From fc809665fd7c292b472d446829cb051d56378bd5 Mon Sep 17 00:00:00 2001 From: Jinyan Chen <93358689+liz-badada@users.noreply.github.com> Date: Fri, 19 Sep 2025 02:42:49 +0800 Subject: [PATCH] [Performance] qwen3-next improve causal conv1d in prefill phase (#10595) Co-authored-by: Jinyan Chen --- .../srt/layers/attention/hybrid_linear_attn_backend.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py b/python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py index 82baea216..0d43a2f8f 100644 --- a/python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py +++ b/python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py @@ -13,6 +13,9 @@ from sglang.srt.layers.attention.fla.fused_recurrent import ( from sglang.srt.layers.attention.fla.fused_sigmoid_gating_recurrent import ( fused_sigmoid_gating_delta_rule_update, ) +from sglang.srt.layers.attention.mamba.causal_conv1d import ( + causal_conv1d_fn as causal_conv1d_fn_sgl, +) from sglang.srt.layers.attention.mamba.causal_conv1d_triton import ( causal_conv1d_fn, causal_conv1d_update, @@ -334,7 +337,7 @@ class MambaAttnBackend(AttentionBackend): mixed_qkv_processed.transpose(1, 2).contiguous().view(seq_len, -1) ) else: - mixed_qkv = causal_conv1d_fn( + mixed_qkv = causal_conv1d_fn_sgl( mixed_qkv.transpose(0, 1), conv_weights, bias,