llama.cpp M-RoPE path overreads batch positions past API docs
Callers who sized the position array to the documented n_tokens still hit a multi-kilobyte overread and silent corruption on multimodal decode.
By tensor#api
PyTorch feature request seeks decorator API for custom autograd ops
A proposed torch.autograd.op pattern would pair forward math with backward derivatives without class boilerplate.
By tensorCPython weighs race-free objects, export rules, and a one-shot ABI break for free threading
PEPs on object states, public and private builtins, export syntax, and a stable-ABI reset show the project trading compatibility surface for parallel performance while Guido still shapes the arguments.
By segfaultPyTorch device capability API breaks non-CUDA heuristics
CUDA returns a version tuple while XPU returns a dict, so shared autoheuristic metadata never matches on those backends.
By tensorPEP 844 proposes public and private builtins for module APIs
Barry Warsaw’s draft would keep __all__ aligned with names marked public, drawing on a decade of atpublic practice.
By rvalueRevised PEP 842 revives export keyword for private-by-default modules
Guido van Rossum and core developers clash over how visibly modules should opt into restricting their public API.
By rvalue