Skip to content

[atom-vllm] remove benchmark override#1070

Draft
zejunchen-zejun wants to merge 1 commit into
dev_perf_0601from
fix/nightly-benchmark-overrides-env-merge
Draft

[atom-vllm] remove benchmark override#1070
zejunchen-zejun wants to merge 1 commit into
dev_perf_0601from
fix/nightly-benchmark-overrides-env-merge

Conversation

@zejunchen-zejun
Copy link
Copy Markdown
Collaborator

No description provided.

BENCHMARK_OVERRIDES.update() replaced the matrix entry's env_vars
wholesale, silently dropping variables that were set only on the matrix
entry. This lost GATED_DELTA_RULE_TRITON_AUTOTUNE=1 for all 8
Qwen3-Next / Qwen3.5 GDN-family models, so the nightly accuracy test ran
their Triton GDN prefill with autotune OFF — diverging from the
benchmark workflow (oot_benchmark_models.json keeps it ON).

Merge env_vars per key (override wins on conflicts, base-only keys
preserved) so no env var is silently dropped. Non-env_vars keys
(extra_args, runner, ...) keep the existing full-replace behavior.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant