DDO VMM allocator

Native CUDA Driver VMM and Python adapters consumed by DDO. This repository owns native arenas, source/cache transport, event lifetimes, pressure/admission, the PyTorch pluggable allocator and accounting. It contains no model-specific runners or weights.

Documentation

• User installation
• Developer architecture and contracts
• Maintainer builds
• Tests and validation
• Experimental package delivery
• Agent instructions

Current scope

Selected Windows, native Linux and WSL workloads passed bounded contracts. Native VMM remains experimental; arbitrary models/drivers, macOS and driver-loss recovery are not promised. Stable DDO linear/NVMe paths must remain untouched. Optional Windows/Linux wheels can embed hash-selected DLL/ELF artifacts; Linux host-wheel assembly passed CPU validation, not broad ABI/CUDA delivery acceptance. Editable Python installation does not compile native code. No binary is built/downloaded during import or inference. Historical probe scaffolding and investigations are preserved locally in ignored .internal/documentation/; current contracts do not depend on those files. Model benchmarks belong to DDO examples.

License

Apache-2.0. Preserve existing copyright/license/attribution notices in source and distribution. Driver/toolkit components retain their own applicable terms.