torchprime

High-performance training for PyTorch on Cloud TPU

torchprime is a reference implementation for training PyTorch models on TPU. It is designed to showcase best practices for large-scale, high-performance model training using torch_xla (project), with minimal changes to model code. It aims to demystify training on XLA-based accelerators, providing clear patterns and best practices to help the PyTorch community unlock top performance and efficiency on Google Cloud TPUs.

torchprime is under active development, and we're eager for feedback and input from the PyTorch community.

Installation

Before installing torchprime, you will need to first install torch_xla following its respective project README.

Install torchprime:

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

Examples

Local training

Here is a simple example of training on a single TPU VM. 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

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

Configuring training

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

In the torch_xla_models directory, 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.

Distributed training

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

torchprime natively supports multi-slice or multi-pod training. --num-slices specifies the number of slices used by the workload. --tpu-type specifies the accelerator type in each slice. To do multi-pod training, simply specify a --tpu-type that is as big as a pod.

After configuring the cluster, prepend tp run to a particular Python file you would like to run remotely, including arguments, e.g.

# 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

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.

Env vars passed to the workload

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

Additional CLI arguments passed to the workload

Besides forwarding your command line arguments, tp run will add:

• profile_dir=[...]: path to a profile directory accessible by the workload

Supported Models

Below are the status of various models. There are five stages for each model:

1. TODO: We need to implement the model.
2. Implemented: The model runs either a training or an inference step.
3. Optimized: We found the best scaling configuration for the model on one or more hardware. One-off performance data is available.
4. Convergence: We tested that the training loss converges to a reasonable value, or that the loss curve tracks an existing reference if exists.
5. 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 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.

Model	Implemented	Optimized	Converges
Llama 3.0 8B	✅	✅	TODO
Llama 3.1 8B	✅	TODO	TODO
Llama 3.1 70B	TODO	TODO	TODO
Llama 3.1 405B	✅	✅	TODO
Llama 4 Scout	TODO	TODO	TODO
Llama 4 Maverick	TODO	TODO	TODO
Mixtral 8x7B	✅	TODO	TODO
Mixtral 8x22B	TODO	TODO	TODO
DeepSeek V3/R1	TODO	TODO	TODO
Stable Diffusion 2.0	TODO	TODO	TODO
Stable Diffusion 2.1	TODO	TODO	TODO

Structure

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 showcase what changes we have done to make it performant on TPU. The original version is not expected to be run.

Contributing

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.

How to run inside the docker container locally

You can also run locally without XPK with docker. When running inside the docker container, it will use the same dependencies and build process as used in the XPK approach, improving the hermeticity and reliability.

tp docker-run torchprime/torch_xla_models/train.py

This will run the torchprime docker image locally. You can also add --use-hf to run HuggingFace model locally.

tp docker-run --use-hf torchprime/hf_models/train.py

Run distributed training with local torch/torch_xla wheel

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.

License

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.