[Installation]: DeepGEMM will not compile: unsupported CUDA architecture 12.0a

[sscckk]

Your current environment

The output of `python collect_env.py`

How you are installing vllm

pip install -vvv vllm

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[sscckk]

Hello.
When I install vLLM, using the sm120 branch of DeepGEMM you provided, it still doesn't get compiled. Is this normal? Will it affect performance?
Thank you!

[jasl]

Thanks. For the current SM12x branch this is expected to be non-fatal for the primary DeepSeek-V4-Flash FP8 path.

DeepGEMM may still fail to compile for 12.0a / SM12x depending on the DeepGEMM and CUDA revision. The current branch does not require a successful DeepGEMM SM120 build for the standard DeepSeek-V4-Flash + FP8 KV-cache path: the DSv4 interfaces that used to be DeepGEMM-only have SM12x fallback paths, and the optimized MHC prenorm path no longer depends on DeepGEMM.

So the short answer is:

• If you are serving DeepSeek-V4-Flash with FP8 KV cache, this should not block startup and should not affect the main tested path.
• It can still affect experimental DeepGEMM-backed / MXFP4 paths, or FP4 indexer-cache paths. Those still need a compatible DeepGEMM build.
• Please make sure you are installing from this branch, not plain PyPI pip install vllm, and try the latest branch head (e585e34df or newer).

If the vLLM install or DeepSeek-V4-Flash FP8 serve path still fails after that, please paste the full build/serve traceback. A useful quick isolation check is to try with VLLM_USE_DEEP_GEMM=0; if that starts, then the remaining issue is isolated to the optional DeepGEMM-backed path rather than the base SM12x DS4 path.