add recive mtp tokenids and req_ids in pp rank 0 from pp rank 1

vllm/zero_overhead/v1/PP2mtp/gpu_model_runner.py

@@ -401,6 +401,13 @@ class V1ZeroModelRunner(GPUModelRunner):
 
             return self.kv_connector_no_forward(scheduler_output)
 
+
+        if get_pp_group().is_first_rank and self.last_draft_token_ids != None:
+            recv_draft_dict = get_pp_group().recv_tensor_dict(
+                    all_gather_group=get_tp_group())
+            self.last_draft_token_ids = recv_draft_dict['draft_token_ids']
+            self.last_sampled_req_ids = recv_draft_dict['sampled_req_ids ']
+
         # Prepare the decoder inputs.
         (attn_metadata, attention_cuda_graphs, logits_indices,
          spec_decode_metadata,