🤗 Optimum Graphcore was designed with one goal in mind: to make training and evaluation straightforward for any 🤗 Transformers user while leveraging the complete power of IPUs.
To install the latest release of 🤗 Optimum Graphcore:
pip install optimum-graphcore
Optimum Graphcore is a fast-moving project, and you may want to install from source.
pip install git+https://github.com/huggingface/optimum-graphcore.git
You need to have the Poplar SDK enabled to use 🤗 Optimum Graphcore. Refer to the Getting Started guide for your system for details on how to enable the Poplar SDK. Also refer to the Jupyter Quick Start guide for how to set up Jupyter to be able to run notebooks on a remote IPU machine.
We have a range of example scripts in /examples
and Jupyter notebooks in /notebooks
which you can refer to to see how specific elements of the Optimum Graphcore library are used.
The main Optimum Graphcore classes are:
IPUTrainer
section for more information.IPUConfig
section for more information.IPUTrainer is very similar to the 🤗 Transformers Trainer class, and you can easily adapt a script that currently uses Trainer
to make it work with IPUs. You will mostly swap Trainer
for IPUTrainer. This is how most of the Optimum Graphcore example scripts were adapted from the original Hugging Face scripts.
-from transformers import Trainer, TrainingArguments
+from optimum.graphcore import IPUConfig, IPUTrainer, IPUTrainingArguments
-training_args = TrainingArguments(
+training_args = IPUTrainingArguments(
per_device_train_batch_size=4,
learning_rate=1e-4,
+ # Any IPUConfig on the Hub or stored locally
+ ipu_config_name="Graphcore/bert-base-ipu",
+)
+
+# Loading the IPUConfig needed by the IPUTrainer to compile and train the model on IPUs
+ipu_config = IPUConfig.from_pretrained(
+ training_args.ipu_config_name,
)
# Initialize our Trainer
-trainer = Trainer(
+trainer = IPUTrainer(
model=model,
+ ipu_config=ipu_config,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
... # Other arguments
We also support the pipeline
API, so you can easily run a model on a given input, for example text image or audio, on IPUs:
->>> from transformers import pipeline
+>>> from optimum.graphcore import pipeline
# Allocate a pipeline for sentiment-analysis
->>> classifier = pipeline('sentiment-analysis', model="distilbert-base-uncased-finetuned-sst-2-english")
+>>> classifier = pipeline('sentiment-analysis', model="distilbert-base-uncased-finetuned-sst-2-english", ipu_config = "Graphcore/distilbert-base-ipu")
>>> classifier('We are very happy to introduce pipeline to the transformers repository.')
[{'label': 'POSITIVE', 'score': 0.9996947050094604}]