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

SDNQ fix a singular bias not getting offloaded

Browse Source
This commit is contained in:
Disty0
2025-10-11 02:38:49 +03:00
parent 1d2775103f
commit 0f785880ee
1 changed files with 3 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sdnq/quantizer.py
+3
View File
@@ -230,8 +230,11 @@ def sdnq_quantize_layer(layer, weights_dtype="int8", torch_dtype=None, group_siz
use_quantized_matmul=use_quantized_matmul,
re_quantize_for_matmul=re_quantize_for_matmul,
)
layer.weight.data = layer.sdnq_dequantizer.pack_weight(layer.weight).to(return_device, non_blocking=non_blocking)
layer.sdnq_dequantizer = layer.sdnq_dequantizer.to(return_device, non_blocking=non_blocking)
if layer.bias is not None and layer.bias.device != "meta":
layer.bias.data = layer.bias.to(return_device, non_blocking=non_blocking)
layer.forward = get_forward_func(layer_class_name, use_quantized_matmul, dtype_dict[weights_dtype]["is_integer"], use_tensorwise_fp8_matmul)
layer.forward = layer.forward.__get__(layer, layer.__class__)