Repository navigation
V3 serving-shape parity: lambda perf041 witness for Qwen3.5-4B at c in {1,4,8,16} #4971
Description
PROPOSE-TICKET from la-0.71.1a (12:43Z). Spec: la-71/v3-serving-shape-parity@adf6c4b4aa.
Done when: one lambda perf041 witness for Qwen3.5-4B (sha256 recorded) measures c in {1,4,8,16} with m_formed == c, divergence_at >= declared_min vs m=1 or top2_margin < 0.05 logit, perf041 PASS per band, and c=4 per-request decode >= 50% of c=1 in the same run.
Falsifier: the same witness with one band at m_formed = c-1, or one divergence without a margin, makes the Rust check exit non-zero.
- V3-a top-2 margin first
- V3-b
- V3-c
- V3-d
Placement: epic E4 #4381 (Qwen 3.5 4B lane #4354) is full; held unlinked and flagged for the first E4 slot, by MOVE. Receipt: handoff/quorum-triage-propose3-q11.md
Activity
noahgift commentedon Oct 10, 2026
Placed under #3597 (verb set V) by MOVE, at the cop's order: E4 #4381 is 9/9 and the epics with room are later trains. #3597 now has 4 open children (≤5). Milestone stays 0.71.1 like #4984, #4985 and #4994. Nothing reads sub-issue membership (no code, workflow or contract), so no gate is affected.
feat(serve): chat logprobs on the quantized CPU path; every other bac…
feat(serve): streamed chat logprobs on the quantized CPU path (PMAT-4…
feat(serve): chat logprobs on the Qwen3.5 hybrid, streamed and not (P…
fix(serve): the chat request names TopLogprobs in scope; DeviceGreedy…
refactor(serve): split the CUDA batch scheduler loop and process_cuda…
feat(serve): chat logprobs on the dense CUDA backend, direct and thro…
feat(apr-cli): `apr test llm shape-check`, the CLI over the V3 shape …
fix(apr-cli): a RED V3 witness exits 4 as InvalidInput, not "Invalid …
feat(perf-gate): the PP-26 witness and its top-2 margin from OpenAI l…
chore(contracts): ONT witness for serving-shape-parity-v1 (PMAT-4971)
feat(perf_gate): V3 reads S7, the near-tie cap and the declared_min f…
docs(contracts): serving-shape-parity-v1 records the S7 run on gx10 (…
9 remaining items
noahgift commentedon Oct 14, 2026
V3 stays on the blessed Qwen3.5-4B; the contract and the check are unchanged. V3 now depends on #5279 (Qwen3.5 on CUDA never batches concurrent requests). Once #5279 lands, re-run the witness on both host classes, and V3 closes when
apr test llm shape-checkadmits it on each.noahgift commentedon Oct 14, 2026
merged into #5312 (master ticket for epic #3994; triage rule 7b, 2026-10-14). It is a checklist row there; reopen this issue when work on it starts or its PR opens.