Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-20 01:31:13 +02:00
Code Issues Packages Projects Releases Wiki Activity

update sdnq

Browse Source
This commit is contained in:
Disty0
2026-01-14 16:23:26 +03:00
parent 4dbdee10e3
commit 784cda80aa
13 changed files with 121 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sdnq/loader.py
+3 -4
View File
@@ -170,7 +170,7 @@ def apply_sdnq_options_to_module(model, dtype: torch.dtype = None, dequantize_fp
return model
for module_name, module in model.named_children():
if hasattr(module, "sdnq_dequantizer"):
layer_class_name = module.__class__.__name__
layer_class_name = module.original_class.__name__
current_use_quantized_matmul = use_quantized_matmul
if current_use_quantized_matmul:
if layer_class_name in conv_types:
@@ -204,7 +204,7 @@ def apply_sdnq_options_to_module(model, dtype: torch.dtype = None, dequantize_fp
module.svd_down.data = module.svd_down.to(dtype=scale_dtype)
if current_use_quantized_matmul is not None and current_use_quantized_matmul != module.sdnq_dequantizer.use_quantized_matmul:
if not module.sdnq_dequantizer.re_quantize_for_matmul:
if not module.sdnq_dequantizer.re_quantize_for_matmul and not dtype_dict[module.sdnq_dequantizer.weights_dtype]["is_packed"]:
module.scale.t_()
module.weight.t_()
if current_use_quantized_matmul:
@@ -215,8 +215,7 @@ def apply_sdnq_options_to_module(model, dtype: torch.dtype = None, dequantize_fp
if module.svd_up is not None:
module.svd_up.data, module.svd_down.data = prepare_svd_for_matmul(module.svd_up.t_(), module.svd_down.t_(), current_use_quantized_matmul)
module.sdnq_dequantizer.use_quantized_matmul = current_use_quantized_matmul
module.forward = get_forward_func(module.__class__.__name__, module.sdnq_dequantizer.quantized_matmul_dtype, current_use_quantized_matmul)
module.forward = module.forward.__get__(module, module.__class__)
module.forward_func = get_forward_func(module.original_class.__name__, module.sdnq_dequantizer.quantized_matmul_dtype, current_use_quantized_matmul)
setattr(model, module_name, module)
else:
setattr(model, module_name, apply_sdnq_options_to_module(module, dtype=dtype, dequantize_fp32=dequantize_fp32, use_quantized_matmul=use_quantized_matmul))