Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-08 22:08:42 +02:00
Code Issues Packages Projects Releases Wiki Activity

refactor(sdnq): clean up dynamic loss threshold slider

Browse Source
This commit is contained in:
CalamitousFelicitousness
2026-06-07 00:16:39 +01:00
parent 09fd3fa3b1
commit f83c1bc731
2 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/ui_definitions.py
+1 -1
View File
@@ -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),
ui/locale/locale_en.json
+1 -1
View File
@@ -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 <b><i>Quantization type</i></b> and steps up to higher precision until its error falls at or below this value.<br>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.<br><br>Only takes effect when <b><i>Use Dynamic quantization</i></b> is enabled.<br><br><b>-1</b> auto-selects a threshold from the base type, about <b>1e-4</b> for 8-bit or <b>1e-2</b> for 4-bit.<br><b>0</b> accepts only layers that quantize losslessly, which on a normal model leaves almost everything at full precision.<br><br>Default is <b>-1</b>.","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"},