Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-18 08:44:33 +02:00
Code Issues Packages Projects Releases Wiki Activity

SDNQ fix int8 matmul with qwen

Browse Source
This commit is contained in:
Disty0
2025-06-18 02:12:34 +03:00
parent 71c2714edf
commit e657cf790d
1 changed files with 5 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sdnq/__init__.py
+5 -2
View File
@@ -56,8 +56,11 @@ def sdnq_quantize_layer(layer, weights_dtype="int8", torch_dtype=None, group_siz
output_channel_size, channel_size = layer.weight.shape
if use_quantized_matmul:
use_quantized_matmul = weights_dtype in quantized_matmul_dtypes and channel_size >= 32 and output_channel_size >= 32
if use_quantized_matmul and not dtype_dict[weights_dtype]["is_integer"]:
use_quantized_matmul = output_channel_size % 16 == 0 and channel_size % 16 == 0
if use_quantized_matmul:
if dtype_dict[weights_dtype]["is_integer"]:
use_quantized_matmul = output_channel_size % 8 == 0 and channel_size % 8 == 0
else:
use_quantized_matmul = output_channel_size % 16 == 0 and channel_size % 16 == 0
if group_size == 0:
if is_linear_type: