Module cuda_backend 

Real CUDA GPU backend using AtlasRegistry.

SBIO IORouter: all CUDA operations flow through GpuBackend. Uses AtlasRegistry for kernel loading/launching and raw CUDA driver API for memory management.

Modules

tensormap

TMA descriptors (CUtensorMap) for cp.async.bulk.tensor loads.

Structs

AtlasCudaBackend

Production GPU backend wrapping AtlasRegistry + raw CUDA driver API.

Functions

cuda_free_memory_bytes

Query GPU free memory without requiring a GpuBackend reference. Safe to call from any thread that shares the CUDA context.

spawn_oom_watchdog

Start the OOM watchdog, or return None if one is already running.