microsoft / BitBLAS

BitBLAS is a library to support mixed-precision matrix multiplications, especially for quantized LLM deployment.
MIT License
190 stars 21 forks source link

[FIX] Must validate ENV settings or wrong gpu selected by nvidia-smi #59

Closed Qubitium closed 1 week ago

Qubitium commented 1 week ago

nvidia-smi uses PCI_BUS_ID order but python program may be launched using default which is not PCI_BUS_ID order for gpu. If the env values do not match, wrong gpu is returned for gpu_id. Validate the env and raise error if issue exists.

TESTS

Qubitium commented 1 week ago
@microsoft-github-policy-service agree

@microsoft-github-policy-service agree

Qubitium commented 1 week ago

@LeiWang1999 Ready for review. The CUDA order ENV must be validated (match nvidia-smi) in multi-gpu env or we get the wrong gpu back.