Expert Parallelism: common C API + NCCL EP backend#3034

Open

phu0ngng wants to merge 21 commits into

NVIDIA:mainfrom

phu0ngng:phuong/ep-2-commwindow

Commits on May 22, 2026

Expert Parallelism: common C API + NCCL EP v0.1 backend
phu0ngng
committed

Commits on May 23, 2026

Expert Parallelism: persistent ncclEpHandle cache with allow_handle_mem_reloc gating
phu0ngng
committed
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed

Commits on May 27, 2026

Commits on May 28, 2026

Expert Parallelism: require token_dtype in NVTEEpGroupConfig and enforce at dispatch
phu0ngng
committed
Expert Parallelism: document ep_comm lifetime, v0.1 single-GPU scope, static layer registration
phu0ngng
committed
Expert Parallelism: drop version label from initialize scope note
phu0ngng
committed
Common: add NVTE_CHECK_NCCL macro and use it in EP tests
phu0ngng
committed
Expert Parallelism tests: replace TensorHandle shim with TensorWrapper
phu0ngng
committed
Expert Parallelism tests: consolidate into single test_ep.cu with essential per-op and pipeline tests
phu0ngng
committed
tests: use test::CudaPtr in DevBuf, check full token rows, simplify bf16 tolerance
phu0ngng
committed
tests: reword EPTensors comment to not imply C-API churn
phu0ngng
committed
EP group config: rename token_dtype to max_token_dtype and allow narrower per-dispatch dtypes
phu0ngng
committed
tests: parameterize FullForwardBackward over bf16, fp16, fp32
phu0ngng
committed
cmake: drop NO_CMAKE_SYSTEM_PATH on TE_LIB lookup and order nvrtc after TE_LIB to remove --no-as-needed
phu0ngng
committed
tests: use MPI for EP distributed tests (bootstrap, build, run script)
phu0ngng
committed
ep.h: add TODO note about struct versioning
phu0ngng
committed

Commits on May 29, 2026