[BugFix] Fix import error on non-blackwell machines

[LucasWilkinson]

FIX #20769 (comment)

Checked vllm serve runs when built on hopper
Checked can run VLLM_ATTENTION_BACKEND=CUTLASS_MLA_VLLM_V1 lm_eval --model vllm --model_args pretrained=deepseek-ai/DeepSeek-V2-Lite-Chat,trust_remote_code=true --tasks gsm 8k --batch_size auto

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[gemini-code-assist]

Code Review

This pull request fixes an import error on non-Blackwell machines by moving the operator implementation registration into a conditionally compiled file. The dispatch key for the workspace calculation function has been corrected to improve the robustness of the solution.

[gemini-code-assist]

The function sm100_cutlass_mla_get_workspace_size is a host-side function that calculates a workspace size and does not involve any GPU operations. Registering it only for the CUDA dispatch key is incorrect and can lead to runtime errors if called in a context where the PyTorch dispatcher selects a different backend. Host-only functions like this should be registered for the CPU dispatch key to ensure they can be called correctly regardless of the context.

TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
  m.impl("sm100_cutlass_mla_decode", &sm100_cutlass_mla_decode);
}

TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CPU, m) {
  m.impl("sm100_cutlass_mla_get_workspace_size", &sm100_cutlass_mla_get_workspace_size);
}

[LucasWilkinson]

GTG:

Checked vllm serve runs when built on hopper
Checked can run VLLM_ATTENTION_BACKEND=CUTLASS_MLA_VLLM_V1 lm_eval --model vllm --model_args pretrained=deepseek-ai/DeepSeek-V2-Lite-Chat,trust_remote_code=true --tasks gsm 8k --batch_size auto on Blackwell

[jeejeelee]

I have tested this PR locally, and it can fix #20769 (comment), thank you