[fix]解决EP开启pd_padding大输入卡住问题

vllm/v1/spec_decode/eagle.py

@@ -226,8 +226,8 @@ class EagleProposer:
 
         with set_forward_context(per_layer_attn_metadata,
                                  self.vllm_config,
-                                 num_tokens=num_input_tokens,
-                                 skip_cuda_graphs=not decoding):
+                                 num_tokens=num_input_tokens):
+                                 #skip_cuda_graphs=not decoding):
             ret_hidden_states = self.model(
                 self.input_ids[:num_input_tokens],
                 self.positions[:num_input_tokens],

vllm/zero_overhead/v1/eagle.py

@@ -150,8 +150,8 @@ class V1ZeroEagleProposer(EagleProposer):
 
         with set_forward_context(per_layer_attn_metadata,
                                  self.vllm_config,
-                                 num_tokens=num_input_tokens,
-                                 skip_cuda_graphs=not decoding):
+                                 num_tokens=num_input_tokens,):
+                                 #skip_cuda_graphs=not decoding):
             ret_hidden_states = self.model(
                 self.input_ids[:num_input_tokens],
                 self.positions[:num_input_tokens],