[None][test] Reduce multi-GPU pre-merge time via shared MPI sessions and HF weight cache - #15872

Closed

sunnyqgg wants to merge 21 commits into

NVIDIA:mainfrom

sunnyqgg:grouped_test_premerge

Commits on Jul 2, 2026

[None][test] Optimize LLM test startup overhead
sunnyqgg
committed
[None][test] Group multi-GPU LLM API tests
sunnyqgg
committed
[None][test] Reuse MPI sessions with pytest fixtures
sunnyqgg
committed
[None][test] Parameterize DeepSeek NVFP4 pre-merge group
sunnyqgg
committed
[None][fix] Reuse PP NCCL communicator across LLMs sharing worker processes
sunnyqgg
committed
[None][test] Harden shared-MPI-session reuse for grouped NVFP4 tests
sunnyqgg
committed
[None][test] Reset torch.compile state between grouped NVFP4 cases
sunnyqgg
committed
[None][test] Extract shared session-reuse helpers into a single in-package module
sunnyqgg
committed
[None][test] Extract hf_weight_cache_env into the shared helper module
sunnyqgg
committed
[None][test] Simplify multi-GPU session plumbing via mpi_kwargs fixtures
sunnyqgg
committed
[None][test] Simplify NVFP4 premerge case table and shared-LLM factory
sunnyqgg
committed
[None][test] Explicitly invalidate HF weight cache on grouped-test teardown
sunnyqgg
committed
[None][test] Organize multi-GPU tests into gpu2/gpu4 classes with class-scoped sessions
sunnyqgg
committed
[None][test] Add CUTEDSL cases to grouped NVFP4 premerge matrix
sunnyqgg
committed
[None][test] Drop carried id from NVFP4 premerge case dicts
sunnyqgg
committed
[None][test] Mark HfWeightLoader weight-cache clear as private
sunnyqgg
committed
[None][test] Group bf16 4-GPU pre-merge cases into the shared 4-GPU session
sunnyqgg
committed
[None][fix] Evict HF weight cache before load on miss to avoid 2x CPU peak
sunnyqgg
committed
[None][test] Dedup shared-session compile-reset teardown; trim docstring
sunnyqgg
committed
[None][test] Use fully-parametrized ids for grouped test-db entries
sunnyqgg
committed
[None][test] Group P0/P1 pre-merge multi-GPU accuracy cases onto shared MPI sessions
sunnyqgg
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[None][test] Reduce multi-GPU pre-merge time via shared MPI sessions and HF weight cache - #15872

[None][test] Reduce multi-GPU pre-merge time via shared MPI sessions and HF weight cache#15872
sunnyqgg wants to merge 21 commits into
NVIDIA:mainfrom
sunnyqgg:grouped_test_premerge

Commits on Jul 2, 2026

[None][test] Optimize LLM test startup overhead

[None][test] Group multi-GPU LLM API tests

[None][test] Reuse MPI sessions with pytest fixtures

[None][test] Parameterize DeepSeek NVFP4 pre-merge group

[None][fix] Reuse PP NCCL communicator across LLMs sharing worker processes

[None][test] Harden shared-MPI-session reuse for grouped NVFP4 tests

[None][test] Reset torch.compile state between grouped NVFP4 cases

[None][test] Extract shared session-reuse helpers into a single in-package module

[None][test] Extract hf_weight_cache_env into the shared helper module

[None][test] Simplify multi-GPU session plumbing via mpi_kwargs fixtures

[None][test] Simplify NVFP4 premerge case table and shared-LLM factory

[None][test] Explicitly invalidate HF weight cache on grouped-test teardown

[None][test] Organize multi-GPU tests into gpu2/gpu4 classes with class-scoped sessions

[None][test] Add CUTEDSL cases to grouped NVFP4 premerge matrix

[None][test] Drop carried id from NVFP4 premerge case dicts

[None][test] Mark HfWeightLoader weight-cache clear as private

[None][test] Group bf16 4-GPU pre-merge cases into the shared 4-GPU session

[None][fix] Evict HF weight cache before load on miss to avoid 2x CPU peak

[None][test] Dedup shared-session compile-reset teardown; trim docstring

[None][test] Use fully-parametrized ids for grouped test-db entries

[None][test] Group P0/P1 pre-merge multi-GPU accuracy cases onto shared MPI sessions