diff --git a/python/sglang/srt/layers/attention/triton_ops/rocm_mla_decode_rope.py b/python/sglang/srt/layers/attention/triton_ops/rocm_mla_decode_rope.py
index 4c3e6396828..915f4ef9276 100644
--- a/python/sglang/srt/layers/attention/triton_ops/rocm_mla_decode_rope.py
+++ b/python/sglang/srt/layers/attention/triton_ops/rocm_mla_decode_rope.py
@@ -20,19 +20,12 @@
 # https://github.com/ModelTC/lightllm/blob/96353e868a840db4d103138caf15ed9dbea8c186/lightllm/models/deepseek2/triton_kernel/gqa_flash_decoding_stage1.py
 # https://github.com/ModelTC/lightllm/blob/96353e868a840db4d103138caf15ed9dbea8c186/lightllm/models/deepseek2/triton_kernel/gqa_flash_decoding_stage2.py
 
-import argparse
-import logging
-import sys
-
-import pytest
-import torch
 import triton
 import triton.language as tl
 
 from sglang.srt.layers.attention.triton_ops.decode_attention import (
     _decode_softmax_reducev_fwd,
 )
-from sglang.srt.layers.rotary_embedding import DeepseekScalingRotaryEmbedding
 
 
 def is_hip():