release 15 - the board stops invalidating daily; the M5/M4 re-mint on one pin, serve cards, and the news - #3909
Merged
Merged
Conversation
…ASLLAMA_RELEASE The one constant conflated three invalidation scopes (tune sidecars, board/records provenance, exchange validity) and was bumped on any kernel change, so every bump voided the whole board and every box's sidecar. It is now DASLLAMA_RELEASE, bumped only on a declared release - a maintainer ruling that bench comparability is broken. Per-change invalidation stays with the finer mechanisms: IMAGE_VERSION and the reflected layout fingerprint (dlims never read this constant - verified). Value stays 15, so every v15 artifact remains valid; wire names (dasllama_version in exchange JSON, sidecar provenance via the new version_key= override on [tune_scope], bench records) are schema-stable. REVIEW.md's kernel-work bump duty is replaced by the declared-release rule; the per-entry kernel-hash hardening is ledgered as followup_general 58. Also: CLAUDE.md's lint-opportunity routing now names REVIEW.das gates for folder-owned patterns; the release_remint_plan.md arc plan lands with the step-0 gate-candidate assessment. Gates: model-free suite 74 files green; compile checks on the renamed roots. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc
…he msl fail-closed sync cell [metal_dispatch] on a class with @Workgroup members now refuses at compile when tgmem= is absent - the generated builder would bind ZERO threadgroup memory and the kernel reads whatever length the encoder last had (the qk_rope bug class; it shipped working only via sticky encoder state). All 93 @Workgroup classes at HEAD already pass tgmem=, so this is pure prevention. Negative+positive controls ride test_metal_misc_kernels as spawned compiles (the tgmem-less fixture must refuse with the message; the tgmem= twin must compile). modules/dasMetal/REVIEW.das grows check_fail_closed_sync (followup 45): the tests/msl/_fail_closed fixture list and test_msl_fail_closed.das's check_rejects asserts must match both directions - either drift is a finding (negative-controlled both ways). REVIEW.md carries the weakening-residue rule. Gates: kernels suite 7/7 files green (misc 4/4 incl. the new lens gate); modules/dasMetal/REVIEW.das OK. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc
Slice A (kernel coverage): cov_tower grows the gemma4a whole-chunk row (GPU mel + convs + 12 Conformer blocks off raw samples), the canary FastConformer row (32 rel-pos XL blocks), the qwen25v single-window row (win-attn + row-gather), the qwen3v GPU-stem row, the qwen3a mel now built from samples so the GPU mel hook dispatches, the gemma3v flash-off leg (head_restride + the slab route rode only sticky luck before) and the crowned twin-W leg (cvt_half + the half GEMMs). Every target family verified dispatching in the census report. Slice B (board plumbing): AsrModelSpec grows metal_served - the whisper pair, canary, qwen3a, gemma4a and omni declare the end-to-end Metal serve on their catalog rows; is_metal_served_family reads the flag (parakeet stays CPU by measurement). The like-for-like GPU ref pairs extend beyond whisper-cli: mtmd gets its -ngl 99 arm, nemo a torch-MPS --device arm (canary_qwen_bench.py). The stale m1/m4 canary rows pinning the retired f16 decoder are removed (m1 retired by ruling - no more m1 runs; m4 re-mints on the q8 artifact in the Efficiency-gate pass); site records regenerated. Followup 45 stays resolved; followup 57 verified no-work (Plane length overloads already return int64) and re-ledgered as such. Gates: census suite green with all new rows dispatching; records/site/specs/ exchange schema tests 72/72. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc
The 2026-08-30 entry: M5 as a first-class target - tensor-API kernels under the per-box crowns, GPU vision towers and speech encoders, the hybrid all-compute CPU pool - with the first official M5 Max board rows quoted from records/m5.json. Regenerated index.html/feed.xml/sitemap.xml also catch the old entries up to their ASCII-normalized _news sources (deploy output already matched). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc (cherry picked from commit 20ac6e6d3399ddc1da0d264028beaa2a6bcbfb24)
…d sibling setup_asr_rig builds a build-metal llama-mtmd-cli beside build-cpu on Apple boxes (the CPU-only build's -ngl is inert - a gpu ref leg through it would stamp a CPU run under a metal label), profile_common resolves it (mtmd_bin_metal, empty = the gpu leg skips loudly), and setup_lcpp_ref builds llama-mtmd-cli beside llama-bench so the sibling rule resolves the image refs on a fresh box. Slice C outcome + slice D predictions logged in the plan. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc
…tier boot-restore A serialized exe restores globals as DATA, so function values arrive null unless a boot-restore [init] heals them (dasllama_config's restore_dlim_source_defaults states the law; dasllama_gpu_tier HAS such an [init]). The vulkan decode-mirror/ffn-tail arc added 12 declaration-defaulted hook globals that never joined it, and put one - g_moe_gpu_attn_dec_claim, via prefill_claim_mirrors - on the UNCONDITIONAL prefill tail. Every standalone-exe forward died with 'invoke null function': all 10 board text models refused at the parity pregate, the four decoder-gguf ASR families' cpu and accel cells crashed, image cells never ran. No -jit gate could see it - script mode runs the declaration initializers - which is why months of green suites proved nothing about the exe rail. Found by git-bisect (first bad 4002388) + a staged exe probe; the 12 join the restore. test_exe_smoke joins the model-free suite: builds _exe_smoke_root.das with -jit -exe and runs the artifact through a 4-token generate (negative-controlled: red on the crash without the fix, green with it). SmolLM2 small tier; skips honestly without it. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc
The full re-mint on the b10659 pin: 10 text models (metal pp 1.07-1.23x / tg to 1.33x ahead; the CPU board debuts vs clean-cpu with the MoE 32-row split at pp 1.60-1.68x, accel to 1.82x), the image board with honest metal ref pairs at last (the patched sibling), the full ASR board with GPU ref twins (das metal ahead of whisper-cli-metal, mtmd-metal and nemo-MPS everywhere but the ledgered canary gb1 CPU front), and the owed board cells: canary q8 decoder + Metal FastConformer, gemma4a/qwen3a metal rails, and the qwen vision carriers - Qwen3VL-4B and Qwen2.5-Omni-3B verified 40/40 against the live oracle, flipped official, their image cells minted (the qwen25v tower reproduces the arc: enc 106.0 vs 118.0 ms). Red 1, gpt-oss pp halved (3989 -> 2047): the followup-53 scaffold migration renumbered the mx4 tensor twin's x/y binds to the family contract (xf@3 y@4) and the HAND-BOUND race arm kept 2/3 - the twin read the zeroed output buffer as X, the race honestly measured garbage (maxdiff == maxref), refused the crown, and serving fell to the simdgroup family. Race arm fixed, both manifests re-minted (moe_mulmm_mx4 tensor-crowned again), cell respotted: pp 3995.1 - the pre-merge number. Third incident of the hand-bound-race class; the census opaque-dispatch gate is now the priority lint candidate. Red 2, the canary nemo ref: NeMo's generate mutates the prompt list in place (warm reps degenerate to a 3-token stub) and torch.set_num_threads beside an MPS device degenerates EVERY rep; the crashed cell then laundered a stale onnx row through the rowless-pipe --out fallback, which never checked the tool. Bench rebuilds prompts per rep, discards degenerate reps loudly, keeps the threads knob CPU-only; the fallback is onnx-only. Honest rows: das metal 2.29x/2.02x ahead on jfk/jfk3, 0.68x on gb1 (the ledgered CPU mel+subsample front). Red 3 (onnx whisper 1/5 clips) closed by-design (single-window adapter, documented SKIP). The two dasllama_math accel-override globals join a boot-restore [init] (REVIEW.das gate green). Site records regenerated; records/site/specs gates green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc
…y gate proven live The retired f16-decoder rows' replacement: canary requanted ON THE BOX (llama-quantize Q8_0 per the recipe; a fresh build-quant dir - the old ref build's SDK path had gone stale), the m4 sidecar re-minted by the release, and the cells landed with a WORKING nemo torch-MPS twin: das metal 2.10x/2.19x ahead on jfk/jfk3, 0.79x on gb1 (the same CPU mel+subsample front residue as the M5 row - one ledgered class, two boxes). The tier-kind gate held live: M4's pool stays 10 lanes (the allowlist refuses the 4 E-cores; the live: real-lanes witness, not the regime intent line). One transient: the sweep-context nemo ref degenerated every rep right after three das children exited (the async-reclaim window) and the new DEGENERATE guard refused to record it - a quiet-box rerun passed 6/6 reps; the guard turned what used to be silent garbage timing into a loud retry. Site records regenerated; records + site schema gates green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc
…arries the final board Qwen3-VL-4B and Qwen2.5-Omni-3B get serve rows (commit-pinned urls per the serving rule - resolve/main floats), each naming its mmproj companion, so dasllama-server's /catalog offers the deepstack tower and the window-ViT tower alongside the gemma line. The dasllama.io entry is rewritten on the re-minted board: every figure now quotes a committed record row (E2B 9074/161 vs 7413/137, Qwen3.8-27B 914/27.8 vs 752/25.4, 30B-A3B 3753/166 vs 3422/137, gpt-oss 3995/163 vs 3477/149; 12B image encode 7.1 vs 95 ms, the qwen25v window ViT 106 vs 118, whisper turbo jfk 140 vs 153), with the M4 refresh and the M1 record named. Gates: model-specs invariants 12/12, openai-server suite 21/21 (the catalog is a view over the spec table, so the serve rows are gated there), site metadata 4/4. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc
…rs section
The dasLLAMA harvester's ledger, ruled and landed: the recovery-file rule
(performance/REVIEW.md - a fallback names the tool that writes its file and reads it
for no other); ARCHITECTURE_MEASUREMENT.md sec.2.20 {#asr-gpu-pairs} (the metal_served
opt-in + anti-sandbag, the three GPU ref arms and the two pairless legs, the sibling
build + timing patch + Metal-ON twin) cited from build_flavor, gen_bench_records main
and mtmd_bin_metal; the scaffold-vs-family-tail binding-number fact appended to
ARCHITECTURE_GPU_PREFILL.md sec.2.2g and cited from race_moe_mulmm_mx4; comment
compressions and drops across the diff per the ledger; the two stale //! docs
(run_asr_mtmd gpu arm, asr_pybench_args device) extended; the mtmd_bin comment
restored onto its function; test_exe_smoke's census entry in tests/CLAUDE.md.
dasMetal's harvest: the gate-cell header drops (law lives in REVIEW.md + finding
strings), ARCHITECTURE.md gains Gate C in the test-gate inventory, and tests/msl's
needle rule shrinks to the machine-checked residue form.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc
…ry census mints The per-checklist audit fan-out (14 auditors: 11 walk checklists + the MEASUREMENT/GPU/TOWER routed companions) over the release-remint branch, every accepted finding applied: Records: the four orphaned pre-rebase commit stamps re-stamped to byte-identical reachable twins (rows' sha, three tune archives re-hashed + renamed with every tune_sha repointed, LKG + the in-tree bench sidecar); gen_site_records --verify OK, all schema gates green. Serving catalog: the two new serve rows' mmproj companion urls commit-pinned (the catalog gate was red on the floating /main/ refs); catalog_idle.json re-captured from a live boot (7 cards), playwright 106/106 after teaching two specs the vision-present behaviors. Tests: the census tower rows gain family_on tags + lane pins and mint IN MEMORY via stage_*+mint_* (a pinned-lane disk bake GC-purges the serving lane's .dlim - it broke test_audio_embedder; new mint_gemma4a_encoder fills the one missing wrapper); the strict whisper/gemma4uv cells pin the simdgroup lane through the new pin_metal_tensor_crowns latch (the re-minted sidecar armed crowns in suites for the first time; a bare setter cannot hold a lane - every load re-applies profile crowns; the trio's crown flags cache at prefill init); the role fixtures gain their tgmem= (the new lens refusal fired ahead of the role error); the exe smoke reads env through env_value_of and spawns Windows-safe. Docs: news entry drops a false cross-box superiority claim and carries record provenance; the checklist self-review repairs land across tests/benchmarks/performance/site/msl/server/ ladder/GPU/TOWER/MEASUREMENT (six exchange rules move to the module checklist, gate-enforced rules collapse to weakening residue, enumerations become properties); followup 60 ledgers the crowned-lane strict-parity question. All three owed suites green on the final tree: model-free, kernels, coverage. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc
… control - test_asr_catalog_invariants pins the metal_served roster per row (the flag decides which rows get GPU board cells and which reference legs pair GPU-vs-GPU; flipping canary reds) - test_model_catalog pins the four curated card names (the count floor alone let a whole card vanish green; renaming qwen3-vl-4b reds) - test_exchange_schema pins the tune scope's version_of/version_key text (the rename's load-bearing compensator - llvm_tune exposes no runtime reader of the resolved key, so the pin is textual; blanking version_key reds) Woodpecker round 2 at the prior tip: zero findings. TDD audit: 19 controlled branches, 12/12 test edits justified; the remaining measurement-orchestrator branches ride the standing coverage-followup PR. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc
…, two off-Apple lint nits - LINT027: the six exchange rules moved into modules/dasLLAMA/REVIEW.md pushed it to 317 lines - they now live in the routed REVIEW_EXCHANGE.md companion (the parent routes by trigger), 294 lines left - LINT012 (linux mirror): the two sample-fed census rows annotate [unused_argument(t, samples)] - off-Apple the gated bodies are empty - tests-interp / tests-jit reds were chained-load transients (three files erroring on builtins that resolve solo and on targeted lane reruns - both lanes PASS clean) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc
- the Canary sentence now carries the gb1 pair (das 6484 ms beside the reference's 4383 on the M5) - the long-audio gap stands on the page next to the wins, and is named the next dig - site/REVIEW.md's first receipt rule no longer contradicts its sibling: a rendered cmd identifies the RUN, and one run's cmd covers the several rows it produced (the ASR reference legs' per-clip rows) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc
Contributor
There was a problem hiding this comment.
Pull request overview
This PR implements dasLLAMA “release 15” rollout mechanics by decoupling the tune/sidecar/board staleness counter from routine kernel changes (DASLLAMA_VERSION → DASLLAMA_RELEASE), re-minting/re-stamping measurement artifacts accordingly, and extending the surrounding gates, tests, serving catalog, and site/news outputs to match the new release semantics.
Changes:
- Rename and re-scope the engine versioning contract to
DASLLAMA_RELEASEwhile keeping the wire/schema keydasllama_versionstable. - Add/extend review gates and tests for (a) exchange + staleness pinning, (b) standalone
-exeserialized-globals hazards, (c) Metal lenstgmem=enforcement, and (d) expanded tower/ASR coverage and catalog invariants. - Refresh serving catalog fixtures/specs and regenerate dasllama.io news/feed/sitemap content for the M5 announcement and updated provenance rules.
Reviewed changes
Copilot reviewed 79 out of 83 changed files in this pull request and generated 1 comment.
Show a summary per file
| File | Description |
|---|---|
| utils/internal/dasllama-ladder/REVIEW.md | Update ladder checklist wording (record-store validation). |
| utils/internal/dasllama-ladder/README.md | Align ladder README with wire stamp terminology. |
| utils/dasllama-server/tests/towers.spec.js | Adjust tower-offer test to ignore capture-box stock state. |
| utils/dasllama-server/tests/fixtures/catalog_idle.json | Refresh captured catalog fixture content. |
| utils/dasllama-server/tests/catalog.spec.js | Expect image_mmproj in config when vision tower present. |
| utils/dasllama-server/test_model_catalog.das | Add roster presence pins for key serve cards. |
| utils/dasllama-server/test_exchange_client.das | Switch fixtures/tests to DASLLAMA_RELEASE while keeping wire key. |
| utils/dasllama-server/REVIEW.md | Tighten fixture/Playwright capture rules wording. |
| utils/dasllama-server/README.md | Document exchange apply gate using DASLLAMA_RELEASE. |
| tests/msl/REVIEW.md | Rephrase fail-closed sync rule to point at gate check. |
| site/REVIEW.md | Clarify receipt rule for shared-run cmd provenance. |
| site/LAWS.md | Record ruling about receipt cmd identifying the run. |
| site-dasllama/sitemap.xml | Update lastmod for deployment date. |
| site-dasllama/REVIEW.md | Clarify “page” scope + comparative-claim and provenance rules. |
| site-dasllama/index.html | Regenerate news section including M5 entry and punctuation tweaks. |
| site-dasllama/feed.xml | Regenerate Atom feed with M5 entry. |
| site-dasllama/build_news.py | Add LF-only writer helper for Python 3.9 compatibility. |
| site-dasllama/_news/2026-08-30-apple-m5.md | Add M5 news source markdown entry. |
| plans/dasllama_io_site.md | Mark old versioning decision as superseded by release split. |
| modules/dasMetal/REVIEW.md | Update wording around fail-closed sync enforcement. |
| modules/dasMetal/REVIEW.das | Add fail-closed fixture/assert sync gate. |
| modules/dasMetal/ARCHITECTURE.md | Document new Gate C for fail-closed sync. |
| modules/dasLLAMA/tests/test_whisper.das | Pin/restore tensor crowns and reset Metal cached flags in a strict cell. |
| modules/dasLLAMA/tests/test_metal_misc_kernels.das | Add spawned-compile test for lens tgmem gate. |
| modules/dasLLAMA/tests/test_kernel_coverage.das | Expand tower coverage rows; use stage+mint to avoid .dlim side effects. |
| modules/dasLLAMA/tests/test_gemma4uv.das | Pin/restore tensor crowns for rung calibration stability. |
| modules/dasLLAMA/tests/test_exe_smoke.das | Add end-to-end -jit -exe smoke test for serialized-globals rail. |
| modules/dasLLAMA/tests/test_exchange_schema.das | Add textual pin test for tune scope version_key stability. |
| modules/dasLLAMA/tests/test_dasllama_lint_escape.das | Update lint-escape test to reference DASLLAMA_RELEASE. |
| modules/dasLLAMA/tests/test_bench_records_schema.das | Use DASLLAMA_RELEASE; pin metal_served roster invariants. |
| modules/dasLLAMA/tests/run.das | Add new exe-smoke test to suite list. |
| modules/dasLLAMA/tests/REVIEW.md | Clarify rules around image-rail loaders, knobs, and stage+mint under pins. |
| modules/dasLLAMA/tests/CLAUDE.md | Document new tests and coverage-family token additions. |
| modules/dasLLAMA/tests/_role_fixtures/_rf_deva_wrong.das | Add tgmem= to fixture dispatch annotation. |
| modules/dasLLAMA/tests/_role_fixtures/_rf_deva_read.das | Add tgmem= to fixture dispatch annotation. |
| modules/dasLLAMA/tests/_exe_smoke_root.das | Add exe smoke root program used by the test. |
| modules/dasLLAMA/REVIEW.md | Update versioning rule to declared-release semantics; route exchange checklist. |
| modules/dasLLAMA/REVIEW.das | Add gate to enforce boot-restore guards for function-typed globals. |
| modules/dasLLAMA/REVIEW_TOWER.md | Broaden routed scope and clarify gate obligations around lens emission. |
| modules/dasLLAMA/REVIEW_MEASUREMENT.md | Refine figure/provenance rules and scope boundaries. |
| modules/dasLLAMA/REVIEW_GPU.md | Update GPU checklist to reflect lens tgmem refusal + binding-arm hazards. |
| modules/dasLLAMA/REVIEW_EXCHANGE.md | Add dedicated exchange checklist companion. |
| modules/dasLLAMA/release_remint_plan.md | Add arc plan and evidence ledger for the release/remint work. |
| modules/dasLLAMA/performance/setup_asr_rig.das | Build Metal-ON mtmd reference sibling on Apple boxes. |
| modules/dasLLAMA/performance/REVIEW.md | Adjust record hygiene rules; add catalog test coupling rule. |
| modules/dasLLAMA/performance/records/m1.json | Remove retired M1 canary rows tied to deleted artifact. |
| modules/dasLLAMA/performance/profile_common.das | Add metal_served flag; add mtmd Metal sibling support; refine pybench args. |
| modules/dasLLAMA/performance/model_specs.das | Add serve cards and commit-pinned URLs for new models/mmprojs. |
| modules/dasLLAMA/performance/gen_bench_records.das | Track measured-cell count; switch ASR GPU pairing logic; harden pybench recovery behavior. |
| modules/dasLLAMA/PERF_LEDGER.md | Add external intel note entry with attribution context. |
| modules/dasLLAMA/LAWS.md | Record maintainers’ rulings (release split, exe restore gate, measurement scoping). |
| modules/dasLLAMA/harness/tuner_noise.das | Stamp release number under wire key dasllama_version. |
| modules/dasLLAMA/followup_general.md | Update/resolve followups related to release scope and fail-closed sync. |
| modules/dasLLAMA/dasllama/dasllama_version.das | Introduce DASLLAMA_RELEASE constant and updated documentation. |
| modules/dasLLAMA/dasllama/dasllama_metal_tower.das | Add getters for tower f16/flash flags. |
| modules/dasLLAMA/dasllama/dasllama_metal_prefill.das | Fix mx4 hand-bound race binding indices; add architecture citation. |
| modules/dasLLAMA/dasllama/dasllama_metal_lens.das | Add compile refusal when @workgroup members exist without tgmem=. |
| modules/dasLLAMA/dasllama/dasllama_math.das | Add boot-restore [init] guards for function-typed globals. |
| modules/dasLLAMA/dasllama/dasllama_math_gen.das | Switch tune scope to DASLLAMA_RELEASE and pin wire version_key. |
| modules/dasLLAMA/dasllama/dasllama_gpu_tier.das | Add additional boot-restore guards for function-typed globals. |
| modules/dasLLAMA/dasllama/dasllama_gemma4a.das | Add stage+mint encoder helper to avoid disk image writes in suites. |
| modules/dasLLAMA/dasllama/dasllama_exchange.das | Update exchange gating/logging to DASLLAMA_RELEASE. |
| modules/dasLLAMA/dasllama/dasllama_common.das | Add crown pin latch + getter; prevent box-profile crowns overriding pins. |
| modules/dasLLAMA/BRINGUP.md | Update bringup doc for declared-release staleness semantics. |
| modules/dasLLAMA/benchmarks/setup_lcpp_ref.das | Ensure mtmd timing patch is applied; build mtmd target as sibling. |
| modules/dasLLAMA/benchmarks/REVIEW.md | Consolidate race-proof rules; clarify scratch recovery-file constraints. |
| modules/dasLLAMA/benchmarks/lcpp_bench.das | Stamp bench rows with DASLLAMA_RELEASE. |
| modules/dasLLAMA/benchmarks/asr/canary_qwen_bench.py | Add device selection; fix prompt mutation and degenerate-rep handling. |
| modules/dasLLAMA/benchmarks/asr/_pybench_rows.txt | Remove tracked recovery file (now gitignored scratch). |
| modules/dasLLAMA/ARCHITECTURE.md | Update measurement doc index reference. |
| modules/dasLLAMA/ARCHITECTURE_MEASUREMENT.md | Document ASR GPU pairing policy and tooling requirements. |
| modules/dasLLAMA/ARCHITECTURE_GPU_PREFILL.md | Document binding-number divergence hazard for mx4 scaffold vs tail. |
| modules/dasLLAMA/ARCHITECTURE_ENGINE.md | Update version constant semantics; describe exe global-restore gate. |
| LAWS.md | Record repo-level ruling about REVIEW.das as a lint/gate landing place. |
| CLAUDE.md | Update guidance on where to land “lint opportunities” (REVIEW.das vs daslib lint). |
| .gitignore | Ignore ASR pybench recovery file path. |
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc
Contributor
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 79 out of 83 changed files in this pull request and generated no new comments.
Suppressed comments (2)
Previously missed (2) — in code that hasn't changed since the last review.
utils/dasllama-server/tests/towers.spec.js:24
- withVisionAbsent assumes the catalog fixture always contains the 'gemma-4-e2b' entry; if the fixture is re-captured without it (or the id changes), this helper throws a vague
Cannot set properties of undefinedinstead of failing with a clear reason. Add an explicit guard with a descriptive error so fixture drift is diagnosed immediately.
utils/dasllama-server/test_model_catalog.das:55 - The presence-pin loop repeatedly calls model_catalog() inside the inner loop, re-building the catalog array for each wanted id. Cache the catalog once so the test is cheaper and asserts against a single snapshot of the roster.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The release split
DASLLAMA_VERSIONbecomesDASLLAMA_RELEASE(value stays 15). The bump duty flips from "any kernel work bumps it" to "bumped only on a declared release" - a maintainer ruling that bench comparability is broken, recorded inLAWS.md- so tune sidecars, board rows, and the exchange stay valid across routine kernel work. Wire and schema field names keep thedasllama_versionspelling (the release number now rides them; no schema break), pinned byversion_key=on the tune scope with a textual gate. Per-change invalidation stays with the finer mechanisms (IMAGE_VERSION, the reflected layout fingerprint); per-kernel identity hashes are ledgered as followup 58.The arc-end board re-mint (ref pin b10659)
Fresh tune manifests minted under the release scope on a quiet box (37/40 winners identical to the prior generation; LKG refreshed), then the full M5 board: text CPU/accel/Metal, image, audio - every cell behind the per-row parity pregate. The M4 canary rows re-mint on the new Q8_0 serving artifact, and the Efficiency-tier gate is proven live on both boxes: M5
pool 18 lanesspanning Super+Performance, M4live: 10 dispatch laneswith the efficiency cores excluded. M1 rows stand as the long-term record; its canary rows retire with the deleted f16 decoder artifact (the box runs no more sweeps by ruling).Sweep reds, all closed under the two-tier rule (a FAIL stops the sweep; a perf red is fixed after and only the red cells re-profiled):
xf@3, y@4, so the arm read its own output as X and un-crowned the correct twin. Rebound, re-raced, re-profiled with the fix in the built exe - the committed 3995.1 cell is the post-fix reading (the fix and the record land in one commit, which is why the row's stamp predates the commit carrying both).set_num_threadson MPS, and a shared recovery file laundered another tool's rows. All three fixed; a DEGENERATE rep guard discards sub-8-id reps.Figure provenance: all figures in this PR:
performance/records/m5.json/m4.json, minted 2026-08-30 (the m4 canary rows stamp 2026-08-31 - the box clock runs UTC+3) on the b10659 ref pin;img:enc/asr:figures arelcpp_bench --image/--asrboard cells against the reference cli/torch legs.Commit-stamp re-stamp (the records reachability rule's second clause): four pre-rebase stamps re-stamped to byte-identical reachable twins (
git diff --stat old new -- modules/dasLLAMA/= 0 lines each): fa67fe165 to 345dcd6, 82e45e0a3 to bbdcf8d, 366669a53 to f5417ba, 54595cfea to 77dfc24. The three content-addressed tune archives re-hashed and renamed, everytune_sharepointed;gen_site_records --verifyOK.Serving catalog + site
Two serve cards join the roster: Qwen3-VL-4B and Qwen2.5-Omni-3B, models and mmproj companions commit-pinned (each pinned url resolved live at pin time), gated by the catalog suite (16/16) and the control-page playwright suite (106/106,
catalog_idlefixture re-captured from a live boot). The dasllama.io news entry announces M5 support with figures quoted from the committed records - including the one loss (canary gb1, 6484 ms beside the reference's 4383), which is the named next dig.Coverage owed from the m5-image arc
Census tower rows for the arc's kernel families: qwen25v window ViT, qwen3v GPU stem, gemma4a Conformer whole-chunk, canary FastConformer, the gemma3v flash-off slab leg and the crowned twin-W leg - family-tagged, lane-pinned, and minting IN MEMORY via
stage_*+mint_*(a pinned-lane disk bake GC-purges the serving lane's.dlim; the rule is now intests/REVIEW.md). The census's assert-free dispatch-diversity design and its numeric fixture claims are ledgered intests/CLAUDE.md- yes, ledgered; likewise the mx4 twin's binding divergence inARCHITECTURE_GPU_PREFILL.md2.2g.The standalone-exe rail gets its gate: a serialized exe restores globals as data, so 14 function-typed globals gain boot-restore
[init]guards,check_exe_fn_global_restorewalks the folder per-diff (reverse control on base sources reports exactly the 14), andtest_exe_smokebuilds and runs a-jit -exeartifact end-to-end. The smoke's generate reaches the decode-claim trio; the rdec/ffn_xf seams are covered statically by the gate, and the Vulkan-armed parity proof is deferred to a Vulkan box (the hooks arrived from the parallel vulkan arc). The lens now refuses a[metal_dispatch]class with@workgroupmembers and notgmem=at compile, with a spawned-compile must-red/must-compile fixture pair.The review round
14 per-checklist auditors (11 walk checklists + the MEASUREMENT/GPU/TOWER routed companions), a worktree tdd audit, and two external reviewer rounds (both zero findings). Real catches applied: the floating mmproj urls (a red catalog gate), the news entry's unsupported cross-box claim, the crowned-lane test drift (the re-minted sidecar armed tensor crowns in suites for the first time - strict token-exact/rung cells now pin the simdgroup lane via the new
pin_metal_tensor_crownslatch; followup 60 holds the open question), and the census dlim purge. The tdd audit's three cheap named gaps closed in-PR with must-red controls (themetal_servedroster pin, serve-card presence pins, theversion_keytextual pin); the measurement-orchestrator branches ride the standing coverage-followup PR.Suites green on the final tree: model-free (75 files), kernels, coverage, exchange/records/site schema gates, catalog + playwright.
preflight --fullran once; its three reds were fixed and validated on targeted lanes (LINT027 splitREVIEW_EXCHANGE.mdout of the module checklist; two off-Apple LINT012 annotations; the interp/jit reds were chained-load transients - the same three files pass solo and both lanes pass clean on targeted rerun).🤖 Generated with Claude Code
https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc