diff --git a/ui/locale/locale_en.json b/ui/locale/locale_en.json
index 1436b289a..0bc67df19 100644
--- a/ui/locale/locale_en.json
+++ b/ui/locale/locale_en.json
@@ -393,8 +393,8 @@
{"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":"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":"Dequantize using torch.compile","localized":"","hint":"Compiles the dequantization step with torch.compile for faster inference. Requires Triton.
Changing this needs a full restart to take effect.
Enabled by default when Triton is available.","ui":"settings_quantization"},
+ {"id":"","label":"Dequantize using full precision","localized":"","hint":"Uses FP32 for the dequantization step for better numerical accuracy, at a small speed cost.
Enabled by default.","ui":"settings_quantization"},
{"id":"","label":"Disabled","localized":"","hint":"","ui":"settings_cuda"},
{"id":"","label":"Dynamic Attention BMM","localized":"","hint":"Performs attention computation in steps instead of all at once. Slower inference times, but greatly reduced memory usage","ui":"settings_cuda"},
{"id":"","label":"Dynamic attention","localized":"","hint":"Adjusts attention computation dynamically per step. Saves VRAM but slows generation.","ui":"settings_cuda"},
@@ -645,7 +645,7 @@
{"id":"","label":"Google Cloud project ID","localized":"","hint":"","ui":"settings_model_options"},
{"id":"","label":"Google Cloud location ID","localized":"","hint":"","ui":"settings_model_options"},
{"id":"","label":"group","localized":"","hint":"","ui":"settings_offload"},
- {"id":"","label":"Group size","localized":"","hint":"","ui":"settings_quantization"},
+ {"id":"","label":"Group size","localized":"","hint":"Number of weight elements that share one quantization scale. Smaller groups improve accuracy at a small size and speed cost; larger groups are leaner but coarser.
0 auto-selects a group size from the Quantization type; with Use quantized MatMul enabled at 6-bit or wider weights, it falls back to one scale per row. -1 uses one scale per row (no grouping).
Default is 0.","ui":"settings_quantization"},
{"id":"","label":"GC threshold","localized":"","hint":"","ui":"settings_backends"},
{"id":"","label":"Grid max rows count","localized":"","hint":"","ui":"settings_saving-images"},
{"id":"","label":"Grid max columns count","localized":"","hint":"","ui":"settings_saving-images"},
@@ -662,6 +662,7 @@
{"id":"","label":"Huggingface","localized":"","hint":"Settings related huggingface access"},
{"id":"","label":"HiDream","localized":"","hint":"","ui":"settings_model_options"},
{"id":"","label":"HyperTile","localized":"","hint":"","ui":"settings_advanced"},
+ {"id":"","label":"Hadamard group size","localized":"","hint":"Number of elements that share one Hadamard rotation group when Use Hadamard rotations is enabled.
Applies only when Use Hadamard rotations is enabled.
Default is 128.","ui":"settings_quantization"},
{"id":"","label":"HiDiffusion","localized":"","hint":"HiDiffusion allows creation of high-resolution images using your standard models without duplicates/distortions and improved performance","ui":"settings_advanced"},
{"id":"","label":"Height","localized":"","hint":"Target height of the output image in pixels.
For generation, this sets the resolution the model produces. For resize and upscale operations, this is the height the input is fitted to.
Should be a multiple of 8 for SD1.x and SDXL latents; newer architectures (Flux, SD3, video models) may require higher multiples (16, 32, or 64). Values that don't match are automatically floored to the nearest valid multiple for the loaded model.","ui":"txt2img"},
{"id":"","label":"HiRes steps","localized":"","hint":"Number of sampling steps for upscaled picture. If 0, uses same as for original","ui":"txt2img"},
@@ -942,8 +943,8 @@
{"id":"","label":"Modules to always offload","localized":"","hint":"","ui":"settings_offload"},
{"id":"","label":"Modules to never offload","localized":"","hint":"","ui":"settings_offload"},
{"id":"","label":"Model types not to quantize","localized":"","hint":"","ui":"settings_quantization"},
- {"id":"","label":"Modules to not convert","localized":"","hint":"","ui":"settings_quantization"},
- {"id":"","label":"Modules dtype dict","localized":"","hint":"","ui":"settings_quantization"},
+ {"id":"","label":"Modules to not convert","localized":"","hint":"Names of modules to leave unquantized (kept in original precision), separated by spaces, commas, or semicolons.
Useful for layers that are sensitive to quantization, such as gate or projection layers. Example: proj_out, x_embedder.
Some models already exclude sensitive modules by default; entries here extend that built-in list rather than replacing it.
Default is empty.","ui":"settings_quantization"},
+ {"id":"","label":"Modules dtype dict","localized":"","hint":"Advanced: JSON mapping a quantization type to a list of module names, to quantize specific modules at a different type than the global Quantization type. Example: {\"uint4\": [\"proj_out\"]}.
Some models already assign certain modules a specific type by default; entries here merge with those built-in mappings rather than replacing them.
Default is empty.","ui":"settings_quantization"},
{"id":"","label":"Math","localized":"","hint":"","ui":"settings_cuda"},
{"id":"","label":"Memory limit","localized":"","hint":"","ui":"settings_backends"},
{"id":"","label":"migraphx","localized":"","hint":"","ui":"settings_compile"},
@@ -1150,13 +1151,15 @@
{"id":"","label":"Qwen layered","localized":"","hint":"","ui":"settings_model_options"},
{"id":"","label":"Quicksettings","localized":"","hint":"","ui":"settings_ui"},
{"id":"","label":"Qwen layered number of layers","localized":"","hint":"","ui":"settings_model_options"},
- {"id":"","label":"Quantization mode","localized":"","hint":"","ui":"settings_quantization"},
- {"id":"","label":"Quantization type","localized":"","hint":"","ui":"settings_quantization"},
- {"id":"","label":"Quantized MatMul type","localized":"","hint":"","ui":"settings_quantization"},
- {"id":"","label":"Quantization type for Text Encoders","localized":"","hint":"","ui":"settings_quantization"},
- {"id":"","label":"Quantized MatMul type for Text Encoders","localized":"","hint":"","ui":"settings_quantization"},
- {"id":"","label":"Quantize convolutional layers","localized":"","hint":"","ui":"settings_quantization"},
- {"id":"","label":"Quantize using GPU","localized":"","hint":"","ui":"settings_quantization"},
+ {"id":"","label":"Quantization enabled","localized":"","hint":"Selects which model components SDNQ quantizes: Model (the diffusion transformer or UNet), TE (pipeline text encoders), LLM (standalone LLM/VLM tools like captioning and prompt enhance), Control (ControlNet), or VAE.
Quantization lowers VRAM use and can speed up inference at some accuracy cost.
An empty selection disables SDNQ.
Default is empty.","ui":"settings_quantization"},
+ {"id":"","label":"Quantization mode","localized":"","hint":"When SDNQ quantizes the model.
- auto: quantizes during load when possible, otherwise after load
- pre: quantizes layer by layer as the model loads, lowering peak memory
- post: quantizes after the full model is loaded
Default is auto.","ui":"settings_quantization"},
+ {"id":"","label":"Quantization type","localized":"","hint":"Weight data type SDNQ quantizes to. Lower bit widths (e.g. uint4) shrink the model and can speed up inference; higher bit widths (e.g. int8) keep more accuracy.
Integer types are widely supported; float types (float8_e4m3fn, etc.) only accelerate on newer GPUs.
Applies to the components selected in Quantization enabled.
Default is int8.","ui":"settings_quantization"},
+ {"id":"","label":"Quantized MatMul type","localized":"","hint":"Compute data type for quantized matrix multiplication, used when Use quantized MatMul is enabled.
auto picks int8 for integer weights or float8_e4m3fn / float16 for float weights.
Default is auto.","ui":"settings_quantization"},
+ {"id":"","label":"Quantization type for Text Encoders","localized":"","hint":"Weight data type for text encoder quantization, separate from the main Quantization type.
Same as model uses the main Quantization type.
Applies to both TE (pipeline text encoders) and LLM (standalone tools like captioning and prompt enhance) when selected in Quantization enabled.
Default is Same as model.","ui":"settings_quantization"},
+ {"id":"","label":"Quantized MatMul type for Text Encoders","localized":"","hint":"Quantized matrix multiplication type for text encoders and standalone LLM tools (captioning, prompt enhance), separate from the main Quantized MatMul type.
Same as model uses the main setting.
Default is Same as model.","ui":"settings_quantization"},
+ {"id":"","label":"Quantize convolutional layers","localized":"","hint":"Also quantizes convolutional layers, such as those in UNet models like SDXL.
Saves more memory, but convolutions are often more sensitive to quantization.
Disabled by default.","ui":"settings_quantization"},
+ {"id":"","label":"Quantize embedding layers","localized":"","hint":"Also quantizes embedding layers in text models.
Saves extra memory at some risk to quality.
Disabled by default.","ui":"settings_quantization"},
+ {"id":"","label":"Quantize using GPU","localized":"","hint":"Runs the quantization computation on the GPU instead of the CPU, which is much faster but uses VRAM during model load.
Disabling it keeps the computation on the CPU when load-time VRAM is limited.
Enabled by default.","ui":"settings_quantization"},
{"id":"","label":"Quantization weights type","localized":"","hint":"","ui":"settings_quantization"},
{"id":"","label":"Quantization activations type","localized":"","hint":"","ui":"settings_quantization"},
{"id":"","label":"Quicksettings list","localized":"","hint":"List of setting names, separated by commas, for settings that should go to the quick access bar at the top instead the setting tab","ui":"settings_ui"}
@@ -1372,9 +1375,9 @@
{"id":"","label":"Search Changelog","localized":"","hint":"","ui":"system_tab_changelog"},
{"id":"","label":"Stage boundary ratio","localized":"","hint":"","ui":"settings_model_options"},
{"id":"","label":"sequential","localized":"","hint":"","ui":"settings_offload"},
- {"id":"","label":"SVD rank size","localized":"","hint":"","ui":"settings_quantization"},
- {"id":"","label":"SVD steps","localized":"","hint":"","ui":"settings_quantization"},
- {"id":"","label":"Shuffle weights in post mode","localized":"","hint":"","ui":"settings_quantization"},
+ {"id":"","label":"SVD rank size","localized":"","hint":"Rank of the low-rank correction added by Use SVD quantization. Higher ranks recover more accuracy but add parameters and compute.
Applies only when Use SVD quantization is enabled.
Default is 32.","ui":"settings_quantization"},
+ {"id":"","label":"SVD steps","localized":"","hint":"Number of iterations used to estimate the low-rank correction for Use SVD quantization. More steps refine the estimate at the cost of longer quantization.
Applies only when Use SVD quantization is enabled.
Default is 8.","ui":"settings_quantization"},
+ {"id":"","label":"Shuffle weights in post mode","localized":"","hint":"In post quantization mode, processes model components through the GPU one at a time, moving each off the GPU before the next, to limit peak VRAM during quantization.
Applies only when Quantization mode is post.
Disabled by default.","ui":"settings_quantization"},
{"id":"","label":"SDXL: Use weighted pooled embeds","localized":"","hint":"","ui":"settings_text_encoder"},
{"id":"","label":"Sana: Use complex human instructions","localized":"","hint":"","ui":"settings_text_encoder"},
{"id":"","label":"Scaled-Dot-Product","localized":"","hint":"Memory optimization. Non-Deterministic unless SDP memory attention is disabled.","ui":"settings_cuda"},
@@ -1557,10 +1560,11 @@
{"id":"","label":"Username","localized":"","hint":"","ui":"component-8823"},
{"id":"","label":"UNET model","localized":"","hint":"","ui":"settings_sd"},
{"id":"","label":"Use torch streams","localized":"","hint":"","ui":"settings_offload"},
- {"id":"","label":"Use SVD quantization","localized":"","hint":"","ui":"settings_quantization"},
- {"id":"","label":"Use Dynamic quantization","localized":"","hint":"","ui":"settings_quantization"},
- {"id":"","label":"Use quantized MatMul","localized":"","hint":"","ui":"settings_quantization"},
- {"id":"","label":"Use quantized MatMul with conv","localized":"","hint":"","ui":"settings_quantization"},
+ {"id":"","label":"Use SVD quantization","localized":"","hint":"Adds a low-rank (SVDQuant) correction on top of SDNQ to recover accuracy lost at low bit widths, at the cost of extra size and compute.
Tuned by SVD rank size and SVD steps.
Disabled by default.","ui":"settings_quantization"},
+ {"id":"","label":"Use Dynamic quantization","localized":"","hint":"Picks a per-layer weight type automatically instead of one type everywhere. Each layer starts at the Quantization type (the minimum) and steps up to higher precision until its error meets the Dynamic loss threshold.
Protects error-sensitive layers at the cost of a larger model and slower load.
Disabled by default.","ui":"settings_quantization"},
+ {"id":"","label":"Use Hadamard rotations","localized":"","hint":"Applies a Hadamard rotation before quantizing to spread out weight outliers, which can improve accuracy at low bit widths.
Group size is set by Hadamard group size.
Disabled by default.","ui":"settings_quantization"},
+ {"id":"","label":"Use quantized MatMul","localized":"","hint":"Runs matrix multiplications in int8 or fp8 instead of bf16 / fp16, which can speed up inference on supported hardware.
Compute type is set by Quantized MatMul type.
Disabled by default.","ui":"settings_quantization"},
+ {"id":"","label":"Use quantized MatMul with conv","localized":"","hint":"Runs quantized matrix multiplication on convolutional layers, such as those in UNets like SDXL. This is the convolution counterpart of Use quantized MatMul and is controlled independently of it.
Only affects conv layers that are quantized, so it needs Quantize convolutional layers.
Disabled by default.","ui":"settings_quantization"},
{"id":"","label":"Use line break as prompt segment marker","localized":"","hint":"","ui":"settings_text_encoder"},
{"id":"","label":"Use zeros for prompt padding","localized":"","hint":"Force full zero tensor when prompt is empty to remove any residual noise","ui":"settings_text_encoder"},
{"id":"","label":"Upcast sampling","localized":"","hint":"Usually produces similar results to --no-half with better performance while using less memory","ui":"settings_cuda"},