Repository navigation
Conversation
vLLM logs "EngineCore failed to start" ahead of its traceback, so the legacy server.log excerpt (two lines of context) ended at the first frame and dropped the ValueError that names the cause. The failure was recorded as server_init_dead, mapped to engine_core_init in the report and left unclassified by the log classifier. - Keep the exceptions named in the 60 lines below a legacy marker. - Add engine_capacity_cause for the Mamba cache block, KV cache memory and empty KV cache errors; baseline and the grid runner assign engine_capacity ahead of server_init_dead and the baseline record carries the parsed numbers in "capacity". - Report maps the class to root_cause_type engine_capacity and quotes the capacity line; the log classifier files the text under resource_constraint, not a kernel or architecture gap.
|
Gate job conclusion: failure (run) on The report below was uploaded by that run; the conclusion above is the result. Code metrics gate: FAILEDJudged: the 11 files this change touches. Files it does not touch never fail it; anything found there is listed under Outside this change for information.
New violations (not in the baseline): 1
Worse than the baseline: 8
Repeated literals: 2
How to fix
Already over at the base (backlog, not this change's): 1
93 findings in files this change does not touchOutside this change (informational): 93
...and 68 more; the job log lists every row.
Thresholds and sources: |
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
EngineCore failed to startbefore its traceback, so the legacyserver.logexcerpt (two lines of context after a_SERVER_DEAD_MARKERShit) ended at the first frame and dropped theValueErrorthat names the cause. The failure was recorded asserver_init_dead, mapped toengine_core_initin the report, and leftunknownby the log classifier. A baseline that dies onmax_num_seqs (256) exceeds available Mamba cache blocks (237)is the example._subprocess_kill: a legacy marker now also keeps the exception lines in the 60 lines below it (nearest first, withinmax_chars); output is unchanged when none is present. Newengine_capacity_cause()recognises the Mamba cache block, KV cache memory and empty KV cache errors and parses their numbers.baselineand the grid runner assignengine_capacityahead ofserver_init_dead(afteraiter_jit_registry_mismatchandcuda_graph_capture_failed, so the eager-fallback retry is unchanged). The baseline record carriescapacity: {kind, requested, available}and theValueErrorline inerror.report:root_cause_typeisengine_capacityand the headline is the capacity line.failure_signature: three patterns added to the existingRESOURCE_CONSTRAINTrule (no new kind,ladder.pyalready maps it), so the text is not a kernel or architecture gap.failure_evidence(staysunknownfor the class, pinned by a test),recipe_journal,bypass_analysis,kernel/conc_sweep.benchmark_*workspace, grid runner watchdog path, reportroot_cause_typeand headline, signature kind, attribution staysunknown. Onmainthey fail (assertion failures, plus an import error forengine_capacity_cause). The six touched test files pass on currentmainwith this change: 352 passed (python -m pyteston those six files).error_class=engine_capacitywith the cause line and the numbers, instead ofserver_init_deadwith an excerpt ending atcore.py ... run_engine_core. Otherserver_init_deadrecords gain the exception lines that followed their marker.server_init_deadconsumers still receive it for every non-capacity failure.engine_capacityflows through the generic failure path (streak, launch log) likeserver_init_deaddid.max_num_seqs256 below 70 GiB, check after KV cache sizing); the missing class is in Hyperloom.