diff options
| author | Daniel Fahey <daniel.fahey@pm.me> | 2025-10-24 21:29:20 +0100 |
|---|---|---|
| committer | Daniel Fahey <daniel.fahey@pm.me> | 2025-10-25 14:57:22 +0100 |
| commit | 59d8696fe6a33d833c7873f11e9cf7e2f45f3d29 (patch) | |
| tree | 038fa623f99c954f809ea37586bb775e0597f6f2 /pkgs/development/python-modules/python-openstackclient | |
| parent | 27fd57665d579734a3e494cbf80426c21e816d56 (diff) | |
python3Packages.vllm: fix Blackwell support
This corrects the approach taken in b1cef6ca8d5dc66099a5f17f26724a1175a0ed72, which disabled Blackwell (SM100+) support due to CUTLASS API incompatibility.
My original diagnosis was incorrect: FlashMLA's SM100 kernels don't need newer CUTLASS APIs; they actually need CUTLASS v3.9.0 (pinned in FlashMLA's git submodule). The incompatibility is with vLLM's FlashMLA fork and CUTLASS v4.x APIs altogether.
This commit packages both CUTLASS versions: v3.9.0 for FlashMLA and v4.0.0 for the main vLLM build, allowing Blackwell support to work correctly. Although this subtlety can easily be expressed in the Nix derivation it was reported upstream in https://github.com/vllm-project/vllm/issues/27425
Diffstat (limited to 'pkgs/development/python-modules/python-openstackclient')
0 files changed, 0 insertions, 0 deletions