minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[XPU] Enable fused GDN QKV split Triton kernel on XPU (#30144)

Browse Source 
Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>
This commit is contained in:
jiayisunx
2026-08-18 10:42:58 +08:00
committed by GitHub
co-authored by Ma Mingfei
parent 7f51e6bba0
commit d6c837489a
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/linear/gdn_backend.py
+2 -2
View File
@@ -29,7 +29,7 @@ if not is_cpu():
CHUNK_SIZE as FLA_CHUNK_SIZE, CHUNK_SIZE as FLA_CHUNK_SIZE,
) )
if is_cuda() or is_hip(): if is_cuda() or is_hip() or is_xpu():
from sglang.kernels.ops.attention.triton_gdn_fused_proj import ( from sglang.kernels.ops.attention.triton_gdn_fused_proj import (
fused_qkv_split_gdn_prefill, fused_qkv_split_gdn_prefill,
) )
@@ -595,7 +595,7 @@ class GDNAttnBackend(MambaAttnBackendBase):
actual_seq_len = mixed_qkv.shape[0] actual_seq_len = mixed_qkv.shape[0]
qkv_dim = layer.q_dim + layer.k_dim + layer.v_dim qkv_dim = layer.q_dim + layer.k_dim + layer.v_dim
if (is_cuda() or is_hip()) and qkv_dim <= MAX_FUSED_QKV_SPLIT_DIM: if (is_cuda() or is_hip() or is_xpu()) and qkv_dim <= MAX_FUSED_QKV_SPLIT_DIM:
query, key, value = fused_qkv_split_gdn_prefill( query, key, value = fused_qkv_split_gdn_prefill(
mixed_qkv, mixed_qkv,
layer.num_q_heads, layer.num_q_heads,