Skip to content

Add NCCL redist support - #14

Draft
cerisier wants to merge 1 commit into
mainfrom
codex/add-nccl-redist-support
Draft

Add NCCL redist support#14
cerisier wants to merge 1 commit into
mainfrom
codex/add-nccl-redist-support

Conversation

@cerisier

@cerisier cerisier commented Jun 5, 2026

Copy link
Copy Markdown
Member

Summary

  • add optional NCCL redist support via nccl_version / nccl_patches on cuda_ext.redist(...)
  • add NCCL wheel metadata for CUDA 12 and CUDA 13 on linux x86_64 and SBSA
  • include all CUDA 12/13 NCCL wheel versions currently carried by rules_ml_toolchain: 2.27.7, 2.28.9, 2.29.2, 2.29.3, 2.29.7, 2.30.4
  • expose @cuda//nccl proxy targets and add an e2e nccl_smoke target

Validation

  • jq empty nccl/nccl_redist_versions.json
  • bazel build //...
  • cd e2e && bazel build //:cudnn_smoke //:nccl_smoke --platforms //:platform_linux_amd64_cuda_13_3_0
  • temporarily selected NCCL 2.27.7, then validated cd e2e && bazel build //:nccl_smoke --platforms //:platform_linux_amd64_cuda_12_9_1
  • temporarily selected NCCL 2.27.7, then validated cd e2e && bazel build //:nccl_smoke --platforms //:platform_linux_amd64_cuda_13_3_0

Notes

  • NCCL smoke is tagged manual because the wheel is large.
  • Keeps existing platform semantics; no linux-sbsa/linux-aarch64 split.

@cerisier
cerisier force-pushed the codex/add-nccl-redist-support branch from 33d5928 to 6399614 Compare June 5, 2026 15:24
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