0.2.9 trial tree.
Off unless the variable is set; STRATA_PREFILL_ROUTED_STREAM_DEPTH sets the copies queued ahead (default 16).
Ported from maxfridbe's pull request #3248 (commits d95a9c4b0, 5fbef34f9, 344e43e1d, c44fb63be of that branch), cherry-picked without the prompt-kernel commits.
- conversation_snapshot: a copy with neither end in device memory (pinned host <-> pageable host) is a memcpy on the CPU; through the queue it runs on the Arc's copy engine, which the PR reports hanging (Engine reset: engine_class=bcs).
- s2_expert_grouped: #pragma clang fp contract(off) in the two transposed grouped kernels (fixups.py 9d keeps it across re-migrations; 9c the snapshot test's sycl::free).
- gr_parity: the STRATA_QFUSE read check is skipped when the SYCL fused GR read writes no q8_1.
- strata-sycl.sh: STRATA_CACHE_DIR defaults to the mounted data root so the int8 crossover cache survives the --rm container.
M5 Max, MTP, 5 interleaved pairs, fixed harness (each run's spec, min-p and context read from the engine's own log), greedy ids identical in every run, over --spec 2:
Q2_0 story 104.4 -> 104.7 (+0.4%), code 114.4 -> 127.3 (+11.5%), JSON 124.6 -> 132.5 (+6.3%), agent turn 115.9 -> 117.9 (+1.9%, 3/5 pairs).
IQ3_S story 94.2 -> 93.3 (-0.9%), code 103.8 -> 106.1 (+2.5%), JSON 105.1 -> 122.6 (+16.7%), agent turn 102.1 -> 107.1 (+4.9%).
The earlier Q2_0 'flat' result and IQ3_S '+1.4..3.6%' came from a harness that gave the spec flags to the first pair of an arm only. Other sets keep --spec 2.
The Qwen3.6 engine (src/qwen35/q36_engine.cu) started its penalty history from the whole prompt (`hist = ids`), so its model-card presence penalty 1.5 counted the prompt's last tokens as in the other engines. The history now starts at the last prompt token and grows with the tokens written (penalty_history_start in include/strata/qwen35/sampling.hpp), the same window as src/program/generate.cpp; batch slots copy that history. STRATA_PENALTY_PROMPT=1 keeps the whole prompt.
Tested: qwen35_sampling_test (276 checks, 9 new for the default and the opt-out), and a CUDA sm_86 build on z440. Replay on z440, one A4000, UD-Q6_K_XL, the reporter's conversation cut to 250 records, 400 tokens: 16 sampled seeds and 2 greedy requests per arm. This model did not return the old JSON in either arm (0/16 with the default, 0/16 with STRATA_PENALTY_PROMPT=1; 0/2 greedy in both), so this replay does not show a difference between the two windows; no answer in either arm started with the old JSON or a code fence.
docs/DETAILS.md: the sentence saying the Qwen3.6 path still counts the prompt is replaced.
M5 Max, MTP, IQ3_S, 5 interleaved pairs (2 + 3), greedy tokens identical in every run: --spec 3 --spec-min-p 0.5 against --spec 2:
story 93.8 vs 92.4 tok/s (+1.4%), code 103.4 vs 100.4 (+3.0%), JSON 105.3 vs 101.6 (+3.6%), agent turn 100.8 vs 98.0 (+2.8%).
Q2_0 measured flat (-1.0% to 0.0%, 5 pairs; 800-token story -0.7%), so it keeps --spec 2; --spec 3/4 without min-p lose on both.
mac_fit.spec_args picks the flags by model; engine_args takes the model name (setup.py and metal/setup_mac.py pass it); a streaming
Mac keeps STREAM_SPEC, a Mac without the draft layer keeps --spec 2. The engine's own default (--spec 2) and an explicit --spec in a
config are unchanged. Test: tools/test_setup_mac.py.
With the model-card presence penalty 1.5 over the last 64 tokens, the prompt's own tokens (the digits of a new task, a quoted JSON answer) were penalised, and a conversation whose last turn asks for something new could get its previous answer back. The window now starts at the end of the prompt: the fed-back last prompt token and the reply's tokens. Earlier assistant turns inside the prompt are not counted. The history is built from the request's prompt ids on every request, so a conversation restored from the conversation cache or a saved slot gets the same window as a cold read. The STRATA_ADMIT_MERGE_FIRST merged first window is not used for a request with a penalty window, since its prompt rows would join the reply window.
Measured with the opt-in switch (618cd7a4b) on a P100, IQ2_XS, the reporter's saved 4-message conversation, 32 seeds per arm: old JSON returned 13/32 as shipped, 4/32 reply-only, 3/32 with presence 0. Qwen IQ3_S on an R9700, 16 seeds: old JSON 8/16 as shipped, 0/16 reply-only. 61-task quality027, greedy, one run per arm: 56 and 55 passed as shipped (two runs), 55 reply-only, 54 with presence 0.
Added penalty_window_test (header-only, CPU) and a line in docs/DETAILS.md.
On Windows the HIP runtime gives every cudaHostRegister'ed range a second address for the GPU and takes any host pointer inside it for the registered memory, but it does not reserve that range in the process's address space. In about one start in ten the stack of the thread that pins the expert arena lay there: small cudaMemcpy calls from local variables then read (and wrote) the arena, and every long prompt was read wrong until a restart.
- pinned.cu: each registered range's alias is reserved (MEM_RESERVE); a range whose alias already covers memory in use is unregistered and treated as refused by the driver, so it stays resident but unpinned (logged).
- generate.cpp: main keeps the free address space below 1 TiB reserved for the life of the process and runs the engine on a new thread, so its stack and later allocations lie above the aliases (STRATA_STACK_ISOLATE=0 turns it off).
- pinned_alias_test, docs/AMD_HIP.md.
Windows HIP only; other backends keep their behaviour. Author's measurements: 0 bad starts in 110 (main + change 0 in 70, main alone 5 in 70) on an RX 7900 XTX. Here: the HIP objects, pinned_alias_test and pin_budget_test build for gfx1100 with ROCm 10.2 (no gfx1100 card to run them on). The hunk in pinned.hpp was applied by hand because v0.2.9 has the shared-file fields between the lines the PR touched.
Claude Code declares ToolSearch and a defer_loading placeholder and names its MCP tools only in the conversation; when the model
calls one of them straight away the call came back as text and the session stopped. Ported from PR #4700 by xrip onto v0.2.9
(the server.py hunk for the parser call moved); the PR's own tests are included.
Co-authored-by: Niko1221 <[email protected]>
The download default stays on the four-shard pin cb02422c; nothing is fetched under the new names. The six-shard files (Hub commit a419d74c,
27 October: five shards a set, the Engram as the sixth) and the four-shard files with the first shard ISTA re-published on 26 October (what the
Hub's last four-shard commit 653a9874 serves; #4237 was refused by the app because only the pinned size of shard 1 was known) are taken from
files already here, by size and SHA-256.
- setup.py: DSV41_ISTA6_FILES (15 files), shards6 on the three sets, dsv41_layout_of / dsv41_parts / dsv41_pack_layout; install_dsv41 picks the
layout of a --gguf-dir folder (both whole: the pinned one), packs the five files with the Engram as --engram, and builds the pack again when it
was made from the other layout's files. The converted Engram (same name and size, another hash) is accepted beside the pinned one.
- strata_control/dsv41_sets.py, models.py: the app's check and `strata models import` know the six-shard names, the re-published shard 1, and
that a folder with both whole layouts of one set is that set once.
- strata_cli/launch: --model-path and the guided screen name what is missing in the layout that is there.
- tools/dsv41/pack_gguf.py: --engram as the sixth shard of a six-shard split; the pinned release's Engram serves the six-shard shards too.
- Tests for each; docs/DEEPSEEK41.md.
Builds on the cherry-picks of #4348 (integer tensors in the GGUF reader) and #4632 (StackArc3's six-shard work, option A), which this
commit reworks so the four-shard pin and files keep working unchanged.
Related: #4237, #4256, #4632.
On 27 October 2026 ISTA-DASLab replaced the four-shard IST-DAS builds on
the Hub's main with six-shard builds (llama.cpp b11870+ layout), and
--model-path with those files failed with "...-00001-of-00004.gguf is
missing". The v0.2.7.1 pin still serves the old files.
- Pin a419d74c: five shards per build named of-00006 and the shared
Engram as the sixth. The Engram was converted in place (same name and
size, new hash).
- pack_gguf.py counts an --engram file as the sixth shard only if it is
one file, not already among the shards, number count-1 of the same
split, and holds the single tensor per_layer_token_embd.weight.
- An Engram whose finish mark names an old hash is fetched again in the
same run. One whose mark has no hash is hashed: a match is kept and
marked, a mismatch is fetched again. A user's own old one is refused
with the reason.
- A pack from an earlier run is rebuilt unless it names this run's files.
- Six-shard only; four-shard files are ignored and stay on disk. The
first-release name tables are now empty. No engine code changes.
Related: #4237, #4256.
Niko1221