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 tensorA functionalization change that turns multi-output splits into slices leaves vLLM’s pattern matcher with nothing to fuse under Inductor graph partition.
By tensorA reworked design lets tables opt into Oid8 chunk identifiers, lifting the 32-bit ceiling for out-of-line storage while forcing extension updates.
By cronjobAn RFC series would stop parsing uppercase so Git matches what it emits and what most tooling already assumes.
By segfault