Hy MT2 30B A3B GGUF

--- license: other base_model: tencent/Hy-MT2-30B-A3B language: - zh - en - multilingual tags: - gguf - llama.cpp - hy-mt2 - hy-v3 - translation - moe - quantized pipeline_tag: text-generation library_name: llama.cpp --- # Hy-MT2-30B-A3B GGUF Models Chinese documentation: [README.zh-CN.md](README.zh-CN.md) This repository contains GGUF conversions and quantizations of Tencent's `tencent/Hy-MT2-30B-A3B` translation model for llama.cpp-compatible inference. ## Files | File | Type / Quantization | Size | Notes | | --- | --- | ---: | --- | | `Hy-MT2-30B-A3B-BF16.gguf` | BF16 / FP16-class source GGUF | ~56 GB | Highest fidelity source GGUF; useful for re-quantization or maximum-quality inference | | `Hy-MT2-30B-A3B-Q2_K.gguf` | Q2_K | ~11 GB | Smallest/lowest-memory option; lowest quality | | `Hy-MT2-30B-A3B-Q3_K_M.gguf` | Q3_K_M | ~14 GB | Very low-memory option, better than Q2_K | | `Hy-MT2-30B-A3B-Q4_K_M.gguf` | Q4_K_M | ~17 GB | Recommended balanced/low-memory option | | `Hy-MT2-30B-A3B-Q5_K_M.gguf` | Q5_K_M | ~20 GB | Better quality than Q4_K_M with moderate extra memory | | `Hy-MT2-30B-A3B-Q6_K.gguf` | Q6_K | ~24 GB | Better quality, higher VRAM/RAM usage | | `Hy-MT2-30B-A3B-Q8_0.gguf` | Q8_0 | ~30 GB | Highest fidelity among the quantized files | SHA256 sidecar files are provided for the GGUF files when available. ## Quantization notes The Q2_K, Q3_K_M, and Q5_K_M files were quantized directly from the BF16 GGUF source, not requantized from another low-bit GGUF. Because Hy-MT2-30B-A3B uses the `hy_v3` architecture, conversion and quantization require llama.cpp tooling that supports `hy_v3`. A generic llama.cpp quantizer may fail with `unknown model architecture: 'hy_v3'`. ## Important compatibility note Hy-MT2-30B-A3B uses the `hy_v3` architecture. It requires a llama.cpp build that supports this architecture. If your llama.cpp build does not support it, loading the GGUF may fail with: ```text unknown model architecture: 'hy_v3' ``` Use a compatible llama.cpp build/branch for Hy-V3/Hy-MT2 models. ## Recommended llama.cpp server usage Example: ```bash ./llama-server \ -m Hy-MT2-30B-A3B-Q4_K_M.gguf \ --alias tencent/Hy-MT2-30B-A3B-GGUF:Q4_K_M \ --host 0.0.0.0 \ --port 18080 \ -c 131072 \ --n-gpu-layers 60 \ --jinja \ -r ' ' ``` Notes: - `--jinja` is recommended so llama.cpp uses the chat template correctly. - `-r ' '` is recommended as a reverse prompt / stop marker because this model may emit the textual EOS marker if the runtime does not treat it as a native EOS token. - Adjust `--n-gpu-layers` according to your GPU memory. - For long context, make sure you have enough VRAM/RAM. KV-cache quantization may be useful on smaller GPUs if supported by your runtime. - BF16/FP16-class GGUF requires much more RAM/VRAM than the quantized files. ## Example OpenAI-compatible API request ```bash curl http://127.0.0.1:18080/v1/chat/completions \ -H 'Content-Type: application/json' \ -d '{ "messages": [ {"role": "user", "content": "Translate to English: 今天天气很好。"} ], "temperature": 0, "max_tokens": 128 }' ``` Expected output is a direct translation such as: ```text The weather is very nice today. ``` ## Source model and license - Upstream model: https://huggingface.co/tencent/Hy-MT2-30B-A3B - These GGUF files are converted and quantized from the upstream model. - Please follow the upstream model license and usage terms. A copy of the upstream `LICENSE.txt` is included when available.