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

OpenVINO add group_size option to compression

Browse Source
This commit is contained in:
Disty0
2025-01-23 00:15:27 +03:00
parent bda45d865e
commit 40e7dcd5a4
2 changed files with 8 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/intel/openvino/__init__.py
+6 -2
View File
@@ -232,7 +232,9 @@ def openvino_compile(gm: GraphModule, *example_inputs, model_hash_str: str = Non
if dont_use_4bit_nncf or shared.opts.nncf_compress_weights_mode == "INT8":
om = nncf.compress_weights(om)
else:
om = nncf.compress_weights(om, mode=getattr(nncf.CompressWeightsMode, shared.opts.nncf_compress_weights_mode), group_size=8, ratio=shared.opts.nncf_compress_weights_raito)
compress_group_size = shared.opts.nncf_compress_weights_group_size if shared.opts.nncf_compress_weights_group_size != 0 else None
compress_ratio = shared.opts.nncf_compress_weights_raito if shared.opts.nncf_compress_weights_raito != 0 else None
om = nncf.compress_weights(om, mode=getattr(nncf.CompressWeightsMode, shared.opts.nncf_compress_weights_mode), group_size=compress_group_size, ratio=compress_ratio)
hints = {}
if shared.opts.openvino_accuracy == "performance":
@@ -290,7 +292,9 @@ def openvino_compile_cached_model(cached_model_path, *example_inputs):
if dont_use_4bit_nncf or shared.opts.nncf_compress_weights_mode == "INT8":
om = nncf.compress_weights(om)
else:
om = nncf.compress_weights(om, mode=getattr(nncf.CompressWeightsMode, shared.opts.nncf_compress_weights_mode), group_size=8, ratio=shared.opts.nncf_compress_weights_raito)
compress_group_size = shared.opts.nncf_compress_weights_group_size if shared.opts.nncf_compress_weights_group_size != 0 else None
compress_ratio = shared.opts.nncf_compress_weights_raito if shared.opts.nncf_compress_weights_raito != 0 else None
om = nncf.compress_weights(om, mode=getattr(nncf.CompressWeightsMode, shared.opts.nncf_compress_weights_mode), group_size=compress_group_size, ratio=compress_ratio)
hints = {'CACHE_DIR': shared.opts.openvino_cache_path + '/blob'}
if shared.opts.openvino_accuracy == "performance":
modules/shared.py
+2 -1
View File
@@ -575,7 +575,8 @@ options_templates.update(options_section(('quantization', "Quantization Settings
"nncf_sep": OptionInfo("<h2>NNCF</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(1.0, "Compress ratio", gr.Slider, {"minimum": 0, "maximum": 1, "step": 0.01, "visible": cmd_opts.use_openvino}),
"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_quantize": OptionInfo([], "OpenVINO enabled", gr.CheckboxGroup, {"choices": ["Model", "VAE", "Text Encoder"], "visible": cmd_opts.use_openvino}),
"nncf_quant_mode": OptionInfo("INT8", "OpenVINO mode", gr.Dropdown, {"choices": ['INT8', 'FP8_E4M3', 'FP8_E5M2'], "visible": cmd_opts.use_openvino}),
"quant_shuffle_weights": OptionInfo(False, "Shuffle weights", gr.Checkbox, {"visible": native}),