Skip to content

fix: support system NCCL installations - #738

Open
janbernloehr wants to merge 1 commit into
deepseek-ai:mainfrom
janbernloehr:system-nccl-main
Open

fix: support system NCCL installations#738
janbernloehr wants to merge 1 commit into
deepseek-ai:mainfrom
janbernloehr:system-nccl-main

Conversation

@janbernloehr

Copy link
Copy Markdown

Summary

DeepEP currently assumes NCCL is installed through the PyPI package
nvidia-nccl-cu12. This fails for standard Ubuntu installations where NCCL
is provided externally, for example:

  • /usr/include/nccl.h
  • /usr/lib/aarch64-linux-gnu/libnccl.so

This change:

  • Discovers NCCL through standard multiarch/system library locations.
  • Preserves the existing PyPI-package discovery path.
  • Links the exact discovered library without overriding the system NCCL version.
  • Avoids adding an unnecessary RPATH for system-installed NCCL.
  • Uses the same discovery logic for runtime checks and JIT compilation.

HybridEP does not require this change: its NIXL path skips NCCL/DOCA, while its
DOCA path intentionally builds the bundled NCCL submodule components.

Testing (internal)

  • Python syntax and whitespace checks passed.
  • System NCCL discovery was validated in the Core Models mbridge build.
  • DeepEP 01dc3aaac82068020353dce2c302e38153c0bfaa built and installed
    successfully with system NCCL.
  • deep_ep._C metadata validation and import deep_ep passed.

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