零消耗添加epsp

vllm/zero_overhead/v1/gpu_model_runner.py

@@ -465,7 +465,7 @@ class V1ZeroModelRunner(GPUModelRunner):
         num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
         
         # make sure that the padded length is divisible by attn_tp_size because we may need reduce-scatter across attn_tp dim.
-        if self.ep_sp:
+        if self.ep_sp or self.enable_dp_attention:
             num_input_tokens = round_up(num_scheduled_tokens, tp_size)
             if (self.use_cuda_graph
                     and num_input_tokens <= self.cudagraph_batch_sizes[-1]):