Revert "[feat]优化mtp相关函数返回类型"

This reverts commit edc45625.

vllm/v1/spec_decode/eagle.py

@@ -98,7 +98,7 @@ class EagleProposer:
         next_token_ids: torch.Tensor,
         common_attn_metadata: CommonAttentionMetadata,
         sampling_metadata: SamplingMetadata,
-    ) -> tuple[torch.Tensor, torch.Tensor]:
+    ) -> torch.Tensor:
         num_tokens = target_token_ids.shape[0]
         batch_size = next_token_ids.shape[0]
         last_token_indices = common_attn_metadata.query_start_loc[1:] - 1
@@ -194,7 +194,7 @@ class EagleProposer:
         # Early exit if there is only one draft token to be generated.
         if self.num_speculative_tokens == 1:
             # [batch_size, 1]
-            return draft_token_ids.view(-1, 1), draft_prob.view(-1, 1, draft_prob.shape[-1])
+            return draft_token_ids.view(-1, 1), draft_probs_list
 
         # TODO: Currently, MTP module released by deepseek only has
         # one layer. Adapt this code to support multiple layers once

vllm/v1/worker/gpu_model_runner.py

@@ -1687,8 +1687,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         aux_hidden_states: Optional[torch.Tensor],
         spec_decode_metadata: Optional[SpecDecodeMetadata],
         common_attn_metadata: CommonAttentionMetadata,
-    ) -> tuple[list[list[int]], torch.Tensor]:
-        draft_probs = None
+    ) -> list[list[int]]:
         num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
         if self.speculative_config.method == "ngram":
             assert isinstance(self.drafter, NgramProposer)