Notifications

No recent security alerts.

PyTorch

OSS

PyTorch 2.10.0 enhances compiler support, adds Intel GPU FP8/complex ops, introduces varlen attention, and improves distributed debugging with LocalTensor.

Repo ZH
Stars
0
Contributors
0
Health Score
0
Last Release
N/A
Overview

PyTorch 2.10.0 adds experimental Python 3.14 and freethreaded build support for torch.compile(), introduces combo-kernels to reduce kernel launch overhead, and adds varlen_attn() for ragged sequences. It expands Intel GPU support (Panther Lake) with FP8, complex MatMul, and SYCL extensions, enables deterministic mode in torch.compile(), and introduces DebugMode for numerical divergence tracking. New features include LocalTensor for single-process distributed simulation, eigenvalue decompositions via DnXgeev, and numerous Inductor, Dynamo, quantization, and ROCm/MPS/XPU backend improvements. Ba