diff --git a/vllm/platforms/cuda.py b/vllm/platforms/cuda.py index 42fff2f07a0a..8e3436a9e73c 100644 --- a/vllm/platforms/cuda.py +++ b/vllm/platforms/cuda.py @@ -64,8 +64,7 @@ def supported_dtypes(self) -> list[torch.dtype]: if self.has_device_capability(80): # Ampere and Hopper or later NVIDIA GPUs. return [torch.bfloat16, torch.float16, torch.float32] - elif (not self.has_device_capability(80) - ) and self.has_device_capability(60): + if self.has_device_capability(60): # Pascal, Volta and Turing NVIDIA GPUs, BF16 is not supported return [torch.float16, torch.float32] # Kepler and Maxwell NVIDIA GPUs, only FP32 is supported,