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

Cleanup SDNQ

Browse Source
This commit is contained in:
Disty0
2026-05-22 17:32:41 +03:00
parent ad9c937e68
commit e484850e9e
3 changed files with 5 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sdnq/common.py
+1 -1
View File
@@ -6,7 +6,7 @@ import torch
from modules import shared, devices
sdnq_version = "0.1.9"
sdnq_version = "0.2.0"
sdnq_keys = {"weight", "scale", "zero_point", "svd_up", "svd_down"}
torch_version = torch.__version__[:4]
modules/sdnq/quant_utils.py
+3 -3
View File
@@ -105,7 +105,7 @@ def get_hadamard(n: int, dtype: torch.dtype | None = None, device: torch.device
@devices.inference_context()
def rotate_hadamard(weight: torch.Tensor, group_size: int = 128, hadamard: torch.Tensor | None = None, is_conv: bool = False) -> torch.Tensor:
def rotate_hadamard(weight: torch.Tensor, group_size: int = 128, hadamard: torch.FloatTensor | None = None, is_conv: bool = False) -> torch.Tensor:
if hadamard is None:
hadamard = get_hadamard(group_size, dtype=weight.dtype, device=weight.device)
else:
@@ -122,7 +122,7 @@ def rotate_hadamard(weight: torch.Tensor, group_size: int = 128, hadamard: torch
@devices.inference_context()
def apply_hadamard(weight: torch.Tensor, group_size: int = 128, hadamard: torch.Tensor | None = None, layer_class_name: str | None = None) -> torch.Tensor:
def apply_hadamard(weight: torch.Tensor, group_size: int = 128, hadamard: torch.FloatTensor | None = None, layer_class_name: str | None = None) -> torch.Tensor:
is_conv = False
use_hadamard = True
if layer_class_name in conv_types or layer_class_name in conv_transpose_types:
@@ -138,7 +138,7 @@ def apply_hadamard(weight: torch.Tensor, group_size: int = 128, hadamard: torch.
if group_size < 4:
use_hadamard = False
if use_hadamard:
weight = rotate_hadamard(weight, group_size=group_size, is_conv=is_conv)
weight = rotate_hadamard(weight, group_size=group_size, hadamard=hadamard, is_conv=is_conv)
return weight, use_hadamard, group_size
modules/sdnq/quantizer.py
+1 -1
View File
@@ -181,7 +181,6 @@ def sdnq_quantize_layer_weight(weight, layer_class_name=None, weights_dtype="int
else:
group_size = -1
cast_scale = True
transpose_weights = False
re_quantize_for_matmul = re_quantize_for_matmul or num_of_groups > 1
@@ -434,6 +433,7 @@ def sdnq_post_load_quant(
):
quantization_config = SDNQConfig(
weights_dtype=weights_dtype,
quantized_matmul_dtype=quantized_matmul_dtype,
hadamard_group_size=hadamard_group_size,
group_size=group_size,
svd_rank=svd_rank,