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).
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.
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.
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 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.
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.
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.
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.
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).
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).
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).
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).
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.
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).
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.
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.
/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]>
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.
--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.
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.
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 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.
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.
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.
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.
Niko1221