fix bfloat mismatch bug when model loading using half()

add dtype changes automatic when extract image embedding

modeling_internlm_xcomposer2.py

@@ -123,7 +123,7 @@ class InternLMXComposer2ForCausalLM(InternLM2PreTrainedModel):
             print ('Batch Size >1 is not supported.')
             assert 0
         #print (img_embeds.shape)
-        img_embeds = self.vision_proj(img_embeds)
+        img_embeds = self.vision_proj(img_embeds.to(self.dtype))
         atts_img = torch.ones(
             img_embeds.size()[:-1], dtype=torch.long).to(img_embeds.device)