Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-19 01:04:32 +02:00
Code Issues Packages Projects Releases Wiki Activity

SDNQ fix weights_dtype getting overwritten on post load quant

Browse Source
This commit is contained in:
Disty0
2025-11-14 16:51:10 +03:00
parent a597f94df8
commit 3a4d7795d8
1 changed files with 17 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sdnq/quantizer.py
+17 -21
View File
@@ -359,27 +359,23 @@ def apply_sdnq_to_module(model, weights_dtype="int8", torch_dtype=None, group_si
if layer_class_name in allowed_types:
if (layer_class_name in conv_types or layer_class_name in conv_transpose_types) and not quant_conv:
continue
else:
continue
weights_dtype = get_minimum_dtype(weights_dtype, param_name, modules_dtype_dict)
module = sdnq_quantize_layer(
module,
weights_dtype=weights_dtype,
torch_dtype=torch_dtype,
group_size=group_size,
svd_rank=svd_rank,
svd_steps=svd_steps,
use_svd=use_svd,
quant_conv=quant_conv,
use_quantized_matmul=use_quantized_matmul,
use_quantized_matmul_conv=use_quantized_matmul_conv,
dequantize_fp32=dequantize_fp32,
non_blocking=non_blocking,
quantization_device=quantization_device,
return_device=return_device,
param_name=param_name,
)
module = sdnq_quantize_layer(
module,
weights_dtype=get_minimum_dtype(weights_dtype, param_name, modules_dtype_dict),
torch_dtype=torch_dtype,
group_size=group_size,
svd_rank=svd_rank,
svd_steps=svd_steps,
use_svd=use_svd,
quant_conv=quant_conv,
use_quantized_matmul=use_quantized_matmul,
use_quantized_matmul_conv=use_quantized_matmul_conv,
dequantize_fp32=dequantize_fp32,
non_blocking=non_blocking,
quantization_device=quantization_device,
return_device=return_device,
param_name=param_name,
)
module = apply_sdnq_to_module(
module,
weights_dtype=weights_dtype,