Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-18 16:54:33 +02:00
Code Issues Packages Projects Releases Wiki Activity

SDNQ fix asym quant formula for dtypes with non zero minimums

Browse Source
This commit is contained in:
Disty0
2025-05-28 17:25:38 +03:00
parent e06cbea7aa
commit dd0dbc476f
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/model_quant_sdnq.py
+2 -2
View File
@@ -187,11 +187,11 @@ def apply_sdnq_to_module(model, weights_dtype="int8", torch_dtype=None, group_si
def get_scale_asymmetric(weight: torch.FloatTensor, reduction_axes: List[int], weights_dtype: str) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
zero_point = torch.amin(weight, dim=reduction_axes, keepdims=True)
scale = torch.amax(weight, dim=reduction_axes, keepdims=True).sub_(zero_point).div_(dtype_dict[weights_dtype]["max"])
scale = torch.amax(weight, dim=reduction_axes, keepdims=True).sub_(zero_point).div_(dtype_dict[weights_dtype]["max"] - dtype_dict[weights_dtype]["min"])
eps = torch.finfo(scale.dtype).eps # prevent divison by 0
scale = torch.where(torch.abs(scale) < eps, eps, scale)
if dtype_dict[weights_dtype]["min"] != 0:
zero_point.add_(dtype_dict[weights_dtype]["min"])
zero_point.sub_(torch.mul(scale, dtype_dict[weights_dtype]["min"]))
return scale, zero_point