fix qwen_vl and load_pretrained patch (#2190)

Author: lvyufeng

mindnlp/transformers/modeling_utils.py

@@ -184,8 +184,8 @@ def wrapper(
         pretrained_model_name_or_path,
         **kwargs,
     ):
-        device_map = kwargs.pop("device_map", None)
-        sharded_metadata = kwargs.pop("sharded_metadata", None)
+        device_map = kwargs.get("device_map", None)
+        sharded_metadata = kwargs.get("sharded_metadata", None)
 
         # if device_map is not None and not initialize distribute module, raise Error.
         if device_map is not None:

mindtorch/nn/functional.py

@@ -1220,7 +1220,7 @@ def scaled_dot_product_attention(query, key, value, attn_mask=None, dropout_p=0.
 
     attn_weight = query @ key.transpose(-2, -1) * scale_factor
     attn_weight += attn_bias
-    attn_weight = softmax(attn_weight, dim=-1, dtype=mindtorch.float32).to(query.dtype)
+    attn_weight = softmax(attn_weight, dim=-1)
     attn_weight = dropout(attn_weight, dropout_p, training=True)
     return attn_weight @ value