diff --git a/python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py b/python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py
index b1935d21c462..6f0ba95a44ff 100644
--- a/python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py
+++ b/python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py
@@ -609,7 +609,7 @@ def replay_prepare(
             buffers.input_ids[num_tokens:static_num_tokens].zero_()
             buffers.positions[num_tokens:static_num_tokens].zero_()
             if self.is_multimodal:
-                buffers.input_embeds[:, num_tokens:static_num_tokens].zero_()
+                buffers.input_embeds[num_tokens:static_num_tokens].zero_()
             if forward_batch.mrope_positions is not None:
                 buffers.mrope_positions[:, num_tokens:static_num_tokens].zero_()