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

[NPU] replace swiglu with custom kernel

Browse Source
This commit is contained in:
JiaruiChang5268
2026-03-10 21:08:37 +08:00
committed by GitHub
parent 9884957c07
commit 5a7c1b8ec6
1 changed files with 5 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/hardware_backend/npu/quantization/fused_moe_method_npu.py
View File

@@ -115,6 +115,8 @@ def npu_fused_experts(
def npu_fused_moe_without_routing_weights_bf16(
layer, hidden_states, group_list_type, group_list, output_dtype
):
from sgl_kernel_npu.activation.swiglu_quant import swiglu_quant
# gmm1: gate_up_proj
hidden_states = torch.ops.npu.npu_grouped_matmul(
x=[hidden_states],
@@ -125,7 +127,9 @@ def npu_fused_moe_without_routing_weights_bf16(
group_list=group_list,
output_dtype=output_dtype,
)[0]
hidden_states = torch.ops.npu.npu_swiglu(hidden_states)
hidden_states, _ = swiglu_quant(
hidden_states, group_list, group_list_type, need_quant=False
)
# gmm2: down_proj
hidden_states = torch.ops.npu.npu_grouped_matmul(
x=[hidden_states],