AngelSlim/Hy3-GGUF

Dedicated to building a more intuitive, comprehensive, and efficient LLMs compression toolkit.

📖 Documentation   |   🤗 Hugging Face   |   🤖 ModelScope   |   💬 WeChat

Hy3 llama.cpp Quantization

Quantize and run Hy3 (hy_v3) on llama.cpp with MTP self-speculative decoding and a thinking/tool-call parser. One script patches and builds llama.cpp; recipes and a chat template are included for quantizing your own low-bit GGUF from a calibration set.

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

---

Quick Start

Build

bash setup_hyv3_llama.sh                    # clone into ./llama.cpp-hyv3, auto CUDA/CPU
bash setup_hyv3_llama.sh /path/to/target    # choose the clone directory
CUDA=0 bash setup_hyv3_llama.sh             # force a CPU-only build

The script pins a verified llama.cpp base commit, applies patches/01 (base arch) then patches/02 (MTP + parser), and builds it. Binaries land in <target>/build/bin/.

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

GPUs	build	MTP	-c (context)	KV cache
1× H20 (96 GB)	IQ1_M	no	-c 65536	-ctk q8_0 -ctv q8_0
2× H20 (192 GB)	IQ1_M	yes	/ (default)	/ (f16)
2× H20 (192 GB)	Q4_K_M	yes	-c 65536	-ctk q8_0 -ctv q8_0 -ctkd q8_0 -ctvd q8_0
4× H20 (384 GB)	Q4_K_M	yes	/ (default)	/ (f16)

Weights: IQ1_M ~83 GiB, Q4_K_M ~166 GiB (MTP adds ~2 GiB plus a draft KV cache). 1 card fits only IQ1_M, and tightly — shrink context, quantize KV, no MTP. 2 cards run IQ1_M+MTP with room to spare (drop -c/KV flags), but Q4_K_M+MTP is tight, so keep -c 65536 and q8_0 KV (main + draft). 4 cards run Q4_K_M+MTP comfortably.

Troubleshooting

• Error: no such instruction: vdpbf16ps (in ggml-cpu's vec.cpp/sgemm.cpp): the CPU has AVX512-BF16 but the system assembler (old binutils) can't encode it. The script auto-detects this and adds -DGGML_NATIVE=OFF; force it with GGML_NATIVE=0 bash setup_hyv3_llama.sh if needed. Unrelated to the patches; CUDA inference is unaffected.
• undefined reference to SSL_get1_peer_certificate: the system OpenSSL is too old. OpenSSL is off by default here (-DLLAMA_OPENSSL=OFF); it only affects the server's HTTPS model download, not local GGUF serving. Set OPENSSL=1 if you have OpenSSL 3.0+ and need it.

---

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 and a minja-compatible chat template 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 (q8_0 / q4_K / q6_K) — cheap in size, and where low bits hurt most.
• Shared experts (ffn_*_shexp, active on every token) stay at q5_K–q6_K.
• Routed experts (ffn_*_exps) carry the aggressive low bits and dominate the file size. In the IQ1_M recipe most layers are iq1_m, with ffn_down a bit higher (iq3_xxs) and sensitive layers upgraded (iq2_xxs) layer-by-layer.

recipe	target	MTP head
recipes/hyv3_q4km_recipe.txt	~Q4_K_M mixed	no
recipes/hyv3_q4km_mtp_recipe.txt	~Q4_K_M mixed	yes
recipes/hyv3_iq1m_recipe.txt	~IQ1_M mixed (extreme)	no
recipes/hyv3_iq1m_mtp_recipe.txt	~IQ1_M mixed (extreme)	yes

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 q8_0 for Q4_K_M (embedding is cheap, no reason to crush it), drop it to q4_K 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

4. (Optional) Embed a minja-compatible chat template

Hy3's built-in chat_template uses Python str.format, which llama.cpp's minja engine can't evaluate → chat/--jinja mode crashes. hyv3_opensource_chat_template.jinja is a static version (suffix hard-coded to :opensource, .format expanded). Bake it into the gguf so --jinja works without --chat-template-file at runtime:

PYTHONPATH=./llama.cpp-hyv3/gguf-py python -m gguf.scripts.gguf_new_metadata --chat-template "$(cat hyv3_opensource_chat_template.jinja)" --force Hy3-IQ1_M-mtp.gguf Hy3-IQ1_M-mtp-tmpl.gguf