imp

From-scratch C++/CUDA inference engine for the NVIDIA RTX 5090 (sm_120a) — the best single-GPU backend for agentic AI: tool calling, long-context loops, reasoning and concurrent sub-agents on top of the fastest single-stream decode on the 5090 (beats llama.cpp, at-or-ahead of vLLM on NVFP4). 100% written by Claude Code. Maintained by kekzl on GitHub, where it has earned 30 stars from the community.

It's actively developed around blackwell, cpp, cuda, and is a solid reference for anyone building with these tools.