Use dynamic dtype for prompts

modeling_chatglm.py

@@ -804,9 +804,9 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
     def set_input_embeddings(self, new_embeddings: torch.Tensor):
         self.word_embeddings = new_embeddings
 
-    def get_prompt(self, batch_size, device):
+    def get_prompt(self, batch_size, device, dtype=torch.half):
         prefix_tokens = self.prefix_tokens.unsqueeze(0).expand(batch_size, -1).to(device)
-        past_key_values = self.prefix_encoder(prefix_tokens).half()
+        past_key_values = self.prefix_encoder(prefix_tokens).type(dtype)
         past_key_values = past_key_values.view(
             batch_size,
             self.pre_seq_len,
@@ -896,9 +896,13 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
         else:
             raise ValueError("You have to specify either input_ids or inputs_embeds")
 
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+
         if past_key_values is None:
             if self.pre_seq_len is not None:
-                past_key_values = self.get_prompt(batch_size=input_ids.shape[0], device=input_ids.device)
+                past_key_values = self.get_prompt(batch_size=input_ids.shape[0], device=input_ids.device,
+                                                  dtype=inputs_embeds.dtype)
             else:
                 past_key_values = tuple([None] * len(self.layers))
 
@@ -927,8 +931,6 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
                     gmask=use_gmask
                 )
 
-        if inputs_embeds is None:
-            inputs_embeds = self.word_embeddings(input_ids)
 
         # [seq_len, batch, hidden_size]
         hidden_states = inputs_embeds.transpose(0, 1)