PyTorch
Open-source deep learning framework with dynamic graphs, torch.compile and torch.distributed, governed by the PyTorch Foundation
Overview
PyTorch describes itself as "Tensors and Dynamic neural networks in Python with strong GPU acceleration." Work on it began at Facebook (now Meta) in 2016 and PyTorch 1.0 shipped in early 2018; today it lives at the PyTorch Foundation, which is hosted by the Linux Foundation and also supports vLLM, DeepSpeed, Ray, Helion and Safetensors. PyTorch 2.15 (October 28, 2026) made a rewritten NCCL backend the default (with a flight recorder for hangs), added NIXL one-sided transports to torch.distributed, gave FSDP2 finer precision and memory control, added NVFP4 decode on Blackwell and FP8 plus faster MPS decode on Apple Silicon, and ships ROCm 10.0 wheels. The release took 3,117 commits from 583 contributors. Around the core sit ExecuTorch for on-device inference, TorchTPU (a native TPU backend built with Google) and TorchNeuron for AWS Trainium. Related on this site: the PyTorch Foundation also hosts Ray and vLLM, and Meta's torchtune (no longer maintained) was a PyTorch-native fine-tuning library.
The Verdict
Who Should Use PyTorch?
Best For
- Researchers and engineers training or fine-tuning neural networks, including LLMs
- Teams that need multi-GPU and multi-node training (FSDP2, torch.distributed)
- Projects that depend on the wider ecosystem (Transformers, vLLM, Ray, ExecuTorch)
Not Ideal For
- Classic tabular ML where scikit-learn or gradient boosting is enough
- Teams that want a TPU-first, purely functional stack (see JAX)
- Users on Python 3.10 or macOS 14 who need MPS (both dropped in 2.15)
What's Great
- Eager, Pythonic execution with torch.compile for speed when you need it
- Broad hardware support: CUDA, ROCm, Apple MPS, plus TPU and Trainium backends
- Vendor-neutral governance under the Linux Foundation
- Fast release cadence: six minor releases between October 2025 and October 2026
Watch Out For
- Complexity and learning curve come up in G2 reviews (small sample)
- Frequent releases drop old Python, CUDA and macOS versions, so plan upgrades
- The pytorch.org install selector showed an old stable version (2.7.0) when we checked; trust the GitHub releases page
Pricing
View all features & details
Key Features (2.15)
- torch.compile, with Dynamo support for Python 3.15
- Rewritten NCCL backend as default, with a flight recorder for hangs
- NIXL one-sided transports in torch.distributed
- FSDP2 precision and memory controls
- NVFP4 decode on Blackwell; FP8 and faster MPS decode on Apple Silicon
- CUDA 13.2 default wheel, ROCm 10.0 wheels, Python 3.11 to 3.15
Ecosystem
- ExecuTorch on-device runtime, with an Arm Cortex-M backend (2026)
- TorchTPU native TPU backend (Google and Meta)
- TorchNeuron for AWS Trainium (public beta, October 2026)
- Foundation projects: vLLM, DeepSpeed, Ray, Helion, Safetensors
Real-World Usage
Community
- PyTorch Conference North America 2026 (San Jose, October 20 to 21): 3,500+ attendees, 100+ sessions
- New Foundation members in 2026 include Nebius, Nscale and Shopify (Platinum)
Named Users
- Shopify: continual-learning loop on PyTorch and vLLM, serving costs cut 96% (company claim)
- Amazon Advertising: inference costs cut 71% (case study on pytorch.org)
- Meta, where the framework started
How It Compares
| Feature | PyTorch | TensorFlow | JAX | MLX |
|---|---|---|---|---|
| Style | Eager, object-oriented, optional compile | Graphs plus Keras | Functional transforms (jit, grad, vmap) | NumPy-like, lazy |
| Hardware | NVIDIA, AMD, Apple MPS, TPU, Trainium | CPU, GPU, TPU, on-device via LiteRT | CPU, GPU, TPU (macOS CPU only) | Apple silicon first |
| 2026 status | Active, releases every few months | Maintenance focus since 2.21 | Active | Active, pre-1.0 |
| License | BSD-style | Apache-2.0 | Apache-2.0 | MIT |
Sources: GitHub · 2.15 release blog · TensorFlow 2.21 post · JAX install docs
Is this your tool? Add a badge