[hsliuustc0106]

Metrics

Item	Value
Kept primary	fcf5e2bef916da81180ac8f1570dff338e8a049d
Primary result	231/231, p50 87.782 ms, p95 291.183 ms
Rejected tip	801ef7b103c9285bef749d54af6cea6f4472088b
Rejected result	231/231, p50 89.172 ms (+1.58% / +1.390 ms), p95 274.790 ms (−5.63% / −16.393 ms)
Gate	both percentiles strictly under → HOLD_PRIMARY at fcf5e2be 87.782 / 291.183
SwiGLU tip	2bbbe4eda71bb745c3b87a1f06c4a23eae68db4d rejected in the owner closeout at p50 97.613 ms, p95 292.795 ms. Not installed
Branch head	509bdd2161ca5bef8f5e158d466514862fac7c9e
Running path	kept primary behavior. mlp=default, fused_swiglu_layers=0, gated_norm=eager

Draft stays open. Do not merge without a separate owner yes. No new latency win is claimed.

Tracks #161.

Required human contributor answers

1. What outcome are you trying to achieve?

   Answer: Reject 801ef7b, keep fcf5e2be, drop always-on FLA fused gated_norm, and push one tip that beats both 87.782 and 291.183 without sacrificing p50. Draft pplx-decider: reference worker and Public231 C1 procedure #166 stays open. No merge. Source: "Decision: reject this tip. ... Protect p50. Drop always-on FLA fused gated_norm ×48 on this path (or any tip that trades p50 for p95). Keep primary GVA+contiguous. Next tip must beat both 87.782 and 291.183 without sacrificing p50. Draft pplx-decider: reference worker and Public231 C1 procedure #166 stays open on fcf5e2be. No merge." The later owner closeout records 2bbbe4ed as rejected and closes the latency story at fcf5e2be.

2. What is the highest-leverage variable for that outcome?

   Answer: Both p50 and p95 strictly under 87.782 / 291.183, protecting p50. Source: "Gate: both must be strictly under → HOLD_PRIMARY" and "Same pattern as prior FLA residual RMSNorm tips — p95 win, p50 lose. Protect p50."

3. How will this work move that variable rather than create busywork?

   Answer: Protect p50. Drop always-on FLA fused gated_norm. Do not trade p50 for p95. The owner did not name the next kernel. Source: "Protect p50. Drop always-on FLA fused gated_norm ×48 on this path (or any tip that trades p50 for p95)."

4. Who is the customer or user, and what do they actually need?

   Answer: The owner of draft pplx-decider: reference worker and Public231 C1 procedure #166. The draft stays open and this tip is not merged. Source: "Draft pplx-decider: reference worker and Public231 C1 procedure #166 stays open on fcf5e2be. No merge." This turn also follows the owner request to address the review on pplx-decider: reference worker and Public231 C1 procedure #166 without merging.

After answering the four questions: is a PR needed at all to achieve this outcome?

Human decision and rationale (or explicit source reference): Yes. Continue draft #166. Do not merge. Source: "Draft #166 stays open on fcf5e2be. No merge." and the owner closeout "Do not merge without a separate yes."

For a bug fix: link actual observed failure evidence.

Evidence: Not a bugfix of the kept primary. 801ef7b ran and missed the gate: 231/231, p50 89.172 ms, p95 274.790 ms. The owner closeout records 2bbbe4ed at p50 97.613 ms, p95 292.795 ms and rejects it. The review comments are code-correctness items, not a new latency result.

Purpose

Serve perplexity-ai/pplx-decider-v1.1-27b at 14c9d90137962d300f3de95633f7b6535e80c5ce.

The kept primary is fcf5e2bef916da81180ac8f1570dff338e8a049d: 231/231, p50 87.782 ms, p95 291.183 ms. 801ef7b and 2bbbe4e are rejected. 509bdd2 does not install either kernel. Startup again matches the packed GVA primary.

Review comments from #166 (comment):

• Unknown PPLX_GDN_PREFILL raises before the capability check, including on CPU and capability 8.9.
• decide() maps only Question branch exceeds the 8192-token limit; no input was truncated. to HTTP 422. Any other ValueError from evaluate stays a 500.
• verify_checkpoint failure paths have weight-free tmp_path tests: missing file, hash mismatch, decision_config field mismatch.
• Falsy admitted prompts ("", [], {}) are pinned by noul-falsy-fallbacks in tests/pplx_decider/fixtures/reference_cases.json, which check_contract.py already walks, plus a unit test.
• tarfile.extractall(..., filter="data") in install_causal_conv1d.py.
• The 12,000,000-character image rejection no longer says "8 MB before base64". The decoded 8 MB limit keeps its own sentence.

origin/main moved to 8228551. git merge-tree --write-tree against that main exited 0, so no conflict commit was made.

Test Plan

• PYTHONPATH=src python3 -m pytest tests/pplx_decider -q
• PYTHONPATH=src python3 -m compileall -q src/models/pplx_decider src/frontend/pplx_decider.py tests/pplx_decider recipe/pplx_decider/install_causal_conv1d.py
• No new Public231 C1. The running path is the kept primary, not a new tip.

System1-Omni Version / Commit: kept primary fcf5e2bef916da81180ac8f1570dff338e8a049d. Head 509bdd2161ca5bef8f5e158d466514862fac7c9e on cursor/pplx-decider-27b-8ee9. Merge base cd302b6e3d7723ebfc33bba9698da2c1e881f203.

Test Result

CPU checks on 509bdd2: pytest 115 passed, compileall success. No NVIDIA device here, so the CUDA forward was not executed.

The falsy-fallback fixture was compared locally to an AST extract of the pinned model.py (describe, options, decision_messages, answer). Messages and the noul answer matched. check_contract.py against a checkpoint directory was not run. mkdocs build --strict was not run.

Diff size and review map

Against merge base cd302b6e, including this head:

Bucket	Changed lines (add + delete)
src/ Python	2602
tests/ Python	1694
recipe/ Python	167
Authored Python	4463
Docs and recipe Markdown	957
Other (fixtures, config)	1073
Total	6493

The diff is one worker, its tests, and its recipe. Splitting it would separate the contract from the checks that pin it. Read in this order:

1. src/models/pplx_decider/contract.py and tests/pplx_decider/test_contract.py — admission, prompts, image limits.
2. src/models/pplx_decider/engine.py and tests/pplx_decider/test_checkpoint.py — pin and startup, including unconditional PPLX_GDN_PREFILL.
3. src/frontend/pplx_decider.py and tests/pplx_decider/test_worker.py — 422 only for the token-limit sentence; other ValueError is 500.
4. gdn.py, attention.py, conv.py, mlp.py, rmsnorm.py, gated_norm.py, swiglu.py, compile_forward.py — kernel binds. swiglu.py and gated_norm.py remain as rejected implementations and are not installed.
5. recipe/pplx_decider/ and the status docs — measured primary fcf5e2b; rejected tips stay rejected.

Demo / evidence

N/A for a video. Latency evidence is the owner pastes and the owner closeout already in this description. Raw responses.jsonl was not attached. This commit has no new latency sample.

Self-review

• I have reviewed the full diff and addressed the issues I found.
• I have checked that the change follows the project's architecture and stays focused on the stated purpose.
• I have run the checks appropriate to this change and reported commands, results, and anything I could not verify above.
• I have checked that the PR description, documentation, and any accuracy or performance claims match the implementation and available evidence.

The PR stays draft. Do not merge.

Optimization closeout (2026-10-25)

Status: ready for owner review — no merge
Close at best-so-far primary: fcf5e2bef916da81180ac8f1570dff338e8a049d
Protocol: Public231 C1, warmup 5, concurrency 1, metric successful_latency
Gain chart: shared in system1 room

Goal

Serve perplexity-ai/pplx-decider-v1.1-27b with a measured Public231 C1 primary, advancing only when both p50 and p95 beat the kept primary (231/231, quality not regress).

Current primary (best so far)

Item	Value
SHA	fcf5e2bef916da81180ac8f1570dff338e8a049d
Success	231/231
p50 / p95	87.782 ms / 291.183 ms
Quality	choice_acc 0.9065 (139), noul_acc 0.8784 (74), score_mae 0.2035 (18)
Lever	FlashInfer GVA (gdn_layout=gva, 48 layers) + contiguous q/k before FLA l2norm; eager residual RMSNorm; 48-slot cu_seqlens ring; compile off
Branch tip head	509bdd2 review fixes. Historical 2bbbe4ed (FLA swiglu) stays REJECT and is not installed

收益 (gains) vs campaign start

Campaign start = earliest remesure baseline 1ba07d15 98.777 / 341.202.

	Start 1ba07d15	Final fcf5e2be	Δ absolute	Δ %	Speedup
p50	98.777 ms	87.782 ms	−10.995 ms	−11.13%	1.13x
p95	341.202 ms	291.183 ms	−50.019 ms	−14.66%	1.17x

Vs immediate prior primary 4b99b73 (88.942 / 295.797): p50 −1.30%, p95 −1.56% (the KEEP that set fcf5e2be).

搜索过程 (search process)

Best-so-far rises only on KEEP . Selected ledger (disk GATE/REPORT where available):

Step	SHA	Lever / hypothesis	p50 / p95 (or fail)	Gate
T+0	1ba07d15	Campaign start baseline	98.777 / 341.202	start
T+1	f968e576	fla-chunk remesure (kept primary for early tips)	94.077 / 300.200†	KEEP
T+2	3c72f02	All-sequence FlashInfer	88.808 / 300.751 (p50↓ p95↑)	REJECT
T+3	a50b9aaa	Short FI / hold path	95.446 / 300.871	REJECT
T+4	0dcaa0ac	cuDNN-first SDPA	138.501 / 332.859	REJECT
T+5	8b4ee326	Fused gate-up MLP ×64	102.437 / 304.970	REJECT
T+6	e36b32ed	torch.compile dynamic	97.751 / 263.124 (p95↓ p50↑)	REJECT
T+7	7e3b1be1	Dynamic-tail compile	100.037 / 257.084	REJECT
T+8	34ba5de	Compile median probe	97.084 / 256.679	REJECT
T+9	ffae182c	Shared cu_seqlens per device+length	92.844 / 296.458	KEEP
T+10	4b99b73	48-slot cu_seqlens ring + all-seq FI	88.942 / 295.797	KEEP
T+11	17a7b120	GDN recurrent-state ring	89.165 / 295.723	REJECT
T+12	3afb472	FLA residual RMSNorm always-on	93.935 / 274.902 (p95↓ p50↑)	REJECT
T+13	705ae089	FLA RMSNorm only ≥512 tok	94.604 / 274.835	REJECT
T+14	a6f4d9a	GDN use_cp heuristic	112.294 / 299.419	REJECT
T+15	3ecdf5ad	GVA layout (no contiguous yet)	warmup fail l2norm non-contiguous, 0/231	REJECT
T+16	fcf5e2be	GVA + contiguous q/k	87.782 / 291.183	KEEP
T+17	801ef7b1	FLA gated_norm ×48	89.172 / 274.790 (p95↓ p50↑)	REJECT
T+18	2bbbe4ed	FLA swiglu_fwd ×64	97.613 / 292.795	REJECT

† First disk remesure of f968e576 was 98.037/302.569 vs 1ba07d15; subsequent tips gated against the settled primary paste 94.077/300.200 recorded in the PR tip chain.

What worked vs rejected (lessons)

• Worked (KEEP chain): fla-chunk bring-up → shared / 48-slot cu_seqlens + FlashInfer → GVA + contiguous q/k (layout beat raw kernel swaps).
• Rejected pattern: many tips improved only p95 (compile, FLA RMSNorm, gated_norm) or only p50 (3c72f02) — dual gate correctly HOLDs. Blind FLA kernel swaps (swiglu, always-on gated_norm) hurt p50.
• Lesson: layout/contiguity > “swap FLA kernel”; smoke before C1 (3ecdf5ad warmup crash); next tips need profile / length buckets, not more unattributed kernel tries.

Closeout

Draft for owner review. Close latency story at fcf5e2be. Head 509bdd2 does not install rejected 2bbbe4ed. Do not merge without a separate yes.

Optimization closeout (2026-10-25)

Status: ready for owner review — no merge
Close at best-so-far primary: fcf5e2bef916da81180ac8f1570dff338e8a049d
Protocol: Public231 C1, warmup 5, concurrency 1, metric successful_latency
Gain chart: shared in system1 room

Goal

Serve perplexity-ai/pplx-decider-v1.1-27b with a measured Public231 C1 primary, advancing only when both p50 and p95 beat the kept primary (231/231, quality not regress).

Current primary (best so far)

Item	Value
SHA	fcf5e2bef916da81180ac8f1570dff338e8a049d
Success	231/231
p50 / p95	87.782 ms / 291.183 ms
Quality	choice_acc 0.9065 (139), noul_acc 0.8784 (74), score_mae 0.2035 (18)
Lever	FlashInfer GVA (gdn_layout=gva, 48 layers) + contiguous q/k before FLA l2norm; eager residual RMSNorm; 48-slot cu_seqlens ring; compile off
Branch tip head (not kept)	2bbbe4ed (FLA swiglu) — measured, REJECT

收益 (gains) vs campaign start

Campaign start = earliest remesure baseline 1ba07d15 98.777 / 341.202.

	Start 1ba07d15	Final fcf5e2be	Δ absolute	Δ %	Speedup
p50	98.777 ms	87.782 ms	−10.995 ms	−11.13%	1.13x
p95	341.202 ms	291.183 ms	−50.019 ms	−14.66%	1.17x

Vs immediate prior primary 4b99b73 (88.942 / 295.797): p50 −1.30%, p95 −1.56% (the KEEP that set fcf5e2be).

搜索过程 (search process)

Best-so-far rises only on KEEP . Selected ledger (disk GATE/REPORT where available):

Step	SHA	Lever / hypothesis	p50 / p95 (or fail)	Gate
T+0	1ba07d15	Campaign start baseline	98.777 / 341.202	start
T+1	f968e576	fla-chunk remesure (kept primary for early tips)	94.077 / 300.200†	KEEP
T+2	3c72f02	All-sequence FlashInfer	88.808 / 300.751 (p50↓ p95↑)	REJECT
T+3	a50b9aaa	Short FI / hold path	95.446 / 300.871	REJECT
T+4	0dcaa0ac	cuDNN-first SDPA	138.501 / 332.859	REJECT
T+5	8b4ee326	Fused gate-up MLP ×64	102.437 / 304.970	REJECT
T+6	e36b32ed	torch.compile dynamic	97.751 / 263.124 (p95↓ p50↑)	REJECT
T+7	7e3b1be1	Dynamic-tail compile	100.037 / 257.084	REJECT
T+8	34ba5de	Compile median probe	97.084 / 256.679	REJECT
T+9	ffae182c	Shared cu_seqlens per device+length	92.844 / 296.458	KEEP
T+10	4b99b73	48-slot cu_seqlens ring + all-seq FI	88.942 / 295.797	KEEP
T+11	17a7b120	GDN recurrent-state ring	89.165 / 295.723	REJECT
T+12	3afb472	FLA residual RMSNorm always-on	93.935 / 274.902 (p95↓ p50↑)	REJECT
T+13	705ae089	FLA RMSNorm only ≥512 tok	94.604 / 274.835	REJECT
T+14	a6f4d9a	GDN use_cp heuristic	112.294 / 299.419	REJECT
T+15	3ecdf5ad	GVA layout (no contiguous yet)	warmup fail l2norm non-contiguous, 0/231	REJECT
T+16	fcf5e2be	GVA + contiguous q/k	87.782 / 291.183	KEEP
T+17	801ef7b1	FLA gated_norm ×48	89.172 / 274.790 (p95↓ p50↑)	REJECT
T+18	2bbbe4ed	FLA swiglu_fwd ×64	97.613 / 292.795	REJECT

† First disk remesure of f968e576 was 98.037/302.569 vs 1ba07d15; subsequent tips gated against the settled primary paste 94.077/300.200 recorded in the PR tip chain.

What worked vs rejected (lessons)

• Worked (KEEP chain): fla-chunk bring-up → shared / 48-slot cu_seqlens + FlashInfer → GVA + contiguous q/k (layout beat raw kernel swaps).
• Rejected pattern: many tips improved only p95 (compile, FLA RMSNorm, gated_norm) or only p50 (3c72f02) — dual gate correctly HOLDs. Blind FLA kernel swaps (swiglu, always-on gated_norm) hurt p50.
• Lesson: layout/contiguity > “swap FLA kernel”; smoke before C1 (3ecdf5ad warmup crash); next tips need profile / length buckets, not more unattributed kernel tries.

Closeout

Draft for owner review. Close latency story at fcf5e2be. Branch head may still be 2bbbe4ed (rejected) — prefer tip head reset to primary before merge review. Do not merge without a separate yes.

  

[psv666]

The next step at head 1ba07d15b11a8c0796657e4a10b7ba3e62354421 should start with load feasibility on the actual operator hardware. The recorded shards + readout total 52,173,413,928 bytes, before runtime/workspace overhead; the body correctly avoids assuming a 48 GB device can load the BF16 resident set. Please record available memory and the actual placement/precision used, including a load failure if that is the result. A different placement/precision must be labeled separately from the intended baseline.

After successful load/warmup, capture one real choice, noul and score request through the worker and Rust frontend, then pinned-reference parity, including images if image support is claimed ready. Preserve the exact-integer and overlength contract checks. Only then run the documented three Public231 C1 traversals, retaining all 693 planned outcomes, raw failures, head/manifest/checkpoint hashes, actual hardware, p50 and p95.

A small reporting clarification would help now: mark the current 0/693 as not run (0 attempted; 693 planned) so it cannot be read as 693 attempted failures. The body already discloses no CUDA inference, and the metrics should convey the same status.

Please also record the scope relationship with #140 (reference-worker initial serving milestone versus any retained Rust processor work), as raised there. Keep this draft and the explicit optimization hold until real measurements arrive. This comment does not claim new model execution or lift that hold.

_{Agent-assisted readiness follow-up using the repository review skill pinned at trusted main cd302b6e3d7723ebfc33bba9698da2c1e881f203 (SHA-256 ae2dfac116376ae2ec41f65122cc1bf5aea10751b8b907198a56154e5d67230e). This comment reconciles current main and the existing discussion; it is not a new full code-review approval or a claim of newly executed model tests.}

[MrDongsls]

Thanks for this. No correctness findings on the main contract path; re-verified locally: 86 passed, compileall clean, CI (incl. strict docs build) green, and the leaderboard / PR-body deltas recompute correctly.

P2 The template self-review checklist is still unchecked and the body does not yet report authored-vs-total diff counts (authored Python ≈3.2k lines, total ≈4.8k). Per CONTRIBUTING.md#self-review-before-requesting-review, a diff above 3,000 lines needs the completed checklist plus split rationale / component map.

P2 On capability 8.9 an unknown PPLX_GDN_PREFILL value is silently ignored: gdn_prefill_mode() is only called inside the capability-9.0 branch (src/models/pplx_decider/engine.py), while the recipe documents that startup raises on unknown values. Would you consider validating it unconditionally so a typo fails closed on 8.9 as well?

P2 decide() maps every ValueError from executor.evaluate to HTTP 422, so a server-side fault inside the checkpoint forward would surface as a client error, and the bench client would book it as a rejected request instead of a failure. Suggest mapping only the known checkpoint sentence (the token-limit ValueError) and letting anything unexpected surface as 500.

P3 verify_checkpoint looks like the only public function in engine.py without a weight-free test; its failure paths (missing file / hash mismatch / decision_config field mismatch) should be straightforward to cover with tmp_path.

P3 The or fallbacks in options() / decision_messages() replace admitted-but-falsy values ("", {}, [] criteria or instructions) with the default prompt text. Probably matches the checkpoint, but untested; one fixture case checked via check_contract.py would pin it either way.

P3 (nits) tarfile.extractall in install_causal_conv1d.py without filter="data" (3.12 deprecation, 3.14 behavior change); the >12,000,000-char image rejection message says "8 MB before base64", which describes the separate decoded-bytes limit.

Provenance: skill .agents/skills/system1-omni-review/SKILL.md @ trusted main cd302b6e (SHA-256 ae2dfac1…). Full diff read at both heads (7e3b1be, 34ba5de); personally run at 34ba5de8: PYTHONPATH=src python3 -m pytest tests/pplx_decider -q (86 passed) and compileall. Not run: local mkdocs build --strict (no mkdocs; CI docs job passed), check_contract.py (no checkpoint), CUDA paths (no GPU).