Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-19 09:14:35 +02:00
Code Issues Packages Projects Releases Wiki Activity

Fix SDNQ Lora

Browse Source
This commit is contained in:
Disty0
2026-05-15 21:25:08 +03:00
parent 58891d3a4e
commit 8e11971572
3 changed files with 20 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
extensions-builtin/sd-extension-system-info
Submodule
+1
Submodule extensions-builtin/sd-extension-system-info added at 006f08f499
extensions-builtin/stable-diffusion-webui-rembg
Submodule
+1
Submodule extensions-builtin/stable-diffusion-webui-rembg added at 988f8b78c7
modules/lora/lora_apply.py
+18 -16
View File
@@ -149,7 +149,7 @@ def network_add_weights(self: torch.nn.Conv2d | torch.nn.Linear | torch.nn.Group
weight, new_weight = None, None
if not bias and hasattr(self, "sdnq_dequantizer"):
try:
from modules.sdnq import sdnq_quantize_layer
from modules.sdnq import SDNQConfig, sdnq_quantize_layer
if hasattr(self, "sdnq_dequantizer_backup"):
use_svd = bool(self.sdnq_svd_up_backup is not None)
dequantize_fp32 = bool(self.sdnq_scale_backup.dtype == torch.float32)
@@ -179,23 +179,25 @@ def network_add_weights(self: torch.nn.Conv2d | torch.nn.Linear | torch.nn.Group
new_weight = dequant_weight.to(devices.device, dtype=torch.float32) + lora_weights.to(devices.device, dtype=torch.float32)
self.weight = torch.nn.Parameter(new_weight, requires_grad=False)
del self.sdnq_dequantizer, self.scale, self.zero_point, self.svd_up, self.svd_down
self.sdnq_dequantizer = self.scale = self.zero_point = self.svd_up = self.svd_down = None
self = sdnq_quantize_layer(
self,
weights_dtype=sdnq_dequantizer.weights_dtype,
quantized_matmul_dtype=sdnq_dequantizer.quantized_matmul_dtype,
torch_dtype=sdnq_dequantizer.result_dtype,
group_size=sdnq_dequantizer.group_size,
svd_rank=sdnq_dequantizer.svd_rank,
use_quantized_matmul=sdnq_dequantizer.use_quantized_matmul,
use_quantized_matmul_conv=sdnq_dequantizer.use_quantized_matmul,
use_svd=use_svd,
dequantize_fp32=dequantize_fp32,
svd_steps=shared.opts.sdnq_svd_steps,
quant_conv=True, # quant_conv is True if conv layers ends up here
non_blocking=False,
quantization_device=devices.device,
return_device=device,
SDNQConfig(
weights_dtype=sdnq_dequantizer.weights_dtype,
quantized_matmul_dtype=sdnq_dequantizer.quantized_matmul_dtype,
torch_dtype=sdnq_dequantizer.result_dtype,
group_size=sdnq_dequantizer.group_size,
svd_rank=sdnq_dequantizer.svd_rank,
use_quantized_matmul=sdnq_dequantizer.use_quantized_matmul,
use_quantized_matmul_conv=sdnq_dequantizer.use_quantized_matmul,
use_svd=use_svd,
dequantize_fp32=dequantize_fp32,
svd_steps=shared.opts.sdnq_svd_steps,
quant_conv=True, # quant_conv is True if conv layers ends up here
non_blocking=False,
quantization_device=devices.device,
return_device=device,
),
param_name=getattr(self, 'network_layer_name', None),
)[0].to(device)
weight = None