AngelSlim/Hy3-GGUF

<p align="center"> <picture> <source media="(prefers-color-scheme: dark)" srcset="https://github.com/Tencent/AngelSlim/blob/main/docs/source/assets/logos/angelslimlogolight.png?raw=true"> <img alt="AngelSlim" src="https://github.com/Tencent/AngelSlim/blob/main/docs/source/assets/logos/angelslim_logo.png?raw=true" width=55%> </picture> </p>

<h3 align="center"> Dedicated to building a more intuitive, comprehensive, and efficient LLMs compression toolkit. </h3>

<p align="center"> 📖 <a href="https://angelslim.readthedocs.io/">Documentation</a>&nbsp&nbsp | &nbsp&nbsp🤗 <a href="https://huggingface.co/AngelSlim">Hugging Face</a>&nbsp&nbsp | &nbsp&nbsp🤖 <a href="https://modelscope.cn/models/AngelSlim/Hy3-GGUF">ModelScope</a>&nbsp&nbsp | &nbsp&nbsp💬 <a href="./docs/source/assets/angelslimwechat.png">WeChat</a> <br> </p>

<br>

Hy3 llama.cpp Quantization

Low-bit quantized Hy3 (hy_v3) GGUFs for llama.cpp — ready to run out of the box, with MTP self-speculative decoding. This repo ships the quantized models, plus the mixed-precision recipes to build your own from a calibration set.

<p align="center"> <img src="assets/benchmark.png" width="95%"/> </p>

Two parts below: Deploy (build & run) and Quantization.

Quick Start

Build

Clone the latest llama.cpp (must include the hy_v3 support merged in PR #25395 — i.e. any commit after 505b1ed; the current master is fine) and build it following the official build guide:

git clone https://github.com/ggml-org/llama.cpp llama.cpp-hyv3
# build per docs/build.md (CUDA / Metal / CPU as appropriate for your machine)

Binaries land in llama.cpp-hyv3/build/bin/.

Already downloaded an older GGUF? Earlier GGUFs here were built against a patched llama.cpp and won't load correctly on today's upstream. To build and run those, follow the previous README (the setup_hyv3_llama.sh + patches flow) instead.

Run

# plain serve (no speculative decoding)
./llama.cpp-hyv3/build/bin/llama-server -m /path/to/Hy3.gguf -ctk q8_0 -ctv q8_0 -fa on -c 65536

# serve + MTP self-speculative decoding (needs an MTP gguf, i.e. converted WITHOUT --no-mtp)
./llama.cpp-hyv3/build/bin/llama-server -m /path/to/Hy3-mtp.gguf --spec-type draft-mtp --spec-draft-n-max 3 --spec-draft-n-min 1 -ctk q8_0 -ctv q8_0 -ctkd q8_0 -ctvd q8_0 -fa on -c 65536

Recommended setups

Weights: IQ1M ~83 GiB, Q4KM ~166 GiB (MTP adds ~2 GiB plus a draft KV cache). 1 card fits only IQ1M, and tightly — shrink context, quantize KV, no MTP. 2 cards run IQ1M+MTP with room to spare (drop `-c`/KV flags), but Q4KM+MTP is tight, so keep `-c 65536` and q80 KV (main + draft). 4 cards run Q4KM+MTP comfortably.

Quantize with Your Own Data

If you have a calibration set, you can compute your own importance matrix and quantize the model yourself. All steps use stock llama.cpp tools; the mixed-precision recipes are shipped in this folder.

The mixed-precision recipes

The recipes in recipes/ spend bits where they matter:

• —Attention (attn_q/k/v) and the token embedding / output head stay high (q80 / q4K / q6_K) — cheap in size, and where low bits hurt most.
• —Shared experts (ffn_*_shexp, active on every token) stay at q5K–q6K.
• —Routed experts (ffn_*_exps) carry the aggressive low bits and dominate the file size. In the IQ1M recipe most layers are iq1m, with ffn_down a bit higher (iq3xxs) and sensitive layers upgraded (iq2xxs) layer-by-layer.

The *_mtp variants add the MTP block (blk.<n>.nextn.* and layer-<n> experts). Use them only for a gguf that still has the MTP head; use the plain ones for a gguf converted with --no-mtp. (The MTP block's experts stay at K-quant, not IQ*, because very-low-bit IQ types need an imatrix and the imatrix only covers the trunk layers.)

1. Convert HF → BF16 GGUF

PYTHONPATH=./llama.cpp-hyv3/gguf-py python ./llama.cpp-hyv3/convert_hf_to_gguf.py /path/to/HYV3-hf --outfile Hy3-BF16.gguf --outtype bf16
#   add --no-mtp to drop the MTP head (then use a non-mtp recipe below)

2. Compute the importance matrix (imatrix)

The calibration file is plain text containing samples with the model's special tokens already applied (one file, fed via -f). The imatrix is independent of the target quant type — compute it once on BF16 and reuse it for any recipe.

./llama.cpp-hyv3/build/bin/llama-imatrix -m Hy3-BF16.gguf -f calib.txt -o imatrix.gguf --output-format gguf --parse-special
#   single GPU can't hold BF16 -> add --cpu-moe (keep experts on CPU) or -ngl 0

3. Quantize with a recipe

Pick the recipe matching your target and whether the gguf has an MTP head (table above). --token-embedding-type differs: keep it at q80 for Q4KM (embedding is cheap, no reason to crush it), drop it to q4K for the extreme IQ1_M build. The examples below are for MTP ggufs — use the non-mtp recipe for a --no-mtp gguf.

# ~Q4_K_M mixed
./llama.cpp-hyv3/build/bin/llama-quantize --imatrix imatrix.gguf --tensor-type-file recipes/hyv3_q4km_mtp_recipe.txt --token-embedding-type q8_0 --output-tensor-type q6_K Hy3-BF16-mtp.gguf Hy3-Q4_K_M-mtp.gguf Q4_K_M

# ~IQ1_M mixed (extreme)
./llama.cpp-hyv3/build/bin/llama-quantize --imatrix imatrix.gguf --tensor-type-file recipes/hyv3_iq1m_mtp_recipe.txt --token-embedding-type q4_K --output-tensor-type q6_K Hy3-BF16-mtp.gguf Hy3-IQ1_M-mtp.gguf IQ1_M