Merge branch 'v0.7.2-fusion-wm' into 'v0.7.2-fusion'

[fix]修复cudagraph和eager分段模式下，开启mla后报错问题

See merge request dcutoolkit/deeplearing/vllm!109

vllm/worker/model_runner.py

@@ -892,6 +892,10 @@ 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.
@@ -1709,7 +1713,7 @@ class ModelRunner(GPUModelRunnerBase[ModelInputForGPUWithSamplingMetadata]):
         # virtual engines share the same kv cache.
         virtual_engine = model_input.virtual_engine
         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][