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

[Intel GPU] DeepSeek V4 4/N: use sgl-kernel implementation of fused_q_norm_rope on XPU (#27790)

Browse Source 
Signed-off-by: P V R K Jyothendra Varma <polisettyvarma@gmail.com>
This commit is contained in:
Polisetty V R K Jyothendra Varma
2026-08-05 16:04:35 -07:00
committed by GitHub
parent c0ef548eef
commit a1cc286062
1 changed files with 6 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/kernels/ops/attention/dsv4/elementwise.py
+6 -2
View File
@@ -17,6 +17,7 @@ _is_xpu = is_xpu()
if _is_xpu: if _is_xpu:
from sgl_kernel import fused_k_norm_rope_flashmla as fused_k_norm_rope_flashmla_xpu from sgl_kernel import fused_k_norm_rope_flashmla as fused_k_norm_rope_flashmla_xpu
from sgl_kernel import fused_q_norm_rope as fused_q_norm_rope_xpu
@cache_once @cache_once
@@ -150,8 +151,11 @@ def fused_q_norm_rope(
freqs_real = torch.view_as_real(freqs_cis).flatten(-2) freqs_real = torch.view_as_real(freqs_cis).flatten(-2)
head_dim = q_input.shape[-1] head_dim = q_input.shape[-1]
rope_dim = freqs_real.shape[-1] rope_dim = freqs_real.shape[-1]
module = _jit_main_q_norm_rope_module(q_input.dtype, head_dim, rope_dim) if _is_xpu:
module.forward(q_input, q_output, freqs_real, positions, eps) fused_q_norm_rope_xpu(q_input, q_output, freqs_real, positions, eps)
else:
module = _jit_main_q_norm_rope_module(q_input.dtype, head_dim, rope_dim)
module.forward(q_input, q_output, freqs_real, positions, eps)
def fused_q_indexer_rope_hadamard_quant( def fused_q_indexer_rope_hadamard_quant(