chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[NPU] bug fix for mtp and w4a8 (#14806)

Browse Source
This commit is contained in:
liupeng374
2025-12-10 22:55:13 +08:00
committed by GitHub
parent d85fecb536
commit b8cfa02c01
3 changed files with 19 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
python/sglang/srt/hardware_backend/npu/quantization/fused_moe_method_npu.py
View File

@@ -567,11 +567,13 @@ class NPUW4A8Int4DynamicMoEMethod(FusedMoEMethodBase):
group_list,
output_dtype,
):
from sgl_kernel_npu.activation.swiglu_quant import swiglu_quant
hidden_states = torch.ops.npu.npu_grouped_matmul(
x=[hidden_states],
weight=[self.w13_weight],
scale=[self.w13_weight_scale],
bias=[self.w13_scale_bias],
weight=[layer.w13_weight],
scale=[layer.w13_weight_scale],
bias=[layer.w13_scale_bias],
per_token_scale=[hidden_states_scale],
group_list=group_list,
split_item=2,
@@ -580,15 +582,15 @@ class NPUW4A8Int4DynamicMoEMethod(FusedMoEMethodBase):
output_dtype=output_dtype,
)[0]
# act_fn: swiglu
hidden_states = torch.ops.npu.npu_swiglu(hidden_states)
hidden_states, swiglu_out_scale = torch.ops.npu.npu_dynamic_quant(hidden_states)
hidden_states, swiglu_out_scale = swiglu_quant(
hidden_states, group_list, group_list_type
)
hidden_states = torch.ops.npu.npu_grouped_matmul(
x=[hidden_states],
weight=[self.w2_weight],
scale=[self.w2_weight_scale],
bias=[self.w2_scale_bias],
weight=[layer.w2_weight],
scale=[layer.w2_weight_scale],
bias=[layer.w2_scale_bias],
per_token_scale=[swiglu_out_scale],
group_list=group_list,
split_item=2,