feat(docs): add platform/model compatibility list
This commit is contained in:
parent
6b2ae2aeab
commit
b22f15600b
|
@ -1,5 +1,7 @@
|
|||
# Very Rough Benchmarks
|
||||
|
||||
CUDA > ROCm > DirectML > drawing it yourself > CPU
|
||||
|
||||
Using 25 steps of Euler A in txt2img, 512x512.
|
||||
|
||||
- CPU:
|
||||
|
@ -9,6 +11,8 @@ Using 25 steps of Euler A in txt2img, 512x512.
|
|||
- 7950X: 3.5s/it, 90sec/image
|
||||
- GPU:
|
||||
- AMD:
|
||||
- 6900XT: 3.5it/s, 9sec/image
|
||||
- 6900XT
|
||||
- Win10, DirectML: 3.5it/s, 9sec/image
|
||||
- Ubuntu 20.04, ROCm 5.2: 4.5it/s, 6sec/image
|
||||
- Nvidia:
|
||||
- 4090: 6.5it/s, 4sec/image
|
||||
|
|
|
@ -0,0 +1,55 @@
|
|||
# Compatibility
|
||||
|
||||
## Contents
|
||||
|
||||
- [Compatibility](#compatibility)
|
||||
- [Contents](#contents)
|
||||
- [Driver Versions](#driver-versions)
|
||||
- [Container/Platform Acceleration](#containerplatform-acceleration)
|
||||
- [Container Notes](#container-notes)
|
||||
- [Model/Platform Acceleration](#modelplatform-acceleration)
|
||||
- [Model Notes](#model-notes)
|
||||
|
||||
## Driver Versions
|
||||
|
||||
- CUDA
|
||||
- 11.6
|
||||
- 11.7
|
||||
- ROCm
|
||||
- 5.2
|
||||
- 5.4 seems like it might work
|
||||
|
||||
## Container/Platform Acceleration
|
||||
|
||||
| Runtime | CUDA | DirectML | ROCm | CPU |
|
||||
| ------- | ---- | -------- | -------- | --- |
|
||||
| docker | yes | no, 1 | maybe, 2 | yes |
|
||||
| podman | 3 | no, 1 | maybe, 2 | yes |
|
||||
|
||||
### Container Notes
|
||||
|
||||
1. no package available: https://github.com/ssube/onnx-web/issues/63
|
||||
2. should work but testing failed: https://github.com/ssube/onnx-web/issues/10
|
||||
3. should work, not tested: https://gist.github.com/bernardomig/315534407585d5912f5616c35c7fe374
|
||||
|
||||
## Model/Platform Acceleration
|
||||
|
||||
| Model | CUDA | DirectML | ROCm | CPU |
|
||||
| ---------------- | ---- | -------- | ----- | --- |
|
||||
| Stable Diffusion | yes | yes | yes | yes |
|
||||
| Real ESRGAN | yes | yes | no, 1 | yes |
|
||||
| GFPGAN | 2 | 2 | 2 | yes |
|
||||
|
||||
### Model Notes
|
||||
|
||||
1. Real ESRGAN running on ROCm crashes with an error:
|
||||
|
||||
```none
|
||||
File "/home/ssube/onnx-web/api/onnx_web/upscale.py", line 67, in __call__
|
||||
output = self.session.run([output_name], {
|
||||
File "/home/ssube/onnx-web/api/onnx_env/lib/python3.8/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 200, in run
|
||||
return self._sess.run(output_names, input_feed, run_options)
|
||||
onnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL : Non-zero status code returned while running FusedConv node. Name:'/body/body.0/rdb1/conv1/Conv' Status Message: MIOPEN failure 1: miopenStatusNotInitialized ; GPU=0 ; hostname=ssube-notwin ; expr=status_;
|
||||
```
|
||||
|
||||
2. GFPGAN seems to always be running in CPU mode
|
|
@ -3,14 +3,16 @@
|
|||
This is the user guide for ONNX web, a web GUI for running ONNX models with hardware acceleration on both AMD and Nvidia
|
||||
system, with a CPU software fallback.
|
||||
|
||||
The API runs on both Linux and Windows and provides access to the major functionality of diffusers, along with metadata
|
||||
about the available models and accelerators, and the output of previous runs. Hardware acceleration is supported on both
|
||||
AMD and Nvidia for both Linux and Windows, with a CPU fallback capable of running on laptop-class machines.
|
||||
The API is written in Python and runs on both Linux and Windows and provides access to the major functionality of
|
||||
diffusers, along with metadata about the available models and accelerators, and the output of previous runs. Hardware
|
||||
acceleration is supported on both AMD and Nvidia for both Linux and Windows, with a CPU fallback capable of running on
|
||||
laptop-class machines.
|
||||
|
||||
The GUI is hosted on Github Pages and runs in all major browsers, including on mobile devices. It allows you to select
|
||||
the model and accelerator being used for each image pipeline. Image parameters are shown for each of the major modes,
|
||||
and you can either upload or paint the mask for inpainting and outpainting. The last few output images are shown below
|
||||
the image controls, making it easy to refer back to previous parameters or save an image from earlier.
|
||||
The GUI is written in Javascript, hosted on Github Pages, and runs in all major browsers, including on mobile devices.
|
||||
It allows you to select the model and accelerator being used for each image pipeline. Image parameters are shown for
|
||||
each of the major modes, and you can either upload or paint the mask for inpainting and outpainting. The last few output
|
||||
images are shown below the image controls, making it easy to refer back to previous parameters or save an image from
|
||||
earlier.
|
||||
|
||||
Please see [the server admin guide](server-admin.md) for details on how to configure and run the server.
|
||||
|
||||
|
@ -92,7 +94,7 @@ will need a simple drawing component, but anything more complicated, like layers
|
|||
the Gimp, Krita, or Photoshop.
|
||||
|
||||
This is _not_ a tool for building new ML models. While I am open to some training features, like Dreambooth and anything
|
||||
needed to convert models, that is not the focus and should be limited features that support the other tabs.
|
||||
needed to convert models, that is not the focus and should be limited to features that support the other tabs.
|
||||
|
||||
### ONNX models
|
||||
|
||||
|
|
Loading…
Reference in New Issue