GitLab

* Fix: Correctly specify non-zero GPUs in multi-GPU environments

This commit resolves an issue where the Nunchaku model could not be
correctly initialized and run on a user-specified non-zero GPU in
multi-GPU systems.

Key changes include:
- Using CUDADeviceContext in the FluxModel constructor to ensure
  the model and its submodules are created within the specified GPU context.
- Modifying the logic in FluxModel::forward for copying residual data
  from CPU back to GPU, ensuring it returns to the correct original GPU device.
- Adding explicit CUDA context management in Tensor::copy_ for data
  copy operations involving CUDA devices (H2D, D2H, D2D) to guarantee
  cudaMemcpyAsync executes on the correct device.

These changes allow users to reliably run Nunchaku on any specified
GPU in a multi-GPU setup.

* finish pre-commit

* fix:txt2_img pipe and controlnet pipe share components

* add the test script

* comfyUI bug fix

* garbage collection

* wrap the test

---------
Co-authored-by: Muyang Li <lmxyy1999@foxmail.com>

* fix the use count

* update

* change the example step to 20

* allow embedding website

* support fp4 on blackwell

* fp4 kontext runnable

* add more examples

* add kontext examples

chore: update the kontext support

* update kontext examples

* update tests

* add tests for kontext

* remove the warning of txt_ids and img_ids

* chore: add kontext to be synced from hf to ms

* add kontext demo

* make linter happy

* style: make linter happy

* update docs

* change the model path

* change the model path

* fix the model paths

* rename facexlib_path to facexlib_dirpath

* add start_timestep and end_timestep

* no need to download the files if the folder exists

* kohya supported

* add a test for the LoRA

* add a gc

* upgrade pulid to 0.9.1,When the EVA CLIP config is not found, I use the built-in JSON parameters instead.The issue of repeated model downloads has been resolved, and the model path specified by ComfyUI will be faithfully used.

style:apply black check

style:apply isort check

* fix callback bugs

style: apply clang-format

* fix:cache issue if offload is set to True

* fix: lint

* bump the version to 0.3.1

* update the t2i demos

* remove the dependencies of GPUtils

* update the html

* update the html

* update the html

* update the html

* fix the demos

* demo runnable again

* support building torch2.8 wheels

* avoid recreating the tag

* rebuild torch2.8 wheels

* support torch2.8 wheels