HF RL Explorer

CUDA backend allocator: Safe stream-aware caching and reuse of heap pages

CUDA backend allocator: Safe stream-aware caching and reuse of heap pages: a task in LegoFlow-SWE (Harbor dataset). The CUDA backend heap allocator does not correctly distinguish same‑stream reuse from cross‑stream reuse. This causes two classes of misbehavior:

The task

The CUDA backend heap allocator does not correctly distinguish same‑stream reuse from cross‑stream reuse. This causes two classes of misbehavior:

Part of Lego-X/LegoFlow-SWE.