Skip to content

Conversation

@denera
Copy link
Collaborator

@denera denera commented Dec 2, 2025

Description

This PR adds support for the NVTE cuBlasMp bindings in the Comm+GEMM overlap API.

Type of change

  • Documentation change (change only to the documentation, either a fix or a new content)
  • Bug fix (non-breaking change which fixes an issue)
  • New feature (non-breaking change which adds functionality)
  • Breaking change (fix or feature that would cause existing functionality to not work as expected)
  • Infra/Build change
  • Code refactoring

Checklist:

  • I have read and followed the contributing guidelines
  • The functionality is complete
  • I have commented my code, particularly in hard-to-understand areas
  • I have made corresponding changes to the documentation
  • My changes generate no new warnings
  • I have added tests that prove my fix is effective or that my feature works
  • New and existing unit tests pass locally with my changes

@denera denera self-assigned this Dec 2, 2025
@denera denera force-pushed the common/tp-overlap-cublasmp branch from 9c8fe95 to 334e8c4 Compare December 2, 2025 19:51
@denera denera force-pushed the common/tp-overlap-cublasmp branch from 908bbc2 to 69cf235 Compare December 2, 2025 20:12

CommOverlapCore::CommOverlapCore(int64_t nccl_comm_ptr, int tp_rank, int tp_size, int num_comm_sm,
bool is_p2p, bool atomic_gemm) {
_with_cublasmp = true;
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This should probably check if TE was built with cublasmp and error out otherwise.


#define NVTE_COMM_OVERLAP_MAX_STREAMS 3

/* \brief Check if TE is built with cuBlasMp.
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nit: cuBLASMp

void ub_barrier(ExtComm comm);

int64_t get_nccl_comm_ptr(std::string comm_name) {
NVTE_CHECK(backend_is_nccl, "Cannot get nccComm_t ptr if backend is not NCCL.");
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This error message could be more descriptive - e.g. something like "chosen backend for the communication-computation overlap (cuBLASMp) requires NCCL communicator, but the passed ProcessGroup uses a different backend."

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.

2 participants