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

Fix the global scale fix does not support EPLB and improve enabling condition (#10369)

Browse Source
This commit is contained in:
fzyzcjy
2025-09-14 16:07:47 +08:00
committed by GitHub
parent abea9250da
commit 2df532ef20
2 changed files with 8 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/layers/quantization/modelopt_quant.py
View File

@@ -999,12 +999,14 @@ class ModelOptNvFp4FusedMoEMethod(FusedMoEMethodBase):
data=torch.empty(layer.num_experts, 2, dtype=torch.float32),
weight_loader=weight_loader,
)
w13_input_scale._sglang_require_global_experts = True
layer.register_parameter("w13_input_scale", w13_input_scale)
w2_input_scale = PerTensorScaleParameter(
data=torch.empty(layer.num_experts, dtype=torch.float32),
weight_loader=weight_loader,
)
w2_input_scale._sglang_require_global_experts = True
layer.register_parameter("w2_input_scale", w2_input_scale)
def swizzle_blockscale(self, scale: torch.Tensor):