PyTorch bounds bug, SYCL graphs, WeatherNext model

A reported CPU memory safety issue in PyTorch triangle operators led the section, while pull requests advanced SYCL graph support in llama.cpp and a weather model in Transformers. Activity stayed technical and centered on correctness, backend performance, and library coverage.

Out-of-bounds write in PyTorch triu and tril

A bug report filed against pytorch/pytorch describes a CPU out-of-bounds write in torch.triu and torch.tril when the out argument has non-contiguous batch dimensions. The report is classified as a security-relevant memory safety defect in core tensor operations. Users who supply custom non-contiguous output tensors to these functions have a concrete condition to watch.

SYCL graph record and replay in llama.cpp

A pull request in ggml-org/llama.cpp adds SYCL graph record and replay support. Contributors shared early Intel GPU benchmarks and discussed tuning questions around the new path. The change matters for anyone targeting Intel GPUs for large-language-model inference with llama.cpp.

WeatherNext 2 forecasting model in Transformers

A pull request in huggingface/transformers introduces the WeatherNext 2 weather forecasting model. Review discussion stayed limited to feature extractors, casts, and repeated continuous-integration runs. The addition extends the library's set of specialized forecasting architectures available to practitioners.