merge-mtp-support-b9109-93ecf68

mtp-pflash-turboquant — Windows CUDA Build

Built from branch merge/mtp-support with CUDA 13.1, targeting sm_89 (RTX 40xx) and sm_120 (RTX 50xx / Blackwell).

Install

1. Extract llama-merge-mtp-support-b9109-93ecf68-bin-win-cuda-13.1-x64.zip — all EXEs, DLLs, and ggml-cuda.dll
2. Extract cudart-llama-merge-mtp-support-b9109-93ecf68-bin-win-cuda-13.1-x64.zip into the same folder — CUDA runtime DLLs

Then run llama-server.exe or llama-cli.exe directly — no build required.

Features

• MTP (Multi-Token Prediction) speculative decoding
• PFlash speculative prefill
• TurboQuant KV cache compression