fix: set fp32 when using cpu bc bf16 is slow

configuration_xlm_roberta.py

@@ -126,3 +126,5 @@ class XLMRobertaFlashConfig(PretrainedConfig):
             self.torch_dtype = getattr(torch, torch_dtype)
         else:
             self.torch_dtype = torch_dtype
+        if not self.use_flash_attn or not torch.cuda.is_available():
+            self.torch_dtype = torch.float32