Infatoshi/MegaQwen

Hyper optimised Qwen3.8-27B inference on one RTX 3090: ReplaySSM, MTP3, reasoning effort, C1-C8 batching, and native Windows and Linux builds.

From-scratch C++/CUDA inference engine for Qwen3-8B

An agent harness that compiles a model into one provably-correct, self-retargeting CUDA megakernel and self-tunes it past cuBLAS at batch-1 LLM decode, paper: https://arxiv.org/abs/2606.09682

Testbed for LLM inference with cutile-rs.

NInfer port for Qwen 3.8 27B on RTX 4090

Qwen3.8-27B on one RTX 4090: full native 262K context via E8 4-bit KV, up to 149 tok/s code decode with MTP3, sm_89-retuned attention prefill, vision, llama.cpp-compatible /metrics + /slots

One line of NInfer for the RTX 3090, RTX 4090 and RTX 5090, consolidated from the forks that carry it: GGUF block formats, ternary Bonsai 2, device route profiles, MTP and DFlash2.

Kernels, of the mega variety :)

Hand-written NVFP4 W4A16 CUDA kernels for Volta