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

use fp32 for e_score_correction_bias in GLM-4.5 (#8729)

Browse Source
This commit is contained in:
Yuxuan Zhang
2025-08-04 01:43:40 +08:00
committed by GitHub
parent 3435a24e81
commit 760286e3d3
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/models/glm4_moe.py
View File

@@ -343,7 +343,7 @@ class Glm4MoeGate(nn.Module):
torch.empty((config.n_routed_experts, config.hidden_size)) torch.empty((config.n_routed_experts, config.hidden_size))
) )
self.e_score_correction_bias = nn.Parameter( self.e_score_correction_bias = nn.Parameter(
torch.empty((config.n_routed_experts)) torch.empty((config.n_routed_experts), dtype=torch.float32)
) )
if _is_cpu and _is_cpu_amx_available: if _is_cpu and _is_cpu_amx_available:
self.quant_method = PackWeightMethod(weight_names=["weight"]) self.quant_method = PackWeightMethod(weight_names=["weight"])