Addressing compatibility points throughout set up | ONNX for NVIDIA GPUs | Hugging Face’s Optimum library
11 hours in the past
This text discusses the ONNX runtime, one of the efficient methods of rushing up Steady Diffusion inference. On an A100 GPU, working SDXL for 30 denoising steps to generate a 1024 x 1024 picture could be as quick as 2 seconds. Nevertheless, the ONNX runtime relies on a number of transferring items, and putting in the proper variations of all of its dependencies could be difficult in a always evolving ecosystem. Take this as a high-level debugging information, the place I share my struggles in hopes of saving you time. Whereas the particular variations and instructions may shortly grow to be out of date, the high-level ideas ought to stay related for an extended time frame.
ONNX can really refer to 2 completely different (however associated) elements of the ML stack:
- ONNX is a format for storing machine studying fashions. It stands for Open Neural Community Trade and, as its identify suggests, its most important purpose is interoperability throughout platforms. ONNX is a self-contained format: it shops each the mannequin weights and structure. Which means a single .onnx file accommodates all the data wanted to run inference. No want to write down any further code to outline or load a mannequin; as an alternative, you merely go it to a runtime (extra on this beneath).
- ONNX can be a runtime to run mannequin which might be in ONNX format. It actually runs the mannequin. You’ll be able to see it as a mediator between the architecture-agnostic ONNX format and the precise {hardware} that runs inference. There’s a separate model of the runtime for every supported accelerator kind (see full listing right here). Observe, nevertheless, that the ONNX runtime will not be the one strategy to run inference with a mannequin that’s in ONNX format — it’s only one manner. Producers can select to construct their very own runtimes which might be hyper-optimized for his or her {hardware}. As an illustration, NVIDIA’s TensorRT is an alternative choice to the ONNX runtime.
This text focuses on working Steady Diffusion fashions utilizing the ONNX runtime. Whereas the high-level ideas are most likely timeless, notice that the ML tooling ecosystem is in fixed change, so the precise workflow or code snippets may grow to be out of date (this text was written in Could 2024). I’ll deal with the Python implementation specifically, however notice that the ONNX runtime may function in different languages like C++, C#, Java or JavaScript.
Professionals of the ONNX Runtime
- Stability between inference pace and interoperability. Whereas the ONNX runtime won’t all the time be the quickest resolution for every type of {hardware}, it’s a quick sufficient resolution for most sorts of {hardware}. That is notably interesting when you’re serving your fashions on a heterogeneous fleet of machines and don’t have the assets to micro-optimize for every completely different accelerator.
- Huge adoption and dependable authorship. ONNX was open-sourced by Microsoft, who’re nonetheless sustaining it. It’s broadly adopted and effectively built-in into the broader ML ecosystem. As an illustration, Hugging Face’s Optimum library means that you can outline and run ONNX mannequin pipelines with a syntax that’s paying homage to their well-liked transformers and diffusers libraries.
Cons of the ONNX Runtime
- Engineering overhead. In comparison with the choice of working inference immediately in PyTorch, the ONNX runtime requires compiling your mannequin to the ONNX format (which might take 20–half-hour for a Steady Diffusion mannequin) and putting in the runtime itself.
- Restricted set of ops. The ONNX format doesn’t help all PyTorch operations (it’s much more restrictive than TorchScript). In case your mannequin is utilizing an unsupported operation, you’ll both must reimplement the related portion, or drop ONNX altogether.
- Brittle set up and setup. For the reason that ONNX runtime makes the interpretation from the ONNX format to architecture-specific directions, it may be difficult to get the proper mixture of software program variations to make it work. As an illustration, if working on an NVIDIA GPU, you should guarantee compatibility of (1) working system, (2) CUDA model, (3) cuDNN model, and (4) ONNX runtime model. There are helpful assets just like the CUDA compatibility matrix, however you may nonetheless find yourself losing hours discovering the magic mixture that works at a given cut-off date.
- {Hardware} limitations. Whereas the ONNX runtime can run on many architectures, it can not run on all architectures like pure PyTorch fashions can. As an illustration, there may be at the moment (Could 2024) no help for Google Cloud TPUs or AWS Inferentia chips (see FAQ).
At first look, the listing of cons seems longer than the listing of professionals, however don’t be discouraged — as proven afterward, the enhancements in mannequin latency could be vital and price it.
Possibility #1: Set up from supply
As talked about above, the ONNX runtime requires compatibility between many items of software program. If you wish to be on the innovative, one of the simplest ways to get the most recent model is to comply with the directions within the official Github repository. For Steady Diffusion specifically, this folder accommodates set up directions and pattern scripts for producing photos. Count on constructing from supply to take fairly some time (round half-hour).
On the time of writing (Could 2024), this resolution labored seamlessly for me on an Amazon EC2 occasion (g5.2xlarge, which comes with a A10G GPU). It avoids compatibility points mentioned beneath by utilizing a Docker picture that comes with the proper dependencies.
Possibility #2: Set up through PyPI
In manufacturing, you’ll probably need a steady model of the ONNX runtime from PyPI, as an alternative of putting in the most recent model from supply. For Python specifically, there are two completely different libraries (one for CPU and one for GPU). Right here is the command to put in it for CPU:
pip set up onnxruntime
And right here is the command to put in it for GPU:
pip set up onnxruntime-gpu
It is best to by no means set up each. Having them each may result in error messages or behaviors that aren’t straightforward to trace again to this root trigger. The ONNX runtime may merely fail to acknowledge the presence of the GPU, which can look stunning on condition that onnxruntime-gpu is certainly put in.
In an excellent world, pip set up onnxruntime-gpu could be the top of the story. Nevertheless, in follow, there are robust compatibility necessities between different items of software program in your machine, together with the working system, the hardware-specific drivers, and the Python model.
Say that you just wish to use the most recent model of the ONNX runtime (1.17.1) on the time of writing. So what stars do we have to align to make this occur?
Listed below are a number of the commonest sources of incompatibility that may enable you arrange your surroundings. The precise particulars will shortly grow to be out of date, however the high-level concepts ought to proceed to use for some time.
CUDA compatibility
In case you are not planning on utilizing an NVIDIA GPU, you may skip this part. CUDA is a platform for parallel computing that sits on high of NVIDIA GPUs, and is required for machine studying workflows. Every model of the ONNX runtime is appropriate with solely sure CUDA variations, as you may see in this compatibility matrix.
In line with this matrix, the most recent ONNX runtime model (1.17) is appropriate with each CUDA 11.8 and CUDA 12. However you should take note of the nice print: by default, ONNX runtime 1.17 expects CUDA 11.8. Nevertheless, most VMs at this time (Could 2024) include CUDA 12.1 (you may test the model by working nvcc --version). For this explicit setup, you’ll have to exchange the standard pip set up onnxruntime-gpu with:
pip set up onnxruntime-gpu==1.17.1 --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/onnxruntime-cuda-12/pypi/easy/
Observe that, as an alternative of being on the mercy of no matter CUDA model occurs to be put in in your machine, a cleaner resolution is to do your work from inside a Docker container. You merely select the picture that has your required model of Python and CUDA. As an illustration:
docker run --rm -it --gpus all nvcr.io/nvidia/pytorch:23.10-py3
OS + Python + pip compatibility
This part discusses compatibility points which might be architecture-agnostic (i.e. you’ll encounter them whatever the goal accelerator). It boils down to creating positive that your software program (working system, Python set up and pip set up) are appropriate along with your desired model of the ONNX runtime library.
Pip model: Except you’re working with legacy code or programs, your most secure wager is to improve pip to the most recent model:
python -m pip set up --upgrade pip
Python model: As of Could 2024, the Python model that’s least seemingly to present you complications is 3.10 (that is what most VMs include by default). Once more, until you’re working with legacy code, you definitely need a minimum of 3.8 (since 3.7 was deprecated in June 2023).
Working system: The truth that the OS model may hinder your capacity to put in the specified library got here as a shock to me, particularly that I used to be utilizing essentially the most commonplace EC2 cases. And it wasn’t simple to determine that the OS model was the perpetrator.
Right here I’ll stroll you thru my debugging course of, within the hopes that the workflow itself is longer-lived than the specifics of the variations at this time. First, I put in onnxruntime-gpu with the next command (since I had CUDA 12.1 put in on my machine):
pip set up onnxruntime-gpu --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/onnxruntime-cuda-12/pypi/easy/
On the floor, this could set up the most recent model of the library obtainable on PyPI. In actuality nevertheless, this can set up the most recent model appropriate along with your present setup (OS + Python model + pip model). For me on the time, that occurred to be onnxruntime-gpu==1.16.0. (versus 1.17.1, which is the most recent). Unknowingly putting in an older model merely manifested within the ONNX runtime being unable to detect the GPU, with no different clues. After considerably by chance discovering the model is older than anticipated, I explicitly requested for the newer one:
pip set up onnxruntime-gpu==1.17.1 --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/onnxruntime-cuda-12/pypi/easy/
This resulted in a message from pip complaining that the model I requested will not be really obtainable (regardless of being listed on PyPI):
ERROR: Couldn't discover a model that satisfies the requirement onnxruntime-gpu==1.17.1 (from variations: 1.12.0, 1.12.1, 1.13.1, 1.14.0, 1.14.1, 1.15.0, 1.15.1, 1.16.0, 1.16.1, 1.16.2, 1.16.3)
ERROR: No matching distribution discovered for onnxruntime-gpu==1.17.1
To know why the most recent model will not be getting put in, you may go a flag that makes pip verbose: pip set up ... -vvv. This reveals all of the Python wheels that pip cycles by with a view to discover the latest one that’s appropriate to your system. Here’s what the output appeared like for me:
Skipping hyperlink: not one of the wheel's tags (cp35-cp35m-manylinux1_x86_64) are appropriate (run pip debug --verbose to point out appropriate tags): https://information.pythonhosted.org/packages/26/1a/163521e075d2e0c3effab02ba11caba362c06360913d7c989dcf9506edb9/onnxruntime_gpu-0.1.2-cp35-cp35m-manylinux1_x86_64.whl (from https://pypi.org/easy/onnxruntime-gpu/)
Skipping hyperlink: not one of the wheel's tags (cp36-cp36m-manylinux1_x86_64) are appropriate (run pip debug --verbose to point out appropriate tags): https://information.pythonhosted.org/packages/52/f2/30aaa83bc9e90e8a919c8e44e1010796eb30f3f6b42a7141ffc89aba9a8e/onnxruntime_gpu-0.1.2-cp36-cp36m-manylinux1_x86_64.whl (from https://pypi.org/easy/onnxruntime-gpu/)
Skipping hyperlink: not one of the wheel's tags (cp37-cp37m-manylinux1_x86_64) are appropriate (run pip debug --verbose to point out appropriate tags): https://information.pythonhosted.org/packages/a2/05/af0481897255798ee57a242d3989427015a11a84f2eae92934627be78cb5/onnxruntime_gpu-0.1.2-cp37-cp37m-manylinux1_x86_64.whl (from https://pypi.org/easy/onnxruntime-gpu/)
Skipping hyperlink: not one of the wheel's tags (cp35-cp35m-manylinux1_x86_64) are appropriate (run pip debug --verbose to point out appropriate tags): https://information.pythonhosted.org/packages/17/cb/0def5a44db45c6d38d95387f20057905ce2dd4fad35c0d43ee4b1cebbb19/onnxruntime_gpu-0.1.3-cp35-cp35m-manylinux1_x86_64.whl (from https://pypi.org/easy/onnxruntime-gpu/)
Skipping hyperlink: not one of the wheel's tags (cp36-cp36m-manylinux1_x86_64) are appropriate (run pip debug --verbose to point out appropriate tags): https://information.pythonhosted.org/packages/a6/53/0e733ebd72d7dbc84e49eeece15af13ab38feb41167fb6c3e90c92f09cbb/onnxruntime_gpu-0.1.3-cp36-cp36m-manylinux1_x86_64.whl (from https://pypi.org/easy/onnxruntime-gpu/)
...
The tags listed in brackets are Python platform compatibility tags, and you may learn extra about them right here. In a nutshell, each Python wheel comes with a tag that signifies what system it might run on. As an illustration, cp35-cp35m-manylinux1_x86_64 requires CPython 3.5, a set of (older) Linux distributions that fall beneath the manylinux1 umbrella, and a 64-bit x86-compatible processor.
Since I needed to run Python 3.10 on a Linux machine (therefore filtering for cp310.*manylinux.*, I used to be left with a single potential wheel for the onnxruntime-gpu library, with the next tag:
cp310-cp310-manylinux_2_28_x86_64
You will get an inventory of tags which might be appropriate along with your system by working pip debug --verbose. Here’s what a part of my output appeared like:
cp310-cp310-manylinux_2_26_x86_64
cp310-cp310-manylinux_2_25_x86_64
cp310-cp310-manylinux_2_24_x86_64
cp310-cp310-manylinux_2_23_x86_64
cp310-cp310-manylinux_2_22_x86_64
cp310-cp310-manylinux_2_21_x86_64
cp310-cp310-manylinux_2_20_x86_64
cp310-cp310-manylinux_2_19_x86_64
cp310-cp310-manylinux_2_18_x86_64
cp310-cp310-manylinux_2_17_x86_64
...
In different phrases, my working system is only a tad too outdated (the utmost linux tag that it helps is manylinux_2_26, whereas the onnxruntime-gpu library’s solely Python 3.10 wheel requires manylinux_2_28. Upgrading from Ubuntu 20.04 to Ubuntu 24.04 solved the issue.
As soon as the ONNX runtime is (lastly) put in, producing photos with Steady Diffusion requires two following steps:
- Export the PyTorch mannequin to ONNX (this will take > half-hour!)
- Go the ONNX mannequin and the inputs (textual content immediate and different parameters) to the ONNX runtime.
Possibility #1: Utilizing official scripts from Microsoft
As talked about earlier than, utilizing the official pattern scripts from the ONNX runtime repository labored out of the field for me. In case you comply with their set up directions, you received’t even must cope with the compatibility points talked about above. After set up, producing a picture is a straightforward as:
python3 demo_txt2img_xl.py "starry night time over Golden Gate Bridge by van gogh"
Underneath the hood, this script defines an SDXL mannequin utilizing Hugging Face’s diffusers library, exports it to ONNX format (which might take as much as half-hour!), then invokes the ONNX runtime.
Possibility #2: Utilizing Hugging Face’s Optimum library
The Optimum library guarantees numerous comfort, permitting you to run fashions on varied accelerators whereas utilizing the acquainted pipeline APIs from the well-known transformers and diffusers libraries. For ONNX specifically, that is what inference code for SDXL seems like (extra in this tutorial):
from optimum.onnxruntime import ORTStableDiffusionXLPipelinemodel_id = "stabilityai/stable-diffusion-xl-base-1.0"
base = ORTStableDiffusionXLPipeline.from_pretrained(model_id)
immediate = "crusing ship in storm by Leonardo da Vinci"
picture = base(immediate).photos[0]
# Remember to save lots of the ONNX mannequin
save_directory = "sd_xl_base"
base.save_pretrained(save_directory)
In follow, nevertheless, I struggled so much with the Optimum library. First, set up is non-trivial; naively following the set up instruction within the README file will run into the incompatibility points defined above. This isn’t Optimum’s fault per se, nevertheless it does add yet one more layer of abstraction on high of an already brittle setup. The Optimum set up may pull a model of onnxruntime that’s conflicting along with your setup.
Even after I received the battle in opposition to compatibility points, I wasn’t capable of run SDXL inference on GPU utilizing Optimum’s ONNX interface. The code snippet above (immediately taken from a Hugging Face tutorial) fails with some form mismatches, maybe because of bugs within the PyTorch → ONNX conversion:
[ONNXRuntimeError] : 1 : FAIL : Non-zero standing code returned whereas working Add node.
Identify:'/down_blocks.1/attentions.0/Add'
Standing Message: /down_blocks.1/attentions.0/Add: left operand can not broadcast on dim 3 LeftShape: {2,64,4096,10}, RightShape: {2,640,64,64}
For a short second I thought of stepping into the weeds and debugging the Hugging Face code (a minimum of it’s open supply!), however gave up once I realized that Optimum has a backlog of greater than 250 points, with points going for weeks with no acknowledgement from the Hugging Face staff. I made a decision to maneuver on and easily use Microsoft’s official scripts as an alternative.
As promised, the trouble to get the ONNX runtime working is price it. On an A100 GPU, the inference time is decreased from 7–8 seconds (when working vanilla PyTorch) to ~2 seconds. That is similar to TensorRT (an NVIDIA-specific various to ONNX), and about 1 second sooner than torch.compile (PyTorch’s native JIT compilation).
Reportedly, switching to much more performant GPUs (e.g. H100) can result in even larger positive aspects from working your mannequin with a specialised runtime.
The ONNX runtime guarantees vital latency positive aspects, nevertheless it comes with non-trivial engineering overhead. It additionally faces the traditional trade-off for static compilation: inference is so much sooner, however the graph can’t be dynamically modified (which is at odds with dynamic adapters like peft). The ONNX runtime and related compilation strategies are price including to your pipeline when you’ve handed the experimentation part, and are able to spend money on environment friendly manufacturing code.
In case you’re taken with optimizing inference time, listed here are some articles that I discovered useful:
