Instructions to use HelloSun/lingtiny30vq3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use HelloSun/lingtiny30vq3 with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf HelloSun/lingtiny30vq3 # Run inference directly in the terminal: llama cli -hf HelloSun/lingtiny30vq3
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf HelloSun/lingtiny30vq3 # Run inference directly in the terminal: llama cli -hf HelloSun/lingtiny30vq3
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-app # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama serve -hf HelloSun/lingtiny30vq3 # Run inference directly in the terminal: ./build/bin/llama cli -hf HelloSun/lingtiny30vq3
Use Docker
docker model run hf.co/HelloSun/lingtiny30vq3
- Llama new
- LM Studio
- Jan
- Ollama
How to use HelloSun/lingtiny30vq3 with Ollama:
ollama run hf.co/HelloSun/lingtiny30vq3
- Unsloth Desktop
- Pi
How to use HelloSun/lingtiny30vq3 with Pi:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf HelloSun/lingtiny30vq3
Configure the model in Pi
# Install Pi: npm install -g @earendil-works/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "llama-cpp": { "baseUrl": "http://localhost:9931/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "HelloSun/lingtiny30vq3" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Docker Model Runner
How to use HelloSun/lingtiny30vq3 with Docker Model Runner:
docker model run hf.co/HelloSun/lingtiny30vq3
- Lemonade
How to use HelloSun/lingtiny30vq3 with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull HelloSun/lingtiny30vq3
Run and chat with the model
lemonade run user.lingtiny30vq3-{{QUANT_TAG}}List all available models
lemonade list
- Hermes Agent
How to use HelloSun/lingtiny30vq3 with Hermes Agent:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf HelloSun/lingtiny30vq3
Configure Hermes
# Install Hermes: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash hermes setup # Point Hermes at the local server: hermes config set model.provider custom hermes config set model.base_url http://127.0.0.1:9931/v1 hermes config set model.default HelloSun/lingtiny30vq3
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use HelloSun/lingtiny30vq3 with OpenClaw:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf HelloSun/lingtiny30vq3
Configure OpenClaw
# Install OpenClaw: npm install -g openclaw@latest # Register the local server and set it as the default model: openclaw onboard --non-interactive --mode local \ --auth-choice custom-api-key \ --custom-base-url http://127.0.0.1:9931/v1 \ --custom-model-id "HelloSun/lingtiny30vq3" \ --custom-provider-id llama-cpp \ --custom-compatibility openai \ --custom-text-input \ --accept-risk \ --skip-health
Run OpenClaw
openclaw agent --local --agent main --message "Hello from Hugging Face"
File size: 10,538 Bytes
511ba50 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 | #!/usr/bin/env python3
"""
build_ling_vq3.py — Build the inclusionAI Ling-3.0-tiny LUT-VQ3-64 hybrid GGUF.
Based on llama-modchip's scripts/build_nemotron_vq3.py, adapted for the
BailingMoeV3 (bailing_hybrid) architecture and for the VQ3 port to upstream
llama.cpp @ 61dbf492 (GGML_TYPE_VQ3_0 = 44 in the port, 24 B / 64 W, 3.00 bpw).
Recipes:
hybrid (default, domain-aware — mirrors the modchip VQ3-64 layout):
blk.*.ffn_gate_exps.weight + blk.*.ffn_up_exps.weight -> LUT-VQ3-64 (3.00 bpw)
blk.*.ffn_down_exps.weight -> Q6_K (6.56 bpw)
attention / shared experts / dense ffn / embeddings -> Q8_0 (8.50 bpw)
max (--down-vq3):
additionally blk.*.ffn_down_exps.weight -> LUT-VQ3-64 (3.00 bpw)
The routed-expert tensors are encoded directly from the uncompressed BF16 master
GGUF (inclusionAI/Ling-3.0-tiny-bf16.gguf) with the modchip's OpenMP AVX-512
encoder (vq3/vq3_enc_lib.c, 128-entry 3D Lloyd-Max codebook).
Usage:
python3 build_ling_vq3.py --base-gguf models/Ling-3.0-tiny-base.gguf \
--bf16-gguf models/Ling-3.0-tiny-bf16.gguf \
--output models/Ling-3.0-tiny-VQ3.gguf
"""
import argparse
import ctypes
import mmap
import os
import struct
import subprocess
import sys
import time
from pathlib import Path
import numpy as np
GGML_TYPE_BF16 = 30
GGML_TYPE_VQ3_0 = 44 # llama-modchip port: appended after upstream GGML_TYPE_F8_E4M3 = 43
def load_vq3_encoder(lib_path: str = "") -> ctypes.CDLL:
here = Path(__file__).resolve().parent
# encoder source lives next to this script (HF repo layout) or in a llama-modchip checkout
so_candidates = [
Path(lib_path) if lib_path else None,
here.parent / "vq3" / "libvq3_enc.so",
here.parent.parent / "llama-modchip" / "vq3" / "libvq3_enc.so",
]
so_path = next((c for c in so_candidates if c and c.is_file()), None)
src_candidates = [
here.parent / "vq3" / "vq3_enc_lib.c",
here.parent.parent / "llama-modchip" / "vq3" / "vq3_enc_lib.c",
]
c_src = next((c for c in src_candidates if c.is_file()), None)
if so_path is None:
if not c_src:
raise FileNotFoundError(f"Missing encoder library and source {src_candidates}")
so_path = c_src.with_suffix(".so").parent / "libvq3_enc.so"
print(f"[*] Compiling {c_src} -> {so_path} ...", flush=True)
subprocess.check_call([
"gcc", "-O3", "-march=native", "-mavx512f", "-mf16c",
"-fopenmp", "-shared", "-fPIC", str(c_src), "-lm", "-o", str(so_path),
])
lib = ctypes.CDLL(str(so_path))
lib.quantize_bf16_to_vq3_0_omp.argtypes = [
ctypes.c_void_p, ctypes.c_void_p, ctypes.c_int64, ctypes.c_int64,
]
lib.quantize_bf16_to_vq3_0_omp.restype = None
return lib
def parse_gguf(f):
"""Minimal GGUF v2/v3 reader: returns metadata KV blob + tensor directory."""
f.seek(0)
magic, version, n_tensors, n_kv = struct.unpack("<4sIQQ", f.read(24))
if magic != b"GGUF":
raise ValueError(f"Not a GGUF file (magic={magic!r})")
def read_str():
length = struct.unpack("<Q", f.read(8))[0]
return f.read(length).decode("utf-8", errors="replace")
def skip_val(vtype):
if vtype in (0, 1, 7):
f.seek(1, 1)
elif vtype in (2, 3):
f.seek(2, 1)
elif vtype in (4, 5, 6):
f.seek(4, 1)
elif vtype in (10, 11, 12):
f.seek(8, 1)
elif vtype == 8:
length = struct.unpack("<Q", f.read(8))[0]
f.seek(length, 1)
elif vtype == 9:
atype, alen = struct.unpack("<IQ", f.read(12))
for _ in range(alen):
skip_val(atype)
else:
raise ValueError(f"Unknown GGUF metadata type {vtype}")
for _ in range(n_kv):
_ = read_str()
vtype = struct.unpack("<I", f.read(4))[0]
skip_val(vtype)
kv_end = f.tell()
tensors = []
for _ in range(n_tensors):
name = read_str()
n_dims = struct.unpack("<I", f.read(4))[0]
dims = [struct.unpack("<Q", f.read(8))[0] for _ in range(n_dims)]
ttype, offset = struct.unpack("<IQ", f.read(12))
tensors.append({
"name": name,
"n_dims": n_dims,
"dims": dims,
"type": ttype,
"offset": offset,
})
header_end = f.tell()
alignment = 32
data_start = (header_end + alignment - 1) // alignment * alignment
return version, n_tensors, n_kv, kv_end, tensors, header_end, data_start
def main() -> int:
ap = argparse.ArgumentParser(description="Build Ling-3.0-tiny LUT-VQ3-64 hybrid GGUF")
ap.add_argument("--base-gguf", required=True,
help="Quantized base GGUF (e.g. Q8_0 with --tensor-type ffn_down_exps=Q6_K)")
ap.add_argument("--bf16-gguf", required=True,
help="Uncompressed BF16 master GGUF (VQ3 encode source)")
ap.add_argument("--output", required=True, help="Output VQ3 hybrid GGUF path")
ap.add_argument("--down-vq3", action="store_true",
help="Also encode ffn_down_exps to VQ3-64 (aggressive 'max' recipe)")
ap.add_argument("--lib", default=os.environ.get("VQ3_ENC_LIB", ""),
help="Optional path to libvq3_enc.so")
args = ap.parse_args()
lib = load_vq3_encoder(args.lib)
vq3_match = ("ffn_gate_exps", "ffn_up_exps") + (("ffn_down_exps",) if args.down_vq3 else ())
t0 = time.time()
with open(args.base_gguf, "rb") as ff, open(args.bf16_gguf, "rb") as fb:
_, _, _, kv_end, f_tensors, f_hdr_end, f_data_start = parse_gguf(ff)
_, _, _, _, b_tensors, _, b_data_start = parse_gguf(fb)
bf16_map = {t["name"]: t for t in b_tensors}
sorted_ft = sorted(f_tensors, key=lambda x: x["offset"])
ff.seek(0, os.SEEK_END)
f_file_size = ff.tell()
new_offset = 0
alignment = 32
vq3_target_total = 0
for t in f_tensors:
n_el = 1
for d in t["dims"]:
n_el *= d
t["n_el"] = n_el
if any(m in t["name"] for m in vq3_match) and t["name"].endswith("exps.weight"):
if n_el % 64:
raise ValueError(f"{t['name']}: {n_el} not divisible by 64")
t["new_type"] = GGML_TYPE_VQ3_0
t["new_bytes"] = (n_el // 64) * 24
t["old_bytes"] = (n_el // 64) * 30 if t["type"] == GGML_TYPE_BF16 else None
if t["old_bytes"] is None:
t["old_bytes"] = next(
(sorted_ft[i + 1]["offset"] - t["offset"] for i in range(len(sorted_ft) - 1)
if sorted_ft[i] is t), f_file_size - (f_data_start + t["offset"]))
vq3_target_total += 1
else:
t["new_type"] = t["type"]
idx_s = sorted_ft.index(t)
if idx_s + 1 < len(sorted_ft):
span = sorted_ft[idx_s + 1]["offset"] - t["offset"]
else:
span = f_file_size - (f_data_start + t["offset"])
t["old_bytes"] = span
t["new_bytes"] = span
t["new_offset"] = new_offset
padded = (t["new_bytes"] + alignment - 1) // alignment * alignment
new_offset += padded
out_gb = (f_data_start + new_offset) / (1024 ** 3)
print(f"Parsed headers: {len(f_tensors)} tensors, encoding {vq3_target_total} expert tensors "
f"to VQ3_0(44), output size ~{out_gb:.2f} GiB", flush=True)
mm_f = mmap.mmap(ff.fileno(), 0, access=mmap.ACCESS_READ)
mm_b = mmap.mmap(fb.fileno(), 0, access=mmap.ACCESS_READ)
with open(args.output, "wb") as fout:
# metadata KV (verbatim from base) + rewritten tensor directory
ff.seek(0)
fout.write(ff.read(kv_end))
for t in f_tensors:
nb = t["name"].encode("utf-8")
fout.write(struct.pack("<Q", len(nb)))
fout.write(nb)
fout.write(struct.pack("<I", t["n_dims"]))
for d in t["dims"]:
fout.write(struct.pack("<Q", d))
fout.write(struct.pack("<IQ", t["new_type"], t["new_offset"]))
assert fout.tell() == f_hdr_end, (fout.tell(), f_hdr_end)
if f_data_start > f_hdr_end:
fout.write(b"\x00" * (f_data_start - f_hdr_end))
vq3_cnt = 0
for t in f_tensors:
target_pos = f_data_start + t["new_offset"]
cur_pos = fout.tell()
if target_pos > cur_pos:
fout.write(b"\x00" * (target_pos - cur_pos))
assert fout.tell() == target_pos
if t["new_type"] == GGML_TYPE_VQ3_0:
bt = bf16_map[t["name"]]
t_type = bt["type"]
assert t_type == GGML_TYPE_BF16, \
f"Expected BF16 ({GGML_TYPE_BF16}) for {t['name']}, got {t_type}"
n_el = t["n_el"]
ncol = t["dims"][0]
nrow = n_el // ncol
b_off = b_data_start + bt["offset"]
src_ptr = np.frombuffer(mm_b, dtype=np.uint8, count=n_el * 2,
offset=b_off).ctypes.data
out_buf = np.empty(t["new_bytes"], dtype=np.uint8)
dst_ptr = out_buf.ctypes.data
ts = time.time()
lib.quantize_bf16_to_vq3_0_omp(src_ptr, dst_ptr, nrow, ncol)
fout.write(out_buf.tobytes())
vq3_cnt += 1
print(f"[{vq3_cnt:2d}/{vq3_target_total}] {t['name']} "
f"({n_el / 1e6:.1f}M w) BF16 -> VQ3_0 in {time.time() - ts:.2f}s",
flush=True)
else:
f_off = f_data_start + t["offset"]
fout.write(mm_f[f_off:f_off + t["old_bytes"]])
rem = fout.tell() % alignment
if rem:
fout.write(b"\x00" * (alignment - rem))
mm_f.close()
mm_b.close()
print(f"DONE in {time.time() - t0:.1f}s: {args.output} "
f"({os.path.getsize(args.output) / (1024 ** 3):.2f} GiB)", flush=True)
return 0
if __name__ == "__main__":
raise SystemExit(main())
|