pytorch

pytorch/pytorch

Python-based GPU-accelerated deep learning tensor libraryready

PyTorch is a Python package that provides tensor computation similar to NumPy and tape-based autograd, with GPU acceleration and dynamic neural network construction as its core features. Researchers and ML engineers use `torch`, `torch.nn`, `torch.autograd`, `torch.jit`, and others to perform model training, inference, data processing, and custom extensions. Since this package's releases directly impact the core runtime, installation distribution, hardware support, and API compatibility, they are crucial for the stability and reproducibility of training and serving environments.

deep-learningpytorchpythongpumachine-learning

Choose version

2.12.13mo ago

pytorch/pytorch PyTorch 2.12.1 Release, bug fix release

Original

Key point

This is a bug fix release that resolves Triton kernel errors and memory access issues on NVIDIA B100/B200 GPU environments.

Details

Bug Fixes

NVIDIA B100/B200 (sm100) related

Fixed nondeterministic output issue in test_batch_invariance occurring when using FLASH_ATTN (resolved via Triton 3.7.1 update)
Fixed illegal memory access error occurring in the Triton convolution2d_bwd_weight kernel (resolved via Triton 3.7.1 update)

Other Fixes

Fixed incorrect fill_ method behavior for byte-dtype views with misaligned storage offset

Releng / Build

Excluded CPython 3.13t support from the binary build matrix

pytorch releases | AI Briefing