Hy-MT2-30B-A3B-4bit-GGUF

Hy-MT2-30B-A3B 4bit GGUF

GGUF quantizations of tencent/Hy-MT2-30B-A3B, with Q4_K_M as the recommended 4bit deployment file.

Files

Important

Current upstream llama.cpp may not support HYV3ForCausalLM / model_type=hy_v3 directly. This repository includes the required patch at:

patches/llama-cpp-hyv3.patch

Clone this model repository first, then prepare a patched llama.cpp checkout:

git lfs install
git clone https://huggingface.co/GrahLnn/Hy-MT2-30B-A3B-4bit-GGUF
cd Hy-MT2-30B-A3B-4bit-GGUF
bash scripts/prepare_llama_cpp.sh ./llama.cpp

For CUDA 13.1 builds:

CUDA_HOME=/usr/local/cuda-13.1 bash scripts/prepare_llama_cpp.sh ./llama.cpp

Run Q4 Server

cd Hy-MT2-30B-A3B-4bit-GGUF
CUDA_VISIBLE_DEVICES=0 bash scripts/run_q4_server.sh

The script runs:

./llama.cpp/build-hyv3-cuda/bin/llama-server \
  -m Hy-MT2-30B-A3B-Q4_K_M.gguf \
  -ngl all \
  --host 127.0.0.1 \
  --port 8080

Test Translation

bash scripts/test_translation.sh

Expected style:

The weather is really nice today.

Use stop strings when serving through the OpenAI-compatible endpoint:

["<eos:6124c78e>", "<｜hy_User｜>", "<｜hy_Assistant｜>"]

Measured Local Runtime

Test machine: RTX 4090, WSL CUDA, patched llama.cpp.

For Q4 server translation tests after warmup, generation was typically around 195-206 tok/s on the local 4090 setup. Actual speed depends on prompt length, context size, batch settings, and build options.

Tested Cases

The Q4 server path was tested with:

• normal Chinese to English translation
• mixed Chinese/English/Japanese text
• code and function names
• Markdown lists and tables
• XML tags
• JSON-like text
• logs, shell commands, and Windows/WSL paths

Structure-heavy prompts should explicitly ask the model to preserve structure, keys, tags, and newlines.

License

This repository contains quantized derivatives of tencent/Hy-MT2-30B-A3B. Follow the upstream model license and usage terms.