Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-19 17:24:32 +02:00
Code Issues Packages Projects Releases Wiki Activity

SDNQ set sdnq_dequantize_fp32 to True by default

Browse Source
This commit is contained in:
Disty0
2026-03-11 14:35:11 +03:00
parent 30d21f420f
commit e202d43801
2 changed files with 8 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sdnq/quantizer.py
+6 -6
View File
@@ -225,7 +225,7 @@ def add_module_skip_keys(model, modules_to_not_convert: list[str] = None, module
@devices.inference_context()
def sdnq_quantize_layer_weight(weight, layer_class_name=None, weights_dtype="int8", quantized_matmul_dtype=None, torch_dtype=None, group_size=0, svd_rank=32, svd_steps=8, use_svd=False, use_quantized_matmul=False, use_stochastic_rounding=False, dequantize_fp32=False, using_pre_calculated_svd=False, skip_sr=False, param_name=None): # pylint: disable=unused-argument
def sdnq_quantize_layer_weight(weight, layer_class_name=None, weights_dtype="int8", quantized_matmul_dtype=None, torch_dtype=None, group_size=0, svd_rank=32, svd_steps=8, use_svd=False, use_quantized_matmul=False, use_stochastic_rounding=False, dequantize_fp32=True, using_pre_calculated_svd=False, skip_sr=False, param_name=None): # pylint: disable=unused-argument
num_of_groups = 1
is_conv_type = False
is_conv_transpose_type = False
@@ -413,7 +413,7 @@ def sdnq_quantize_layer_weight(weight, layer_class_name=None, weights_dtype="int
@devices.inference_context()
def sdnq_quantize_layer_weight_dynamic(weight, layer_class_name=None, weights_dtype="int2", quantized_matmul_dtype=None, torch_dtype=None, group_size=0, svd_rank=32, svd_steps=8, dynamic_loss_threshold=1e-2, use_svd=False, use_quantized_matmul=False, use_dynamic_quantization=False, use_stochastic_rounding=False, dequantize_fp32=False, param_name=None): # pylint: disable=unused-argument
def sdnq_quantize_layer_weight_dynamic(weight, layer_class_name=None, weights_dtype="int2", quantized_matmul_dtype=None, torch_dtype=None, group_size=0, svd_rank=32, svd_steps=8, dynamic_loss_threshold=1e-2, use_svd=False, use_quantized_matmul=False, use_dynamic_quantization=False, use_stochastic_rounding=False, dequantize_fp32=True, param_name=None): # pylint: disable=unused-argument
if torch_dtype is None:
torch_dtype = weight.dtype
if weight.dtype != torch.float64:
@@ -465,7 +465,7 @@ def sdnq_quantize_layer_weight_dynamic(weight, layer_class_name=None, weights_dt
@devices.inference_context()
def sdnq_quantize_layer(layer, weights_dtype="int8", quantized_matmul_dtype=None, torch_dtype=None, group_size=0, svd_rank=32, svd_steps=8, dynamic_loss_threshold=1e-2, use_svd=False, quant_conv=False, use_quantized_matmul=False, use_quantized_matmul_conv=False, use_dynamic_quantization=False, use_stochastic_rounding=False, dequantize_fp32=False, non_blocking=False, modules_to_not_convert=None, modules_dtype_dict=None, quantization_device=None, return_device=None, param_name=None): # pylint: disable=unused-argument
def sdnq_quantize_layer(layer, weights_dtype="int8", quantized_matmul_dtype=None, torch_dtype=None, group_size=0, svd_rank=32, svd_steps=8, dynamic_loss_threshold=1e-2, use_svd=False, quant_conv=False, use_quantized_matmul=False, use_quantized_matmul_conv=False, use_dynamic_quantization=False, use_stochastic_rounding=False, dequantize_fp32=True, non_blocking=False, modules_to_not_convert=None, modules_dtype_dict=None, quantization_device=None, return_device=None, param_name=None): # pylint: disable=unused-argument
layer_class_name = layer.__class__.__name__
if layer_class_name in conv_transpose_types or layer_class_name in conv_types:
if not quant_conv:
@@ -549,7 +549,7 @@ def sdnq_quantize_layer(layer, weights_dtype="int8", quantized_matmul_dtype=None
@devices.inference_context()
def apply_sdnq_to_module(model, weights_dtype="int8", quantized_matmul_dtype=None, torch_dtype=None, group_size=0, svd_rank=32, svd_steps=8, dynamic_loss_threshold=1e-2, use_svd=False, quant_conv=False, use_quantized_matmul=False, use_quantized_matmul_conv=False, use_dynamic_quantization=False, use_stochastic_rounding=False, dequantize_fp32=False, non_blocking=False, modules_to_not_convert: list[str] = None, modules_dtype_dict: dict[str, list[str]] = None, modules_quant_config: dict[str, dict] = None, quantization_device=None, return_device=None, full_param_name=""): # pylint: disable=unused-argument
def apply_sdnq_to_module(model, weights_dtype="int8", quantized_matmul_dtype=None, torch_dtype=None, group_size=0, svd_rank=32, svd_steps=8, dynamic_loss_threshold=1e-2, use_svd=False, quant_conv=False, use_quantized_matmul=False, use_quantized_matmul_conv=False, use_dynamic_quantization=False, use_stochastic_rounding=False, dequantize_fp32=True, non_blocking=False, modules_to_not_convert: list[str] = None, modules_dtype_dict: dict[str, list[str]] = None, modules_quant_config: dict[str, dict] = None, quantization_device=None, return_device=None, full_param_name=""): # pylint: disable=unused-argument
has_children = list(model.children())
if not has_children:
return model, modules_to_not_convert, modules_dtype_dict
@@ -642,7 +642,7 @@ def sdnq_post_load_quant(
use_quantized_matmul_conv: bool = False,
use_dynamic_quantization: bool = False,
use_stochastic_rounding: bool = False,
dequantize_fp32: bool = False,
dequantize_fp32: bool = True,
non_blocking: bool = False,
add_skip_keys:bool = True,
quantization_device: torch.device = None,
@@ -1070,7 +1070,7 @@ class SDNQConfig(QuantizationConfigMixin):
use_static_quantization: bool = True,
use_dynamic_quantization: bool = False,
use_stochastic_rounding: bool = False,
dequantize_fp32: bool = False,
dequantize_fp32: bool = True,
non_blocking: bool = False,
add_skip_keys: bool = True,
quantization_device: torch.device = None,
modules/ui_definitions.py
+2 -2
View File
@@ -150,7 +150,7 @@ def create_settings(cmd_opts):
"sdnq_quantize_weights_group_size": OptionInfo(0, "Group size", gr.Slider, {"minimum": -1, "maximum": 4096, "step": 1}),
"sdnq_svd_rank": OptionInfo(32, "SVD rank size", gr.Slider, {"minimum": 1, "maximum": 512, "step": 1}),
"sdnq_svd_steps": OptionInfo(8, "SVD steps", gr.Slider, {"minimum": 1, "maximum": 128, "step": 1}),
"sdnq_dynamic_loss_threshold": OptionInfo(1e-2, "Dynamic loss threshold", gr.Slider, {"minimum": 1e-4, "maximum": 1e-1, "step": 1e-4}),
"sdnq_dynamic_loss_threshold": OptionInfo(1e-2, "Dynamic loss threshold", gr.Slider, {"minimum": 1e-8, "maximum": 1, "step": 1e-4}),
"sdnq_use_svd": OptionInfo(False, "Use SVD quantization", gr.Checkbox),
"sdnq_use_dynamic_quantization": OptionInfo(False, "Use Dynamic quantization", gr.Checkbox),
"sdnq_quantize_conv_layers": OptionInfo(False, "Quantize convolutional layers", gr.Checkbox),
@@ -158,7 +158,7 @@ def create_settings(cmd_opts):
"sdnq_use_quantized_matmul": OptionInfo(False, "Use quantized MatMul", gr.Checkbox),
"sdnq_use_quantized_matmul_conv": OptionInfo(False, "Use quantized MatMul with conv", gr.Checkbox),
"sdnq_quantize_with_gpu": OptionInfo(True, "Quantize using GPU", gr.Checkbox),
"sdnq_dequantize_fp32": OptionInfo(False, "Dequantize using full precision", gr.Checkbox),
"sdnq_dequantize_fp32": OptionInfo(True, "Dequantize using full precision", gr.Checkbox),
"sdnq_quantize_shuffle_weights": OptionInfo(False, "Shuffle weights in post mode", gr.Checkbox),
"nunchaku_sep": OptionInfo("<h2>Nunchaku Engine</h2>", "", gr.HTML),