Add NPU Support

[wangshuai09]

Description

This PR enables users to leverage the Ascend NPU for using features including txt2img, img2img, etc.

#13263

Screenshots/videos:

Verified on NPU, prompt is panda:

Verified on Windows with cpu, prompt is sitting panda:

Checklist:

• I have read contributing wiki page
• I have performed a self-review of my own code
• My code follows the style guidelines
• My code passes tests

(python310) root@ascend-01:/root/project/own/stable-diffusion-webui# python -m pytest -vv --verify-base-url test                 
=========================================================================== test session starts ============================================================================
platform linux -- Python 3.10.6, pytest-7.4.4, pluggy-1.4.0 -- /root/miniconda/envs/python310/bin/python
cachedir: .pytest_cache
baseurl: http://127.0.0.1:7862/
rootdir: /root/project/own/stable-diffusion-webui
configfile: pyproject.toml
plugins: cov-4.1.0, base-url-2.0.0
collected 29 items                                                                                                                                                         

test/test_extras.py::test_simple_upscaling_performed PASSED                                                                                                          [  3%]
test/test_extras.py::test_png_info_performed PASSED                                                                                                                  [  6%]
test/test_extras.py::test_interrogate_performed PASSED                                                                                                               [ 10%]
test/test_img2img.py::test_img2img_simple_performed PASSED                                                                                                           [ 13%]
test/test_img2img.py::test_inpainting_masked_performed PASSED                                                                                                        [ 17%]
test/test_img2img.py::test_inpainting_with_inverted_masked_performed PASSED                                                                                          [ 20%]
test/test_img2img.py::test_img2img_sd_upscale_performed PASSED                                                                                                       [ 24%]
test/test_txt2img.py::test_txt2img_simple_performed PASSED                                                                                                           [ 27%]
test/test_txt2img.py::test_txt2img_with_negative_prompt_performed PASSED                                                                                             [ 31%]
test/test_txt2img.py::test_txt2img_with_complex_prompt_performed PASSED                                                                                              [ 34%]
test/test_txt2img.py::test_txt2img_not_square_image_performed PASSED                                                                                                 [ 37%]
test/test_txt2img.py::test_txt2img_with_hrfix_performed PASSED                                                                                                       [ 41%]
test/test_txt2img.py::test_txt2img_with_tiling_performed PASSED                                                                                                      [ 44%]
test/test_txt2img.py::test_txt2img_with_restore_faces_performed PASSED                                                                                               [ 48%]
test/test_txt2img.py::test_txt2img_with_vanilla_sampler_performed[PLMS] PASSED                                                                                       [ 51%]
test/test_txt2img.py::test_txt2img_with_vanilla_sampler_performed[DDIM] PASSED                                                                                       [ 55%]
test/test_txt2img.py::test_txt2img_with_vanilla_sampler_performed[UniPC] PASSED                                                                                      [ 58%]
test/test_txt2img.py::test_txt2img_multiple_batches_performed PASSED                                                                                                 [ 62%]
test/test_txt2img.py::test_txt2img_batch_performed PASSED                                                                                                            [ 65%]
test/test_utils.py::test_options_write PASSED                                                                                                                        [ 68%]
test/test_utils.py::test_get_api_url[sdapi/v1/cmd-flags] PASSED                                                                                                      [ 72%]
test/test_utils.py::test_get_api_url[sdapi/v1/samplers] PASSED                                                                                                       [ 75%]
test/test_utils.py::test_get_api_url[sdapi/v1/upscalers] PASSED                                                                                                      [ 79%]
test/test_utils.py::test_get_api_url[sdapi/v1/sd-models] PASSED                                                                                                      [ 82%]
test/test_utils.py::test_get_api_url[sdapi/v1/hypernetworks] PASSED                                                                                                  [ 86%]
test/test_utils.py::test_get_api_url[sdapi/v1/face-restorers] PASSED                                                                                                 [ 89%]
test/test_utils.py::test_get_api_url[sdapi/v1/realesrgan-models] PASSED                                                                                              [ 93%]
test/test_utils.py::test_get_api_url[sdapi/v1/prompt-styles] PASSED                                                                                                  [ 96%]
test/test_utils.py::test_get_api_url[sdapi/v1/embeddings] PASSED                                                                                                     [100%]

====================================================================== 29 passed in 870.87s (0:14:30) ======================================================================

[AUTOMATIC1111]

Some things:

• the way you added dependencies, they are installed for everyone. This shouldn't be the case. If they're needed for Ascend NPU (are they?) they should only be installed if Ascend NPU is present.
• I don't like the added bits of torch.npu.set_device(0) code in multiple places. Can You only do it once, when torch is imported for the first time?

[wangshuai09]

@AUTOMATIC1111 , thanks for you quick response.
I've extracted NPU releated dependency packages into the requirements_npu.txt and these will only be installed in NPU env.
torch.npu.set_device(0) is needed because the context couldn't share in different thread, abstracted this as a function now and will remove after it was fixed.

[wangshuai09]

@AUTOMATIC1111, Hi, I have wrote a wiki about Install and run on Ascend NPUs, Should I update a new wiki page in this repo or link to the Install and run on Ascend NPUs.

[wangshuai09]

@why-not-ava Hi, make sure your code is up to data and then follow the Install and run on Ascend NPUs, the 'webui.sh' will auto install torch_npu

[why-not-ava]

hello, i always get this error with python 3.10.6 when i run [./webui.sh -f --skip-torch-cuda-test --no-half --listen --port 12000], i have no idea how to solve it.

[wangshuai09]

Did you install it using following command? @why-not-ava

conda create -n python310 python=3.10.6
conda activate python310
./webui.sh --skip-torch-cuda-test --no-half

[FuturePrayer]

hi,can the Atlas 200I DK A2 or Orange Pi AIpro run in this way?

[wangshuai09]

@FuturePrayer Hi, it can run mainly based on the ability of torch_npu. You can create a issuse for device support at ascend/pytorch.

[cleansely]

请问下，有测试过sdxl和sd1.5吗，我在800-9000上测试，半精度，xl需要10s

[wangshuai09]

The speed is related to sampling steps, width, height, etc. I've tested and it need 9s using sdxl with 20 sampling steps, 512 witdth, 1024 height.

[cleansely]

@wangshuai09 我这边测试是20step，1024*1024，altas 800环境。您那边应该是altas 800 A2吧，我看代码没有加上torch_npu.npu.set_compile_mode(jit_compile=False)

[CyanWoods]

Did you install it using following command? @why-not-ava

conda create -n python310 python=3.10.6
conda activate python310
./webui.sh --skip-torch-cuda-test --no-half

I come with the same problem, with these commands runned.

[wangshuai09]

@cleansely 我不清楚 altas 800 A2 和 altas 800 的 NPU 型号有区别吗，torch_npu.npu.set_compile_mode(jit_compile=False) 这个参数你设置了吗，对速度影响多大呢