Skip to content

Conversation

kushanam
Copy link

This PR adds Blackwell archs and updates the CUTLASS version.

@kushanam kushanam changed the title adding priimary Blackwell support adding preliminary Blackwell support Feb 25, 2025
@LucasWilkinson
Copy link
Collaborator

Can we add a guard to make sure we don't try to build these arches on less than CUDA 12.8 something like:
https://github.com/vllm-project/vllm/blob/b382a7f28f739f3b120e5495fd029089d0399428/CMakeLists.txt#L382

@LucasWilkinson
Copy link
Collaborator

Probably not for this PR but we may want to compile the FA3 SM80 kernels for Blackwell https://github.com/vllm-project/flash-attention/blob/main/hopper/flash_fwd_kernel_sm80.h

@kushanam
Copy link
Author

kushanam commented Mar 4, 2025

@LucasWilkinson added the guard, and thanks for pointing out FAv3. I will add it in a separate PR

@LucasWilkinson
Copy link
Collaborator

@LucasWilkinson added the guard, and thanks for pointing out FAv3. I will add it in a separate PR

Thanks, LGTM now 👍

@LucasWilkinson LucasWilkinson merged commit 6f27a05 into vllm-project:main Mar 4, 2025
1 check failed
LucasWilkinson added a commit that referenced this pull request Mar 5, 2025
* adding priimary Blackwell support

* guarding blackwell build against cuda < 12.8

Signed-off-by: Lucas Wilkinson <[email protected]>
LucasWilkinson added a commit that referenced this pull request Mar 5, 2025
* adding priimary Blackwell support

* guarding blackwell build against cuda < 12.8

Signed-off-by: Lucas Wilkinson <[email protected]>
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.

2 participants