akhaliq/llama-cpp-ricursive-cuda120

llama.cpp + ricursive (Ripple-65B) — prebuilt CUDA binaries

Prebuilt binaries of llama.cpp with support for the ricursive architecture (ricursiveAI/Ripple-65B-Preview), which is not in upstream llama.cpp yet (PR #31692).

Built from llama.cpp @ 9e2a04a

• AtomicChat's ricursive-port.patch, with CUDA 12.8 for sm_120 (RTX PRO 6000 Blackwell) only, CPU backend with runtime-dispatched variants. The exact build script is build.sh in this repo (reproducible with bash build.sh in an nvidia/cuda:12.8.1-devel-ubuntu22.04 container).

Files

File	What	Use
llama-ricursive-cuda128-sm120.tar.gz	llama-server + libggml* shared libs (bin/)	Untar, run bin/llama-server -m <gguf> --jinja -c 65536 -ngl all. This is what akhaliq/ripple-65b-zerogpu downloads at startup
llama_cpp_python-0.3.36+cu128.ricursive-py3-none-linux_x86_64.whl	llama-cpp-python 0.3.36 with the same patched llama.cpp	pip install https://huggingface.co/akhaliq/llama-cpp-ricursive-cuda120/resolve/main/llama_cpp_python-0.3.36%2Bcu128.ricursive-py3-none-linux_x86_64.whl

Notes

• Linux x86_64, glibc (built on Ubuntu 22.04). GPU: Blackwell sm_120 only — other cards need a rebuild with different CMAKE_CUDA_ARCHITECTURES.
• GGUFs that work with this: AtomicChat/Ripple-65B-Preview-GGUF (BF16 / Q8_0 / Q4_K_M) and prometheusAIR/Ripple-65B-Preview-GGUF (Q8_0 / Q6_K_XL / Q6_K) — note prometheusAIR ship their own patch; binaries here follow AtomicChat's port, both load the same files per AtomicChat's card.
• Once PR #31692 merges upstream, prefer official llama.cpp releases; this repo is a stopgap.

MIT (llama.cpp, llama-cpp-python) · GGUF weights are Apache-2.0 via Ricursive.