Skip to content

python3Packages.vllm: disable Blackwell GPU support to fix CUDA build#448965

Merged
GaetanLepage merged 1 commit intoNixOS:masterfrom
daniel-fahey:fix-vllm
Oct 7, 2025
Merged

python3Packages.vllm: disable Blackwell GPU support to fix CUDA build#448965
GaetanLepage merged 1 commit intoNixOS:masterfrom
daniel-fahey:fix-vllm

Conversation

@daniel-fahey
Copy link
Contributor

@daniel-fahey daniel-fahey commented Oct 5, 2025

Fixes type of build failure seen here.

  • vLLM v0.11.0's FlashMLA kernels for Blackwell (SM100+) GPUs require CUTLASS v4.2.1+ APIs that are not available in the bundled v4.0.0
  • Temporarily disable these architectures until upstream upgrades CUTLASS

Things done

  • Built on platform:
    • x86_64-linux
    • aarch64-linux
    • x86_64-darwin
    • aarch64-darwin
  • Tested, as applicable:
  • Ran nixpkgs-review on this PR. See nixpkgs-review usage.
  • Tested basic functionality of all binary files, usually in ./result/bin/.
  • Nixpkgs Release Notes
    • Package update: when the change is major or breaking.
  • NixOS Release Notes
    • Module addition: when adding a new NixOS module.
    • Module update: when the change is significant.
  • Fits CONTRIBUTING.md, pkgs/README.md, maintainers/README.md and other READMEs.

Add a 👍 reaction to pull requests you find important.

@nixpkgs-ci nixpkgs-ci bot added 10.rebuild-darwin: 0 This PR does not cause any packages to rebuild on Darwin. 10.rebuild-linux: 0 This PR does not cause any packages to rebuild on Linux. 6.topic: python Python is a high-level, general-purpose programming language. labels Oct 5, 2025
@daniel-fahey daniel-fahey force-pushed the fix-vllm branch 2 times, most recently from cd48d25 to 2dad101 Compare October 5, 2025 22:54
Copy link
Contributor

@GaetanLepage GaetanLepage left a comment

Choose a reason for hiding this comment

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

Have you tried bumping cutlass to 4.2.1 instead?

@daniel-fahey
Copy link
Contributor Author

Have you tried bumping cutlass to 4.2.1 instead?

Not yet, definitely worth a go.

@daniel-fahey
Copy link
Contributor Author

daniel-fahey commented Oct 6, 2025

Attempt 0da1974

nix-build -I nixpkgs=https://github.com/NixOS/nixpkgs/archive/0da19742a9d0b7041c791cb043ffaaccffb6090d.tar.gz --arg config '{ allowUnfree = true; cudaSupport = true; }' -A python313Packages.vllm
this derivation will be built:
  /nix/store/f2bxd9hgy6gdhyqf3a6wyrii4kpy5zc1-python3.13-vllm-0.11.0.drv
building '/nix/store/f2bxd9hgy6gdhyqf3a6wyrii4kpy5zc1-python3.13-vllm-0.11.0.drv' on 'ssh://eu.nixbuild.net'...

failing

[nixbuild.net] Build 4586852 finished in 27 minutes with status 'build_failed'. Max memory usage was 73.0 of 128.5 GB. Average CPU utilization was 84% of 15 CPUs.

See this link for more details and logs.

@daniel-fahey
Copy link
Contributor Author

Earlier attempt cd48d25 also failed.

nix-build -I nixpkgs=https://github.com/NixOS/nixpkgs/archive/2dad10142fae482e3b7859118467bfbf38fbb4d5.tar.gz --arg config '{ allowUnfree = true; cudaSupport = true; }' -A python313Packages.vllm
[nixbuild.net] Build 4588139 finished in 38 minutes with status 'build_failed'. Max memory usage was 73.7 of 128.5 GB. Average CPU utilization was 84% of 15 CPUs.

See this link for more details and logs.

- vLLM v0.11.0's FlashMLA kernels for Blackwell (SM100+) GPUs require CUTLASS v4.2.1+ APIs that are not available in the bundled v4.0.0
- Temporarily disable these architectures until upstream upgrades CUTLASS
@daniel-fahey
Copy link
Contributor Author

daniel-fahey commented Oct 6, 2025

b1cef6c

nix-build --pure https://github.com/NixOS/nixpkgs/archive/b1cef6ca8d5dc66099a5f17f26724a1175a0ed72.tar.gz --arg config '{ allowUnfree = true; cudaSupport = true; }' --attr python313Packages.vllm
unpacking 'https://github.com/NixOS/nixpkgs/archive/b1cef6ca8d5dc66099a5f17f26724a1175a0ed72.tar.gz' into the Git cache...
this derivation will be built:
  /nix/store/1hqrs2hflnpppacgmwgghndqsk6f55mb-python3.13-vllm-0.11.0.drv
building '/nix/store/1hqrs2hflnpppacgmwgghndqsk6f55mb-python3.13-vllm-0.11.0.drv' on 'ssh://eu.nixbuild.net'...
copying 2 paths...
copying path '/nix/store/02mndfh9j65g9g3dg3y76yvfwik0539s-vllm-flash-attn-2.7.2.post1' to 'ssh://eu.nixbuild.net'...
copying path '/nix/store/a1vglvbl2lg951y0izm6bmzp8ggggxfq-flashmla-1.0.0' to 'ssh://eu.nixbuild.net'...
[nixbuild.net] Build 4590363 queued at 25-10-06 15:54:26Z. Allocated 15 CPUs, 128.5 GB memory.

Successfully built:

[nixbuild.net] Build 4590363 finished in an hour with status 'built'. Max memory usage was 84.8 of 128.5 GB. Average CPU utilization was 93% of 15 CPUs.[....]
copying 2 paths...
copying path '/nix/store/bnc6a9ds9hbcsaxdkrh1mwdnjrf24fb5-python3.13-vllm-0.11.0' from 'ssh://eu.nixbuild.net'...
copying path '/nix/store/djhlaai2s38dax2i85zasxffffjjgbn7-python3.13-vllm-0.11.0-dist' from 'ssh://eu.nixbuild.net'...
/nix/store/bnc6a9ds9hbcsaxdkrh1mwdnjrf24fb5-python3.13-vllm-0.11.0

See this link for more details and logs.

@daniel-fahey daniel-fahey changed the title python3Packages.vllm: fix CUDA build python3Packages.vllm: disable Blackwell GPU support to fix CUDA build Oct 6, 2025
@daniel-fahey daniel-fahey marked this pull request as ready for review October 6, 2025 18:54
@GaetanLepage
Copy link
Contributor

nixpkgs-review result

Generated using nixpkgs-review.

Command: nixpkgs-review pr 448965 --extra-nixpkgs-config '{ allowUnfree = true; cudaSupport = true; }'
Commit: b1cef6ca8d5dc66099a5f17f26724a1175a0ed72


x86_64-linux

✅ 13 packages built:
  • nixpkgs-manual
  • python312Packages.kserve
  • python312Packages.kserve.dist
  • python312Packages.torchrl
  • python312Packages.torchrl.dist
  • vllm (python312Packages.vllm)
  • vllm.dist (python312Packages.vllm.dist)
  • python313Packages.kserve
  • python313Packages.kserve.dist
  • python313Packages.torchrl
  • python313Packages.torchrl.dist
  • python313Packages.vllm
  • python313Packages.vllm.dist

Copy link
Contributor

@GaetanLepage GaetanLepage left a comment

Choose a reason for hiding this comment

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

Thanks

@GaetanLepage GaetanLepage added this pull request to the merge queue Oct 7, 2025
Merged via the queue into NixOS:master with commit 500193c Oct 7, 2025
33 of 35 checks passed
@daniel-fahey daniel-fahey deleted the fix-vllm branch October 17, 2025 13:38
@daniel-fahey daniel-fahey mentioned this pull request Oct 20, 2025
3 tasks
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

6.topic: python Python is a high-level, general-purpose programming language. 10.rebuild-darwin: 0 This PR does not cause any packages to rebuild on Darwin. 10.rebuild-linux: 0 This PR does not cause any packages to rebuild on Linux.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants