Skip to content

[Bugfix] Scope allocator settings during SM70 compile warmup - #840

Merged
yangzhuxinyzx merged 1 commit into
mainfrom
codex/scoped-compile-allocator-20261003
Oct 3, 2026
Merged

yangzhuxinyzx merged 1 commit into
mainfrom
codex/scoped-compile-allocator-20261003

Conversation

@yangzhuxinyzx

Copy link
Copy Markdown
Contributor

Cold compilation can leave small persistent allocations inside large split segments. Use the existing temporary allocator setting during the first SM70 profile forward, then restore the complete configuration before measuring the steady-state forward or returning an explicit KV budget. This extracts the allocator-lifetime fix from #690; projection arithmetic and cache layouts are unchanged.

Validation: tools/merge_gate.sh tests/v1/worker/test_sm70_warmup_allocator.py tests/v1/worker/test_gpu_worker_allocator.py tests/v1/worker/test_gpu_worker_memory_profile.py passed: 95 CPU checks, five hardware-dependent tests skipped. Scoped pre-commit passed. The added cases exercise compile exceptions, explicit KV budgets, ordinary profiling, allocator restoration, existing allocator options and other-platform fallback. Actual peak-memory measurements were not repeated for this revision.

The hybrid-page portion of #690 is already integrated through #829. The tiled projection changes remain separate. This is a split of that contribution rather than a duplicate implementation. AI assistance was used.

Scope compile-time allocation settings and restore the complete configuration before steady-state profiling or returning an explicit KV budget. Extracts the allocator lifetime fix from #690 without changing projection arithmetic.

Co-authored-by: OpenAI Codex <noreply@openai.com>
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.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