EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Performance] qwen3-next improve causal conv1d in prefill phase (#10595)

Browse Source 
Co-authored-by: Jinyan Chen <jinyanc@nvidia.com>
This commit is contained in:
Jinyan Chen
2025-09-19 02:42:49 +08:00
committed by GitHub
parent 6fd4816d9f
commit fc809665fd
1 changed files with 4 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py
View File

@@ -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,