empty selection for *Hires VAE / TE* means 'use modules built-in to the model file', same as top-row VAE / TE selection. But this must be written to infotext as something, to distinguish from the non-HiRes case.
adds selection of none/same/different modules for hiresfix
('Use same choices' default option has priority over other selections made at same time.)
includes saving/loading from infotext
- `/sdapi/v1/options` GET now calls `get_config()` from **sysinfo** module, instead of from its own version of the function.
- Defined a new, flexible and more robust `set_config()` function in **sysinfo** module, which:
- obsoletes redundant code
- skips updating values that are unchanged
- has flexible args for both API and UI use
- `/sdapi/v1/options` POST and `override_settings` now use the new `set_config()` function. `set_config()` could possibly obsolete additional functions, but I'm not going to get into that just yet.
- Options for `forge_additional_modules` can now be provided either as the file path, or just the module name.
- Most importantly, `refresh_model_loading_parameters()` is now only called ONCE per request, and **only** if necessary.
- It is now much easier to call `shared.opts.save()` as needed
* Do refresh load params for modules
* Adjust call order for model mgmt/prompt cache
* new function `manage_model_and_prompt_cache()` to improve code clarity
#1760, original solution by cmdr2. extended to 2 other locations where setting is read. I could replicate the issue only by manually entering a decimal value.
Possibly also #1764. Though I don't know how the clipskip setting became a string.
1. fix several problems related to layerdiffuse not unloaded
2. fix several problems related to Fooocus inpaint
3. Slightly speed up on-the-fly LoRAs by precomputing them to computation dtype
1. Add an option to allow users to use UNet in fp8/gguf but lora in fp16.
2. All FP16 loras do not need patch. Others will only patch again when lora weight change.
3. FP8 unet + fp16 lora are available (somewhat only available) in Forge now. This also solves some “LoRA too subtle” problems.
4. Significantly speed up all gguf models (in Async mode) by using independent thread (CUDA stream) to compute and dequant at the same time, even when low-bit weights are already on GPU.
5. View “online lora” as a module similar to ControlLoRA so that it is moved to GPU together with model when sampling, achieving significant speedup and perfect low VRAM management simultaneously.
100% reproduce all previous results, including TI embeddings, LoRAs in CLIP, emphasize settings, BREAK, timestep swap scheduling, AB mixture, advanced uncond, etc
Backend is 85% finished
