dasLLAMA: M5 pp deep-dense - board v8 with every metal cell at or ahead

[borisbat]

DASLLAMA_VERSION bumps to 11 - the sidecar exchange stops serving v10 sidecars, so each box re-tunes its new kernels on the next run.

Board v7 left five M5 pp512 cells behind upstream, all deep-dense or MoE models. Three mechanisms, each fixed here. The dev-W tile picker capped N-column tiles at 8, so a 300 MB+ FFN panel could never tile under the cache knee and the whole FFN fell to the staged fallback; the cap is now 32. The q5_1 MoE weight format - forced on experts whose FFN width does not divide 256 - had no tensor twin at all; it gets the full T/TH/TH128/THR family. And a K-quant attention site whose f16 panel would be DRAM-resident loses by materializing it; new tall 128-row stamps dequantize their own W tile in threadgroup memory, gated on panel bytes (96 MiB).

Board v8, re-minted at ref pin b10659: every M5 metal cell is at or ahead of upstream in both columns for the first time - pp 1.006-1.202, tg 1.003-1.332. The five former reds close at Mistral-24B 1.117, gemma-4-12B 1.138, gemma-4-26B 1.051, Qwen3.8-27B 1.202, Qwen3.6-27B 1.185.

The prefill driver was also harvested into the document system: kernel-form facts now live in ARCHITECTURE_GPU_PREFILL.md behind [arch] anchors, the driver's bans landed in REVIEW_GPU.md, and REVIEW_COMMON.md gains the [arch] triangle audit rules (a diff touching a cited function, a citation, or an anchored section audits the rest of that anchor's triangle). REVIEW.md's measurement family split into REVIEW_MEASUREMENT.md. Two new mechanical checks ride along: make-pr's stamp-reach gate (a rebase orphans record-store engine stamps; the gate catches it where it appears) and a records/ corpus sweep in performance/REVIEW.das plus a dl-* selector-parity gate in the new site/REVIEW.das. The external review round found the dense tall-kq gate ignoring the set_metal_prefill_tall A/B seat - fixed as a pure predicate (tallkq_pick) with a device-free test binding the override clause.

Where to look: modules/dasLLAMA/dasllama/dasllama_metal_prefill.das (the pick ladders and the new kernel templates), ARCHITECTURE_GPU_PREFILL.md sec.2.2c/2.2d for the form ladder and knee map.

Validation, claims, ledger

Validation

• Board v8 is a full metal-leg gen_bench_records.das sweep at ref pin b10659; the sweep re-released the rig exe first (daspkg release, its own gate refuses a stale exe), warm-cache r5 protocol, all rows noise: ok, remote_desktop: off.
• The branch rebase orphaned the sweep's stamped commit. Rows and sidecar re-stamped to 667d4c1, whose modules/dasLLAMA/ tree is byte-identical to the measured tree; the rebase's only engine-adjacent delta is dasLLVM's darwin in-memory dtor arm - setup/teardown only, and the timed cells are Metal PSOs the LLVM JIT never compiles. Sidecar snapshot renamed by content hash (m5.tune.d30db41e2278.json); site records regenerated; gen_site_records --verify, test_bench_records_schema (37/37) and test_exchange_schema (21/21) green after the re-stamp.
• Suites on the M5 box (Metal-4 toolchain present): run.das -- --suite model-free (46 files green); --suite kernels --arm all (all seven files green; the prefill file's 4 tests include the two new device-free predicate tests); --suite prefill --arm base,kq (q8 and Q4_K_M carriers, prefill GPU-served with rising encode counters, 40-token continuations token-exact vs the CPU control); --suite decode --arm arm1,arm10.
• Negative controls: the q51 twin decode, the REM floor, and the tall MT=128 substitution were each mutated and reddened exactly their named test, then restored; the new selector-parity gate was mutation-controlled the same way.
• Full preflight ran once at the final tip; later fixes validate targeted.

Claims - stated, not tested

• The selection constants (DEVW_MAX_TILES, the 96 MiB TALLKQ_MIN_PANEL, the q51 pick) were adopted on release-grade full-model A/B racing (lcpp_bench -p 512 -n 0 -r 5 --ngl 99, gate forced per run, cv-gated) - two processes, not the in-process interleaved instrument, which is the queued per-box knob-tune arc's first deliverable. A mispicked form is a performance regression only; the bit-exact kernel gates pin correctness.
• The q5_1 model-level pick ladder is exercised by the board's gemma-4-26B cell (its parity pregate is stamped on the row). The MoE tall pair needs npos x experts-per-token at or past 128 x experts - beyond that model's serving window - so its correctness witness is the bit-exact kernel-unit pair gate, including the [136,32] fixture crossing the 128 floor.
• MT=128 reaching the MSL is witnessed behaviorally: m/128-group dispatches compare bit-exact at m=256 and at m=320 with the remainder pair at nonzero x/y offsets.
• The devw_cvt knockout arms are debug instruments (served output is garbage by design, announced at both spellings); they are untested like the sibling knockout arms.

Not done

• The five MoE tensor-twin templates share one scaffold by hand; consolidation into the base-template pattern is followup 53.
• The dupe report's hits are the per-domain stats-struct family and standalone-test-file helpers - established sibling patterns, no extraction.
• The hardware probe stamps remote_desktop on daemon presence, not session activity; 47 historical m1/zen2 rows stamped by a dormant daemon were hand-corrected to off (dormant is harmless - an active session is what taints a measurement), and the probe refinement is followup 54.
• The external review round ran once; its single finding's fix is a predicate extraction plus an override clause - no material logic rewrite, so no re-round.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR advances dasLLAMA’s Metal prefill/perf infrastructure to “board v8” (and bumps DASLLAMA_VERSION to 11), alongside tightening the repository’s review/documentation gates around record-store provenance, selector parity, and kernel coverage.

Changes:

• Added a new make-pr mechanical gate (stamp-reach) to ensure newly-added record-store/sidecar engine stamps remain reachable after rebases.
• Added/expanded REVIEW contracts and REVIEW.das mechanical checks (site dl-* selector parity; performance records/ corpus sweep), and documented new [arch] audit triangle rules repo-wide.
• Expanded Metal kernel tests/coverage lists and updated site/performance record artifacts for the new M5 board refresh.

Reviewed changes

Copilot reviewed 28 out of 29 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
utils/internal/make-pr/main.das	Adds stamp-reach gate to validate reachability of newly-added engine stamps in record stores/sidecars.
utils/internal/make-pr/gates.das	Registers the new stamp-reach gate name.
skills/review_md.md	Updates REVIEW.md opening-line guidance for folders with multiple architecture doc companions.
skills/internal/make_pr.md	Documents the new stamp-reach step in the standard make-pr gate chain.
site/REVIEW.md	Tightens site checklist wording and updates bench_records.json placement rules.
site/REVIEW.das	New mechanical gate to enforce dl-* selector body parity between CSS and inline HTML <style>.
site/files/dasllama/bench_records.json	Updates site-served bench records (dates/SHAs/tune hashes/remote_desktop normalization).
REVIEW_COMMON.md	Adds repo-wide [arch(at=...)] “audit triangle” rules (function/citation/anchor duties).
modules/dasLLAMA/tests/test_metal_prefill_kernels.das	Adds Metal-4 toolchain probe, device-free predicate tests, and q5_1 MoE twin/tall-pair test legs.
modules/dasLLAMA/tests/test_metal_gemm_kernels.das	Extends KQ GEMM gate to cover tall TH128 stamps and remainder-pair encoding.
modules/dasLLAMA/tests/test_kernel_coverage.das	Marks new/unreachable kernel classes as never-dispatched, documenting why.
modules/dasLLAMA/tests/REVIEW.md	Adds/clarifies test weakening bans and kernel-class coverage rules.
modules/dasLLAMA/REVIEW.md	Routes measurement-related diffs to REVIEW_MEASUREMENT.md.
modules/dasLLAMA/REVIEW_MEASUREMENT.md	New measurement checklist extracted from REVIEW.md.
modules/dasLLAMA/REVIEW_GPU.md	Updates GPU checklist wording and adds new correctness/constraint rules.
modules/dasLLAMA/performance/REVIEW.md	Refines performance record/sidecar provenance rules (reachability, noise, boot fallthrough).
modules/dasLLAMA/performance/REVIEW.das	Adds records/ corpus sweep to mechanically enforce remote_desktop/noise invariants.
modules/dasLLAMA/performance/records/zen2.json	Normalizes stored runs’ hardware.remote_desktop to off for historical rows.
modules/dasLLAMA/performance/records/m5.json	Updates M5 record-store rows for refreshed measurements and provenance.
modules/dasLLAMA/performance/records/m1.json	Normalizes stored runs’ hardware.remote_desktop to off for historical rows.
modules/dasLLAMA/PERF_LEDGER.md	Adds follow-up item(s) derived from the prefill-driver/document harvest.
modules/dasLLAMA/followup_general.md	Adds follow-up entries tracking future consolidation/probe refinements.
modules/dasLLAMA/dasllama/dasllama_version.das	Bumps DASLLAMA_VERSION from 10 → 11 with updated arc note.
modules/dasLLAMA/ARCHITECTURE.md	Updates companion-document routing to include ARCHITECTURE_GPU_PREFILL.md.
modules/dasLLAMA/ARCHITECTURE_GPU.md	Refines wording around sanctioned float-A stamps.
modules/dasLLAMA/ARCHITECTURE_GPU_PREFILL.md	New architecture companion documenting the Metal prefill driver (forms, knees, rails).
LAWS.md	Records the repo-wide ruling introducing the [arch] audit triangle rules.

Suppressed comments (1)

utils/internal/make-pr/main.das:144

• Same prefix-matching issue as the tune-sidecar branch: added |> find(sha) can treat a shorter SHA from elsewhere in the file as “added” if a longer SHA with the same prefix was added in the diff. Match the quoted SHA to ensure exact stamp detection.

                    let sha = "{r?["sha"] ?? ""}"
                    if (!empty(sha) && (added |> find(sha) >= 0) && stamps |> find_index(sha) < 0) {
                        stamps |> push(sha)

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.