Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-19 17:24:32 +02:00
Code Issues Packages Projects Releases Wiki Activity

SDNQ override conv dtype

Browse Source
This commit is contained in:
Disty0
2025-06-03 02:37:29 +03:00
parent e9ff242e03
commit 79ec23dd9d
1 changed files with 8 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/model_quant_sdnq.py
+8
View File
@@ -61,6 +61,10 @@ def sdnq_quantize_layer(layer, weights_dtype="int8", torch_dtype=None, group_siz
use_quantized_matmul = False
is_conv_type = True
output_channel_size, channel_size = layer.weight.shape[:2]
if dtype_dict[weights_dtype]["num_bits"] < 4:
weights_dtype = "uint4"
elif dtype_dict[weights_dtype]["num_bits"] < 6:
weights_dtype = "int6"
elif layer_class_name in conv_transpose_types:
if not quant_conv:
return layer
@@ -69,6 +73,10 @@ def sdnq_quantize_layer(layer, weights_dtype="int8", torch_dtype=None, group_siz
use_quantized_matmul = False
is_conv_transpose_type = True
channel_size, output_channel_size = layer.weight.shape[:2]
if dtype_dict[weights_dtype]["num_bits"] < 4:
weights_dtype = "uint4"
elif dtype_dict[weights_dtype]["num_bits"] < 6:
weights_dtype = "int6"
else:
is_linear_type = True
reduction_axes = -1