with pytorch cpu only version. without --no_cuda, using --bf16 will trigger...

with pytorch cpu only version. without --no_cuda, using --bf16 will trigger error like "Your setup doesn't support bf16/gpu. You need torch>=1.10, using Ampere GPU with cuda>=11.0" (#20445)

src/transformers/training_args.py

@@ -1086,7 +1086,7 @@ class TrainingArguments:
                 if self.no_cuda and not is_torch_bf16_cpu_available():
                     # cpu
                     raise ValueError("Your setup doesn't support bf16/cpu. You need torch>=1.10")
-                elif not self.no_cuda and not is_torch_bf16_gpu_available():
+                elif not self.no_cuda and torch.cuda.is_available() and not is_torch_bf16_gpu_available():
                     # gpu
                     raise ValueError(
                         "Your setup doesn't support bf16/gpu. You need torch>=1.10, using Ampere GPU with cuda>=11.0"