diff --git a/python/sglang/jit_kernel/flash_attention.py b/python/sglang/jit_kernel/flash_attention.py index c6a67757d..22c7dc9bc 100644 --- a/python/sglang/jit_kernel/flash_attention.py +++ b/python/sglang/jit_kernel/flash_attention.py @@ -230,6 +230,7 @@ def flash_attn_varlen_func( softcap=0.0, num_splits=1, pack_gqa=None, + only_qv=False, sm_margin=0, return_softmax_lse=False, sinks=None, @@ -262,6 +263,7 @@ def flash_attn_varlen_func( softcap=softcap, num_splits=num_splits, pack_gqa=pack_gqa, + only_qv=only_qv, sm_margin=sm_margin, return_softmax_lse=return_softmax_lse, sinks=sinks,