Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-14 02:28:43 +02:00
Code Issues Packages Projects Releases Wiki Activity

SDNQ remove use_contiguous_mm path in re_quant

Browse Source
This commit is contained in:
Disty0
2025-10-04 19:17:05 +03:00
parent df142afe81
commit f2e12a682f
1 changed files with 2 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sdnq/dequantizer.py
+2 -6
View File
@@ -4,7 +4,7 @@ from typing import Tuple
import torch
from .common import dtype_dict, compile_func, use_contiguous_mm
from .common import dtype_dict, compile_func
from .packed_int import pack_int_symetric, unpack_int_symetric, pack_int_asymetric, unpack_int_asymetric
@@ -51,11 +51,7 @@ def quantize_fp8(input: torch.FloatTensor, dim: int = -1) -> Tuple[torch.Tensor,
def re_quantize_int8(weight: torch.FloatTensor) -> Tuple[torch.CharTensor, torch.FloatTensor]:
if weight.ndim > 2: # convs
weight = weight.flatten(1,-1)
if use_contiguous_mm:
weight, scale = quantize_int8(weight.t(), dim=0)
else:
weight, scale = quantize_int8(weight, dim=-1)
weight, scale = weight.t_(), scale.t_()
weight, scale = quantize_int8(weight.t(), dim=0)
return weight, scale