[fix]修复cudagraph和eager分段模式开启mla后报错问题

vllm/worker/model_runner.py

@@ -888,6 +888,10 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
             max_encoder_seq_len=max_encoder_seq_len)
 
         batch_size = len(input_tokens)
+        
+        if batch_size + cuda_graph_pad_size >= self.runner.enforce_eager_bs_threshould:
+            cuda_graph_pad_size = -1
+
         if cuda_graph_pad_size != -1:
             # If cuda graph can be used, pad tensors accordingly.
             # See `capture_model` API for more details.
@@ -1717,7 +1721,7 @@ class ModelRunner(GPUModelRunnerBase[ModelInputForGPUWithSamplingMetadata]):
         virtual_engine = model_input.virtual_engine
         previous_hidden_states = kwargs.get("previous_hidden_states")
         if prefill_meta is None and decode_meta.use_cuda_graph and \
-                model_input.input_tokens.shape[0] <= self.enforce_eager_bs_threshould:
+                model_input.input_tokens.shape[0] < self.enforce_eager_bs_threshould:
             assert model_input.input_tokens is not None
             graph_batch_size = model_input.input_tokens.shape[0]
             model_executable = self.graph_runners[virtual_engine][