[fix]解决v1 engine mtp只能设为1的问题

vllm/v1/spec_decode/eagle.py

@@ -196,7 +196,12 @@ class EagleProposer:
         draft_token_ids_list = [draft_token_ids]
 
         positions = target_positions[last_token_indices]
+
+        if self.method == "deepseek_mtp":
+            hidden_states = last_hidden_states[last_token_indices]
+        else:
             hidden_states = hidden_states[last_token_indices]
+
         if self.use_cuda_graph and \
             batch_size <= self.cudagraph_batch_sizes[-1]:
             input_batch_size = self.vllm_config.pad_for_cudagraph(batch_size)
@@ -256,12 +261,18 @@ class EagleProposer:
             with set_forward_context(per_layer_attn_metadata,
                                      self.vllm_config,
                                      num_tokens=input_batch_size):
-                last_hidden_states, hidden_states = self.model(
+                ret_hidden_states = self.model(
                     self.input_ids[:input_batch_size],
                     self.positions[:input_batch_size],
                     self.hidden_states[:input_batch_size],
                 )
+                if self.method == "deepseek_mtp":
+                    last_hidden_states = ret_hidden_states
+                    hidden_states = last_hidden_states[:batch_size]
+                else:
+                    last_hidden_states, hidden_states = ret_hidden_states
                     hidden_states = hidden_states[:batch_size]
+
             logits = self.model.compute_logits(last_hidden_states[:batch_size],
                                                None)