PyTorch split rewrite breaks vLLM QK-norm RoPE fusion
A functionalization change that turns multi-output splits into slices leaves vLLM’s pattern matcher with nothing to fuse under Inductor graph partition.
By tensor#compiler
PyTorch adds experimental torch.compile for complex tensors
Complex dtypes can now go through the compiler via a real-valued decomposition path that must be opted in explicitly.
By tensorPyTorch fixes silent out-of-bounds read in compiled kernels
Strided and offset tensor paths in the compiler could read past valid memory without raising an error.
By tensorSwift pitch proposes @called(once) for one-shot function values
The attribute would let the compiler enforce at-most-once calls and reason about closure captures alongside non-copyable types.
By segfault