Đang tải bảng tin…
Hacker News0 Hot0 bình luận5 phút đọc
PyTorch: A Reference Language
PyTorch is increasingly playing a dual role: as both a reference language for deep learning and a production implementation language. As kernel DSLs have made it easier to write optimized operators, LLMs are now enabling a similar pattern for training steps — generating explicit forward-backward code that can be optimized independently from the PyTorch reference implementation. The proposed workflow keeps traditional autograd-based PyTorch code as a reference spec, uses LLMs to generate optimized explicit backward passes, and employs a verifier (bitwise equivalence or translation validation) to ensure correctness. This approach sidesteps the limitations of implicit autograd graphs at scale while keeping PyTorch central to frontier training research.
Nguồn: https://docs.pytorch.org/devlogs/compiler/2026-07-25-pytorch-a-reference-language. 8 Sync News chỉ tóm tắt và dẫn link; bản quyền nội dung thuộc tác giả và nguồn gốc.
