The PyTorch team released PyTorch 2.0, introducing torch.compile (TorchDynamo + TorchInductor + Triton codegen) to capture Python execution graphs and emit fused GPU kernels automatically while preserving eager-mode ergonomics.
PyTorch
Also known as: torch · torch.compile · FSDP
Plain English
The dominant open-source Python framework for building, training, and running neural networks, prized for feeling like ordinary Python code while still executing at full GPU speed.
Technical Overview
Originated at Meta AI and now governed by the PyTorch Foundation, PyTorch pairs eager-mode automatic differentiation with a C++ / CUDA tensor runtime, torch.compile (TorchDynamo + TorchInductor) graph compilation, and native distributed training primitives (DDP, FSDP2, Tensor Parallelism). It serves as the primary training and model-authoring substrate beneath Hugging Face, vLLM, SGLang, and TensorRT-LLM.
Chronological Entries (2)
Paszke et al. published the landmark PyTorch paper (An Imperative Style, High-Performance Deep Learning Library), documenting how dynamic define-by-run autograd and a C++ tensor engine let researchers iterate in native Python without sacrificing GPU throughput.
Direct Relations (4)
- Created byMeta AI
- UsesCUDA
- EnablesDistributed Training
- EnablesHugging Face