增加mtp>1的pad

vllm/v1/spec_decode/eagle.py

@@ -282,6 +282,10 @@ class EagleProposer:
                 seq_lens=seq_lens,
             )
 
+        #增加mtp>1的pad 
+        num_pad, num_tokens_across_dp = self.get_dp_padding(input_batch_size)
+        input_batch_size += num_pad
+                
         for i in range(self.num_speculative_tokens - 1):
             # Update the inputs.
             # cast to int32 is crucial when eagle model is compiled.

vllm/zero_overhead/v1/eagle.py

@@ -203,6 +203,10 @@ class V1ZeroEagleProposer(EagleProposer):
                 seq_lens=seq_lens,
             )
 
+        #增加mtp>1的pad 
+        num_pad, num_tokens_across_dp = self.get_dp_padding(input_batch_size)
+        input_batch_size += num_pad
+        
         for i in range(self.num_speculative_tokens - 1):
             # Update the inputs.
             # cast to int32 is crucial when eagle model is compiled.