May 02, 2024
Announcing PyTorch Docathon June, 2024
We are thrilled to announce the upcoming PyTorch Docathon in June! The Docathon, akin to a hackathon, is an event dedicated to enhancing the quality of the PyTorch documentation with the invaluable assistance of our community. Documentation is a vital component of any technology. By refining it, we can simplify the process for new users to get started with PyTorch, guide them in effectively utilizing its features, and ultimately expedite the transition from research to production in machine l...
April 30, 2024
ExecuTorch Alpha: Taking LLMs and AI to the Edge with Our Community and Partners
We are excited to announce the release of ExecuTorch alpha, focused on deploying large language models (LLMs) and large ML models to the edge, stabilizing the API surface, and improving our installation processes. It has been an exciting few months from our 0.1 (preview) release in collaboration with our partners at Arm, Apple, and Qualcomm Technologies, Inc.
April 24, 2024
PyTorch 2.3 Release Blog
We are excited to announce the release of PyTorch® 2.3 (release note)! PyTorch 2.3 offers support for user-defined Triton kernels in torch.compile, allowing for users to migrate their own Triton kernels from eager without experiencing performance regressions or graph breaks. Tensor Parallelism improves the experience for training Large Language Models using native PyTorch functions, which has been validated on training runs for 100B parameter models. As well, semi-structured sparsity implemen...
April 16, 2024
torchtune: Easily fine-tune LLMs using PyTorch
We’re pleased to announce the alpha release of torchtune, a PyTorch-native library for easily fine-tuning large language models.
March 13, 2024
Maximizing training throughput using PyTorch FSDP
In this blog, we demonstrate the scalability of FSDP with a pre-training exemplar, a 7B model trained for 2T tokens, and share various techniques we used to achieve a rapid training speed of 3,700 tokens/sec/GPU, or 40B tokens/day on 128 A100 GPUs. This translates to a model FLOPS utilization (MFU) and hardware FLOPS utilization (HFU) of 57%. Additionally, we have observed near linear scaling of FSDP to 512 GPUs, implying that training a 7B model on 512 GPUs to 2T tokens using this method wou...