diff --git a/modules/ui_definitions.py b/modules/ui_definitions.py index e7ba1ac26..17c181e50 100644 --- a/modules/ui_definitions.py +++ b/modules/ui_definitions.py @@ -162,7 +162,7 @@ def create_settings(cmd_opts): "sdnq_hadamard_group_size": OptionInfo(128, "Hadamard group size", gr.Slider, {"minimum": 4, "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-8, "Dynamic loss threshold", gr.Slider, {"minimum": -1e-8, "maximum": 1e-1, "step": 1e-8}), + "sdnq_dynamic_loss_threshold": OptionInfo(-1, "Dynamic loss threshold", gr.Slider, {"minimum": -1, "maximum": 0.1, "step": 1e-4}), "sdnq_use_svd": OptionInfo(False, "Use SVD quantization", gr.Checkbox), "sdnq_use_hadamard": OptionInfo(False, "Use Hadamard rotations", gr.Checkbox), "sdnq_use_dynamic_quantization": OptionInfo(False, "Use Dynamic quantization", gr.Checkbox), diff --git a/ui/locale/locale_en.json b/ui/locale/locale_en.json index 6714b66ad..1436b289a 100644 --- a/ui/locale/locale_en.json +++ b/ui/locale/locale_en.json @@ -392,7 +392,7 @@ {"id":"","label":"Diffusers load using Run:ai streamer","localized":"","hint":"","ui":"settings_sd"}, {"id":"","label":"Disable accelerate","localized":"","hint":"","ui":"settings_sd"}, {"id":"","label":"Disable T5 text encoder","localized":"","hint":"","ui":"settings_model_options"}, - {"id":"","label":"Dynamic loss threshold","localized":"","hint":"","ui":"settings_quantization"}, + {"id":"","label":"Dynamic loss threshold","localized":"","hint":"Target per-layer quantization error (normalized MSE) for dynamic quantization. Each layer starts at the base Quantization type and steps up to higher precision until its error falls at or below this value.
Lower values keep more layers at higher precision for a larger, more accurate model; higher values let more layers stay at the base type for a smaller one.

Only takes effect when Use Dynamic quantization is enabled.

-1 auto-selects a threshold from the base type, about 1e-4 for 8-bit or 1e-2 for 4-bit.
0 accepts only layers that quantize losslessly, which on a normal model leaves almost everything at full precision.

Default is -1.","ui":"settings_quantization"}, {"id":"","label":"Dequantize using torch.compile","localized":"","hint":"","ui":"settings_quantization"}, {"id":"","label":"Dequantize using full precision","localized":"","hint":"","ui":"settings_quantization"}, {"id":"","label":"Disabled","localized":"","hint":"","ui":"settings_cuda"},