[serge] integration failure triage - 2026-08-18

[github-actions]

Automated integration-failure triage for the daily CI window 2026-08-12 → 2026-08-18.

This issue was generated by AI-assisted automation. The grouping, summaries, and recommended follow-up can be incomplete or misleading; verify the failures before acting.

Serge dispatched one task per failure group below — each opens or updates its own PR on a serge/fix/itf-<fingerprint> branch. This table is refreshed in place as Serge runs: a group links its #<pr> when opened, shows 🚫 no fix when Serge found no safe change, ⚠️ task failed on error, or (pending) while still running (a late PR links on the next nightly run).

Dispatched failure groups

Model	Error	Occurrences	PR
kimi_k25 (regressed by PR #47573)	mixed — other (2), tensor values differ (2)	4	🚫 no fix
generation	output_mismatch — list output differs (4)	4	#48070
qwen3_moe	other — other (4)	4	🚫 no fix
exaone4	OOM — other (1)	1	🚫 no fix
gpt_oss	import_or_config — other (1)	1	🚫 no fix
cosmos3_omni (regressed by PR #47096)	mixed — list output differs (2)	2	#48060
deepseek_vl	other — other (3)	3	⚠️ task failed
qwen3_vl_moe	OOM — other (1)	1	🚫 no fix
voxtral_realtime (regressed by PR #47622)	mixed — other (2)	2	⚠️ task failed
phimoe	other — other (2)	2	🚫 no fix

Not dispatched — environment / dependency

These groups were triaged but NOT handed to Serge: their failure mode is a property of the runner or the environment, so no minimal source patch can fix them. They need a human (runner capacity, a dependency pin).

5 models ran out of device memory (13 failures) — needs runner capacity, not a patch, so none of these were dispatched: bamba (4), llama4 (3), cohere2_vision (2), gemma4 (2), pi0 (2).

Outcome recap

Why each group that opened no PR ended without one, and what it cost — surfaced here from the Serge dashboard. Failing tests links straight to each test's dashboard view, since these are the groups a human has to pick up.

Group	Reason	LLM	Tokens (in / out)	Failing tests
kimi_k25 (regressed by PR #47573)	[not_fixed] GPU verification did not confirm the fix (not_fixed): the patch did not turn the targeted tests green.	moonshotai/Kimi-K2.7-Code	2,067,305 / 3,048	test_model_logits · test_model_logits_batched
qwen3_moe	[reproduced] GPU reproduce + classify: this is an ENVIRONMENT issue (The traceback shows a 4-bit quantized 15B model being dispatched to CPU/disk because the two-GPU CI runner lacks enough GPU RAM, which is a hardware/environment capacity problem, not a library bug or stale test expectation.) — no source patch can fix it, so no investigation and no PR.	moonshotai/Kimi-K2.7-Code	— / —	test_model_15b_a2b_generation · test_model_15b_a2b_logits · test_model_15b_a2b_long_prompt_sdpa · +1 more
exaone4	[reproduced] GPU reproduce + classify: this is an ENVIRONMENT issue (The failure is a CUDA out-of-memory error during a large-model generation integration test, which is a property of the GPU memory capacity and allocation state rather than a bug in the model or test logic.) — no source patch can fix it, so no investigation and no PR.	moonshotai/Kimi-K2.7-Code	— / —	test_model_generation_beyond_sliding_window
gpt_oss	[reproduced] GPU reproduce + classify: this is an ENVIRONMENT issue (The traceback is a clean ImportError stating the kernels package is missing or has an incompatible version, which is an environment/dependency problem, not a bug in the library code or test expectations.) — no source patch can fix it, so no investigation and no PR.	moonshotai/Kimi-K2.7-Code	— / —	test_model_outputs_02

Generated 2026-08-18T23:02:10+00:00.