[Bugfix] RoBERTa position_id accumulation in CUDA graph padding region (vllm-project#37873)

Signed-off-by: dass90 <3053034939@qq.com>

Author: yanghui1-arch

vllm/v1/worker/gpu_model_runner.py

@@ -3084,6 +3084,8 @@ def _preprocess(
             positions = self.xdrope_positions.gpu[:, :num_input_tokens]
         else:
             positions = self.positions.gpu[:num_input_tokens]
+            if num_input_tokens > num_scheduled_tokens:
+                self.positions.gpu[num_scheduled_tokens:num_input_tokens].zero_()
 
         if is_first_rank:
             intermediate_tensors = None