Note: this repository is under experimental status. We don't yet have correctness guarantees for the model implementations.

TorchPrime is a reference model implementation for PyTorch on TPU/GPU using torch_xla and torchax. It is designed to showcase best practices for high-performance model training with these frameworks.

Here is a simple example of training on a single TPU VM. It assumes that you have already installed torch_xla 1 and torchax 2 following their respective project READMEs.

Install torchprime:

git clone https://github.com/AI-Hypercomputer/torchprime.git cd torchprime pip install -e '.[dev]'

Train Llama 3 8B using torch_xla:

export HF_TOKEN='...your huggingface token...' XLA_IR_DEBUG=1 XLA_HLO_DEBUG=1 python3 torchprime/torch_xla_models/train.py

Train Llama 3 8B using torchax:

python3 torchprime/experimental/torchax_models/run.py global_batch_size=16

Refer to README.md in torchprime/torch_xla_models and torchprime/experimental/torchax_models for more details.

torchprime uses hydra to read configurations (e.g. model name, batch size) from the command line and .yaml files.

In both torch_xla_models and torchax_models directories, you'll find a configs/default.yaml. That specifies the default configuration for the trainer. You may override configs on the command line with a key=value syntax. For example, the following command will train Mixtral 8x7B with a global batch size of 256, and set the FSDP SPMD ICI mesh axis length to 64:

python3 torchprime/torch_xla_models/train.py \ model=mixtral-8x7b \ global_batch_size=256 \ ici_mesh.fsdp=64

You may refer to the hydra docs for other ways to specify configs.

torchprime uses xpk as the standard path for iterating on distributed training code.

First teach torchprime about the XPK cluster it is using, the artifact storage location, etc. You only need to do this on first clone or when switching to a different topology or cluster. Example:

tp use \ --cluster <XPK CLUSTER NAME> \ --project my-gcp-project \ --zone us-east5-b \ --num-slices 1 \ --tpu-type v6e-256 \ --artifact-dir gs://bucket/dir

Then prepend tp run to a particular Python file you would like to run remotely, including arguments, e.g.

torch_xla example:

# Train Llama 3.0 8B on 256 chips tp run torchprime/torch_xla_models/train.py \ model=llama-3-8b \ global_batch_size=256 \ ici_mesh.fsdp=256

torchax example:

tp run torchprime/experimental/torchax_models/run.py global_batch_size=256

tp run will broadcast the specified command to all VMs in the XPK cluster, which is the convention for running SPMD distributed workloads. See tp run --help for more advanced features.

tp run will pick up these environment variables locally and proxy them to the distributed workload, if found:

• HF_TOKEN: HuggingFace token
• XLA_IR_DEBUG: torch_xla debugging flag
• XLA_HLO_DEBUG: torch_xla debugging flag
• LIBTPU_INIT_ARGS: XLA flags that affect compilation and execution behavior

Here are the status of various models. In general, there are five stages for each model:

• TODO: We need to implement the model.
• Implemented: The model runs either a training or an inference step.
• Optimized: We found the best scaling configuration for the model on one or more hardware. One-off performance data is available.
• Convergence: We tested that the training loss converges to a reasonable value, or that the loss curve tracks an existing reference if exists.
• Production: Not only is the model optimized and converges, its performance is also continuously monitored. This is a good state for using the model in production.

All implemented models will at least have unit tests to verify basic numerical correctness, and the convergence verification stage serves as an additional correctness guarantee.

If a model is at least implemented, you'll also find a training recipe linked from the checkmark emoji in the table. If a model is optimized, you'll also find MFU numbers linked from the table. Note that a model may continue to receive ongoing optimization thereafter.

This repo will contain a set of reference models that we have optimized and runs well on TPU. The best performing scaling configuration (parallelism techniques, checkpointing, etc.) for a model on various hardwares will be provided for ease of reproducibility.

docs contains guides for optimizing performance and debugging issues.

torchprime/launcher contains scripts to train a model on a large TPU cluster.

torchprime/data contains dataset and data loading utilities.

torchprime/torch_xla_models contains model implementations using torch_xla.

torchprime/experimental/torchax_models contains model implementations using torchax.

Finally, each model may also provide a GPU "original" version that illustrates and attributes where this model code came from, if any. This also helps to show case what changes we have done to make it performant on TPU. The original version is not expected to be run.

Contributions are welcome! Please feel free to submit a pull request.

When developing, use pip install -e '.[dev]' to install dev dependencies such as linter and formatter.

How to run tests:

pytest

How to run some of the tests, and re-run them whenever you change a file:

tp -i test ... # replace with path to tests/directories

How to format:

ruff format

How to lint:

ruff check [--fix]

You can install a Ruff VSCode plugin to check errors and format files from the editor.

Torchprime supports running with user specified torch and torch_xla wheels placed under local_dist/ directory. The wheel will be automatically installed in the docker image when use tp run command. To use the wheel, add flag --use-local-wheel to tp run command:

tp run --use-local-wheel torchprime/hf_models/train.py

The wheels should be built inside a PyTorch/XLA development docker image or the PyTorch/XLA VSCode Dev Container to minimize compatibility issues.

This project is licensed under the New BSD License - see the LICENSE file for details.

For more information on PyTorch/XLA, visit the official documentation.