Quickstart

🤗 Optimum Neuron was designed with one goal in mind: to make training and inference straightforward for any 🤗 Transformers user while leveraging the complete power of AWS Accelerators.

Training

There are three main classes one needs to know:

NeuronArgumentParser: inherits the original HfArgumentParser in Transformers with additional checks on the argument values to make sure that they will work well with AWS Trainium instances.
NeuronTrainingArguments: extends the original TrainingArguments with Trainium-specific parameters to optimize performance on AWS Trainium hardware.
NeuronTrainer: the trainer class that takes care of compiling and distributing the model to run on Trainium Chips, and performing training and evaluation.

The NeuronTrainer is very similar to the 🤗 Transformers Trainer, and adapting a script for Trainium instances will mostly consist in:

Swapping the Trainer class for the NeuronTrainer one
Using NeuronTrainingArguments instead of TrainingArguments

That’s how most of the example scripts were adapted from their original counterparts:

-from transformers import TrainingArguments
+from optimum.neuron import NeuronTrainingArguments
-from transformers import Trainer
+from optimum.neuron import NeuronTrainer as Trainer
-training_args = TrainingArguments(
+training_args = NeuronTrainingArguments(
  # training arguments...
)
# A lot of code here
# Initialize our Trainer
trainer = Trainer(
    model=model,
    args=training_args,
    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,
)

All Trainium instances come at least with 2 Neuron Cores. To leverage those we need to launch the training with torchrun. Below you see an example of how to launch a training script on a trn1.2xlarge instance using a bert-base-uncased model.

torchrun --nproc_per_node=2 huggingface-neuron-samples/text-classification/run_glue.py \
--model_name_or_path bert-base-uncased \
--dataset_name philschmid/emotion \
--do_train \
--do_eval \
--bf16 True \
--per_device_train_batch_size 16 \
--learning_rate 5e-5 \
--num_train_epochs 3 \
--output_dir ./bert-emotion

Inference

You can compile and export your 🤗 Transformers models to a serialized format before inference on Neuron devices:

optimum-cli export neuron \
  --model distilbert-base-uncased-finetuned-sst-2-english \
  --batch_size 1 \
  --sequence_length 32 \
  --auto_cast matmul \
  --auto_cast_type bf16 \
  distilbert_base_uncased_finetuned_sst2_english_neuron/

The command above will export distilbert-base-uncased-finetuned-sst-2-english with static shapes: batch_size=1 and sequence_length=32, and cast all matmul operations from FP32 to BF16. Check out the exporter guide for more compilation options.

Then you can run the exported Neuron model on Neuron devices with NeuronModelForXXX classes which are similar to AutoModelForXXX classes in 🤗 Transformers:

from transformers import AutoTokenizer
-from transformers import AutoModelForSequenceClassification
+from optimum.neuron import NeuronModelForSequenceClassification

# PyTorch checkpoint
-model = AutoModelForSequenceClassification.from_pretrained("distilbert-base-uncased-finetuned-sst-2-english")
+model = NeuronModelForSequenceClassification.from_pretrained("distilbert_base_uncased_finetuned_sst2_english_neuron")

tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased-finetuned-sst-2-english")
inputs = tokenizer("Hamilton is considered to be the best musical of past years.", return_tensors="pt")

logits = model(**inputs).logits
print(model.config.id2label[logits.argmax().item()])
# 'POSITIVE'