Commit b1cef6ca authored by Daniel Fahey's avatar Daniel Fahey
Browse files

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

- 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
parent 8ea360ac

pkgs/development/python-modules/vllm/default.nix

+13 −10
Original line number Diff line number Diff line
@@ -200,16 +200,19 @@ let 
        "8.9"

        "9.0"

        "9.0a"

        "10.0"

        "10.0a"

        "10.1"

        "10.1a"

        "10.3"

        "10.3a"

        "12.0"

        "12.0a"

        "12.1"

        "12.1a"

        # Blackwell (SM100+) capabilities temporarily disabled due to CUTLASS API incompatibility

        # FlashMLA kernels require CUTLASS v4.2.1+ APIs not available in bundled v4.0.0

        # TODO: Re-enable when vLLM upgrades CUTLASS (see https://github.com/vllm-project/vllm/pull/24673)

        # "10.0"

        # "10.0a"

        # "10.1"

        # "10.1a"

        # "10.3"

        # "10.3a"

        # "12.0"

        # "12.0a"

        # "12.1"

        # "12.1a"

      ];

      ptx = lists.map (x: "${x}+PTX") real;

    in