Optimum for Intel Gaudi - a.k.a. optimum-habana - is the interface between the Transformers and Diffusers libraries and Intel Gaudi AI Accelerators (HPU).
It provides a set of tools enabling easy model loading, training and inference on single- and multi-HPU settings for different downstream tasks.
The list of officially validated models and tasks is available here. Users can try other of the thousands of Hugging Face models on Intel Gaudi accelerators and tasks with only few changes.
HPUs offer fast model training and inference as well as a great price-performance ratio. Check out this blog post about BLOOM inference and this post benchmarking Intel Gaudi 2 and NVIDIA A100 GPUs for BridgeTower training for concrete examples.
Please refer to the Intel Gaudi AI Accelerator official installation guide.
Tests should be run in a Docker container based on Intel Gaudi Docker images.
The current version has been validated for SynapseAI 1.19.
To install the latest stable release of this package
pip install --upgrade-strategy eager optimum[habana]
The --upgrade-strategy eager option is needed to ensure optimum-habana is upgraded to the latest stable release.
To use the example associated with the latest stable release, run:
git clone https://github.com/huggingface/optimum-habana cd optimum-habana && git checkout v1.15.0with
v1.15.0the version number of this release.
Optimum for Intel Gaudi is a fast-moving project, and you may want to install it from source and get the latest scripts :
pip install git+https://github.com/huggingface/optimum-habana.git
git clone https://github.com/huggingface/optimum-habanaThe transformers_future branch is regularly updated with the latest changes from the main branches of Optimum Habana and Transformers. This enables you to try out new Transformers features that have not been merged into the main branch yet.
Warning
The transformers_future branch may have some regressions or bugs and may be less stable than the main branch.
pip install git+https://github.com/huggingface/optimum-habana.git@transformers_future
git clone -b transformers_future https://github.com/huggingface/optimum-habanaTo use DeepSpeed on HPUs, you also need to run the following command:
pip install git+https://github.com/HabanaAI/[email protected]
To install the requirements for every example:
cd <example-folder> pip install -r requirements.txt
Optimum for Intel Gaudi was designed with one goal in mind: to make training and inference straightforward for Transformers and Diffusers users, while fully leveraging the power of Intel Gaudi AI Accelerators.
There are two main classes one needs to know:
- GaudiTrainer: the trainer class that takes care of compiling and distributing the model to run on HPUs, and performing training and evaluation.
- GaudiConfig: the class that enables to configure Habana Mixed Precision and to decide whether optimized operators and optimizers should be used or not.
The GaudiTrainer is very similar to the Transformers Trainer, and adapting a script using the Trainer to make it work with Intel Gaudi accelerators will mostly consist in simply swapping the Trainer class for the GaudiTrainer one.
That's how most of the example scripts were adapted from their original counterparts.
Here is an example:
- from transformers import Trainer, TrainingArguments
+ from optimum.habana import GaudiConfig, GaudiTrainer, GaudiTrainingArguments
- training_args = TrainingArguments(
+ training_args = GaudiTrainingArguments(
# training arguments...
+ use_habana=True,
+ use_lazy_mode=True, # whether to use lazy or eager mode
+ gaudi_config_name=path_to_gaudi_config,
)
# A lot of code here
# Initialize our Trainer
- trainer = Trainer(
+ trainer = GaudiTrainer(
model=model,
args=training_args, # Original training arguments.
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
compute_metrics=compute_metrics,
tokenizer=tokenizer,
data_collator=data_collator,
)where gaudi_config_name is the name of a model from the Hub (Intel Gaudi configurations are stored in model repositories) or a path to a local Intel Gaudi configuration file (you can see here how to write your own).
You can generate images from prompts using Stable Diffusion on Intel Gaudi using the GaudiStableDiffusionPipeline class and the [GaudiDDIMScheduler] which have been both optimized for HPUs. Here is how to use them and the differences with the Diffusers library:
- from diffusers import DDIMScheduler, StableDiffusionPipeline
+ from optimum.habana.diffusers import GaudiDDIMScheduler, GaudiStableDiffusionPipeline
model_name = "CompVis/stable-diffusion-v1-4"
- scheduler = DDIMScheduler.from_pretrained(model_name, subfolder="scheduler")
+ scheduler = GaudiDDIMScheduler.from_pretrained(model_name, subfolder="scheduler")
- pipeline = StableDiffusionPipeline.from_pretrained(
+ pipeline = GaudiStableDiffusionPipeline.from_pretrained(
model_name,
scheduler=scheduler,
+ use_habana=True,
+ use_hpu_graphs=True,
+ gaudi_config="Habana/stable-diffusion",
)
outputs = generator(
["An image of a squirrel in Picasso style"],
num_images_per_prompt=16,
+ batch_size=4,
)With the upgrade to PyTorch 2.5, users may experience some performance degradation due to changes in the handling of FP16/BF16 inputs. The note from PyTorch 2.5 states:
"A naive SDPA math backend, when using FP16/BF16 inputs, can accumulate significant numerical errors due to the usage of low-precision intermediate buffers. To mitigate this issue, the default behavior now involves upcasting FP16/BF16 inputs to FP32. Computations are performed in FP32/TF32, and the final FP32 results are then downcasted back to FP16/BF16. This will improve numerical accuracy of the final output for the math backend with FP16/BF16 inputs, but increases memory usages and may cause the performance regressions in the math backend as computations shift from FP16/BF16 BMM to FP32/TF32 BMM/Matmul."
For scenarios where reduced-precision reductions are preferred for speed, they can be enabled with the following setting:
torch.backends.cuda.allow_fp16_bf16_reduction_math_sdp(True)Additionally, the next release of Optimum Habana will include a Gaudi-specific safe_softmax implementation that will also improve performance.
More info:
Check out the documentation of Optimum for Intel Gaudi for more advanced usage.
The following model architectures, tasks and device distributions have been validated for Optimum for Intel Gaudi:
In the tables below, ✔️ means single-card, multi-card and DeepSpeed have all been validated.
- Transformers:
- Diffusers:
| Architecture | Training | Inference | Tasks |
|---|---|---|---|
| Stable Diffusion | |||
| Stable Diffusion XL | |||
| Stable Diffusion Depth2img | |||
| LDM3D | |||
| FLUX.1 | |||
| Text to Video |
- PyTorch Image Models/TIMM:
| Architecture | Training | Inference | Tasks |
|---|---|---|---|
| FastViT |
- TRL:
| Architecture | Training | Inference | Tasks |
|---|---|---|---|
| Llama 2 | ✔️ | ||
| Llama 2 | ✔️ | ||
| Stable Diffusion | ✔️ |
Other models and tasks supported by the Transformers and Diffusers libraries may also work. You can refer to this section for using them with Optimum for Intel Gaudi. In addition, this page explains how to modify any example from the Transformers library to make it work with Optimum for Intel Gaudi.
If you find any issues while using those, please open an issue or a pull request.
After training your model, feel free to submit it to the Intel leaderboard which is designed to evaluate, score, and rank open-source LLMs that have been pre-trained or fine-tuned on Intel Hardwares. Models submitted to the leaderboard will be evaluated on the Intel Developer Cloud. The evaluation platform consists of Gaudi Accelerators and Xeon CPUs running benchmarks from the Eleuther AI Language Model Evaluation Harness.
Check the contributor guide for instructions.
