diff --git a/python/sglang/srt/layers/attention/wave_ops/decode_attention.py b/python/sglang/srt/layers/attention/wave_ops/decode_attention.py index aadcb3079991..cf804780f63d 100644 --- a/python/sglang/srt/layers/attention/wave_ops/decode_attention.py +++ b/python/sglang/srt/layers/attention/wave_ops/decode_attention.py @@ -78,7 +78,6 @@ def get_wave_kernel( max_kv_splits, input_dtype=input_dtype, output_dtype=output_dtype, - mha=mha, logit_cap=logit_cap, ) hyperparams_0.update(get_default_scheduling_params())