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

NNCF add number of quantization groups

Browse Source
This commit is contained in:
Disty0
2025-05-11 05:55:58 +03:00
parent fbaca247ef
commit 03a6d7f9bf
2 changed files with 65 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/shared.py
+6 -5
View File
@@ -542,16 +542,17 @@ options_templates.update(options_section(('quantization', "Quantization Settings
"nncf_compress_sep": OptionInfo("<h2>NNCF: Neural Network Compression Framework</h2>", "", gr.HTML),
"nncf_compress_weights": OptionInfo([], "Quantization enabled", gr.CheckboxGroup, {"choices": ["Model", "Transformer", "VAE", "TE", "Video", "LLM", "ControlNet"], "visible": native}),
"nncf_compress_mode": OptionInfo("post", "Quantization mode", gr.Dropdown, {"choices": ['pre', 'post'], "visible": native}),
"nncf_compress_mode": OptionInfo("post", "Quantization mode", gr.Dropdown, {"choices": ['pre', 'post'], "visible": native and not cmd_opts.use_openvino}),
"nncf_compress_weights_mode": OptionInfo("INT8_SYM", "Quantization type", gr.Dropdown, {"choices": ['INT8', 'INT8_SYM', 'INT4_ASYM', 'INT4_SYM', 'NF4'] if cmd_opts.use_openvino else ['INT8', 'INT8_SYM', 'INT4', 'INT4_SYM']}),
"nncf_compress_weights_raito": OptionInfo(0, "Compress ratio", gr.Slider, {"minimum": 0, "maximum": 1, "step": 0.01, "visible": cmd_opts.use_openvino}),
"nncf_compress_weights_group_size": OptionInfo(0, "Group size", gr.Slider, {"minimum": -1, "maximum": 512, "step": 1, "visible": cmd_opts.use_openvino}),
"nncf_compress_weights_num_of_groups": OptionInfo(1, "Number of quantization groups", gr.Slider, {"minimum": 1, "maximum": 128, "step": 1, "visible": native and not cmd_opts.use_openvino}),
"nncf_quantize": OptionInfo([], "OpenVINO enabled", gr.CheckboxGroup, {"choices": ["Model", "VAE", "TE"], "visible": cmd_opts.use_openvino}),
"nncf_quantize_mode": OptionInfo("INT8", "OpenVINO activations mode", gr.Dropdown, {"choices": ['INT8', 'FP8_E4M3', 'FP8_E5M2'], "visible": cmd_opts.use_openvino}),
"nncf_quantize_conv_layers": OptionInfo(False, "Quantize the convolutional layers", gr.Checkbox, {"visible": native}),
"nncf_decompress_fp32": OptionInfo(False, "Decompress using full precision", gr.Checkbox, {"visible": native}),
"nncf_decompress_compile": OptionInfo(devices.has_triton(), "Decompress using torch.compile", gr.Checkbox, {"visible": native}),
"nncf_quantize_shuffle_weights": OptionInfo(False, "Shuffle weights in post mode", gr.Checkbox, {"visible": native}),
"nncf_quantize_conv_layers": OptionInfo(False, "Quantize the convolutional layers", gr.Checkbox, {"visible": native and not cmd_opts.use_openvino}),
"nncf_decompress_fp32": OptionInfo(False, "Decompress using full precision", gr.Checkbox, {"visible": native and not cmd_opts.use_openvino}),
"nncf_decompress_compile": OptionInfo(devices.has_triton(), "Decompress using torch.compile", gr.Checkbox, {"visible": native and not cmd_opts.use_openvino}),
"nncf_quantize_shuffle_weights": OptionInfo(False, "Shuffle weights in post mode", gr.Checkbox, {"visible": native and not cmd_opts.use_openvino}),
"layerwise_quantization_sep": OptionInfo("<h2>Layerwise Casting</h2>", "", gr.HTML),
"layerwise_quantization": OptionInfo([], "Layerwise casting enabled", gr.CheckboxGroup, {"choices": ["Model", "Transformer", "TE"], "visible": native}),