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

add new optimum-quanto on-the-fly and simplify quantization loading

Browse Source 
Signed-off-by: Vladimir Mandic <mandic00@live.com>
This commit is contained in:
Vladimir Mandic
2025-03-16 21:45:05 -04:00
parent d4dff967b3
commit 4f56f4aa33
14 changed files with 104 additions and 148 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/shared.py
+6 -2
View File
@@ -513,7 +513,11 @@ options_templates.update(options_section(('quantization', "Quantization Settings
"bnb_quantization_type": OptionInfo("nf4", "Quantization type", gr.Dropdown, {"choices": ['nf4', 'fp8', 'fp4'], "visible": native}),
"bnb_quantization_storage": OptionInfo("uint8", "Backend storage", gr.Dropdown, {"choices": ["float16", "float32", "int8", "uint8", "float64", "bfloat16"], "visible": native}),
"optimum_quanto_sep": OptionInfo("<h2>Optimum Quanto</h2>", "", gr.HTML),
"quanto_quantization_sep": OptionInfo("<h2>Optimum Quanto</h2>", "", gr.HTML),
"quanto_quantization": OptionInfo([], "Quantization enabled", gr.CheckboxGroup, {"choices": ["Model", "VAE", "Text Encoder"], "visible": native}),
"quanto_quantization_type": OptionInfo("int8", "Quantization weights type", gr.Dropdown, {"choices": ["float8", "int8", "int4", "int2"], "visible": native}),
"optimum_quanto_sep": OptionInfo("<h2>Optimum Quanto: post-load</h2>", "", gr.HTML),
"optimum_quanto_weights": OptionInfo([], "Quantization enabled", gr.CheckboxGroup, {"choices": ["Model", "VAE", "Text Encoder", "ControlNet"], "visible": native}),
"optimum_quanto_weights_type": OptionInfo("qint8", "Quantization weights type", gr.Dropdown, {"choices": ['qint8', 'qfloat8_e4m3fn', 'qfloat8_e5m2', 'qint4', 'qint2'], "visible": native}),
"optimum_quanto_activations_type": OptionInfo("none", "Quantization activations type ", gr.Dropdown, {"choices": ['none', 'qint8', 'qfloat8_e4m3fn', 'qfloat8_e5m2'], "visible": native}),
@@ -524,7 +528,7 @@ options_templates.update(options_section(('quantization', "Quantization Settings
"torchao_quantization_mode": OptionInfo("pre", "Quantization mode", gr.Dropdown, {"choices": ['pre', 'post'], "visible": native}),
"torchao_quantization_type": OptionInfo("int8_weight_only", "Quantization type", gr.Dropdown, {"choices": ['int4_weight_only', 'int8_dynamic_activation_int4_weight', 'int8_weight_only', 'int8_dynamic_activation_int8_weight', 'float8_weight_only', 'float8_dynamic_activation_float8_weight', 'float8_static_activation_float8_weight'], "visible": native}),
"nncf_compress_sep": OptionInfo("<h2>NNCF</h2>", "", gr.HTML),
"nncf_compress_sep": OptionInfo("<h2>NNCF: Neural Network Compression Framework</h2>", "", gr.HTML),
"nncf_compress_weights": OptionInfo([], "Quantization enabled", gr.CheckboxGroup, {"choices": ["Model", "VAE", "Text Encoder", "ControlNet"], "visible": native}),
"nncf_compress_weights_mode": OptionInfo("INT8", "Quantization type", gr.Dropdown, {"choices": ['INT8', 'INT8_SYM', 'INT4_ASYM', 'INT4_SYM', 'NF4'] if cmd_opts.use_openvino else ['INT8']}),
"nncf_compress_weights_raito": OptionInfo(0, "Compress ratio", gr.Slider, {"minimum": 0, "maximum": 1, "step": 0.01, "visible": cmd_opts.use_openvino}),