add bfloat16 to inference support dtypes

to allow running inference tasks using bfloat16

deepspeed/inference/engine.py

@@ -324,7 +324,7 @@ def _validate_args(self, mpu, replace_with_kernel_inject):
         if self._config.checkpoint is not None and not isinstance(self._config.checkpoint, (str, dict)):
             raise ValueError(f"checkpoint must be None, str or dict, got {type(self._config.checkpoint)}")
 
-        supported_dtypes = [None, torch.half, torch.int8, torch.float]
+        supported_dtypes = [None, torch.half, torch.int8, torch.float, torch.bfloat16]
         if self._config.dtype not in supported_dtypes:
             raise ValueError(f"{self._config.dtype} not supported, valid dtype: {supported_dtypes}")