Repository navigation
lambda run/chat cells time out: unchunked dense prefill OOM, qk-norm forced serial, chat KV 2048, Qwen3.5 no F16/IQ4_XS dequant #4621
Description
Child of #3715 (release-readiness SHACL gate): the lambda run/chat cells time out at 600 s, so the gate cannot get Pass receipts for them.
Root causes (aprender-ec, fix PR #4620):
- Unchunked dense prefill runs out of memory at 8k–20k context.
- qk-norm is forced onto the serial path.
- Chat KV cache is capped at 2048.
- Qwen3.5 has no F16/IQ4_XS dequant.
Evidence: rg7 on a5dc47a: 8B 8k 243 s GPU, 8B chat 22 s, 1.7B 20k 42 s, 0.5B-IQ3_M 20k 51 s, no CPU fallback.
Out of scope, follow-up: UD-IQ2_XXS needs 5 more dequant kernels.
Activity
test(cuda): the fit-check no-op cases assert the path of the profile …
fix(#4621): a chunked prefill's later chunks never replay the prefill…
evidence(pr-review): #4620 Arm-4 receipt for split head 31cd3f6 (PMAT…
evidence(pr-review): #4620 post-split quorum (3/3 AGREED, scope A) + …
fix(mutants-gate): a mutant in code the build never compiles is NOT_M…
ci(mutants): cuda-gated mutants are NOT_MEASURED on the CPU gate and …
test(cuda): model-free kill tests for qwen35_dequant_f32 and prefill_…
fix(mutants): the feature shard passed --lib to cargo test twice (#4621)
fix(mutants): --lib must reach the feature shard's build, once (#4621)
fix(mutants,roadmap): --cargo-arg --lib spelling for bashrs 7.4.1; re…
fix(ci): the fat driver's results output dropped section outputs — th…
fix(mutants): --defer-unbuilt file's directory is made before it is w…
merge origin/main into ec/4621-chat-session-kv
fold: #4630 into #4623 -- serve sizes its GPU KV to the model context…
noahgift commentedon Oct 14, 2026
P0 to P1 (2026-10-14 P0 triage): P0 now means it blocks the next release or carries the operator's own words on it; this issue does neither, so it stays on the 0.71.0 milestone at P1.
noahgift commentedon Oct 14, 2026
Folded into #3715 as a checklist row (TRIAGE-100, 2026-10-14): the open-ticket cap is 100, so this ticket is now tracked on that master and closed as not planned. Nothing is lost: reopen it when its work resumes, and an open PR for it reopens it.
Agent: aprender-clerk