hqq-quantization

Other ways of shrinking an AI model want a pile of sample text first and hours of processing; this one needs neither — point it at the model and a much smaller version comes back in minutes.

WHEN TO USE

SCOPE

USE FOR

Serve a shrunk model through vLLM so many people can query it at once.
Push a model down to 2-bit or 1-bit precision for extreme compression experiments.
Apply a different bit-width to attention versus MLP layers in the same model.
Fine-tune an HQQ-quantized model with LoRA using the standard HuggingFace Trainer.
Switch to a faster backend, like Marlin or TorchAO, once quantization is done.

NOT FOR

For calibration-based accuracy in production serving, use AWQ instead.
For maximum accuracy when calibration data is available, use GPTQ instead.
For simple 8-bit/4-bit quantization without custom backends, use bitsandbytes instead.
For CPU or Apple Silicon deployment, use llama.cpp/GGUF instead.

This is a saved copy of the skill's SKILL.md instructions. It may differ from the upstream file or the version you install.

In this document (12)

HQQ - Half-Quadratic Quantization

Fast, calibration-free weight quantization supporting 8/4/3/2/1-bit precision with multiple optimized backends.

When to use HQQ

Use HQQ when:

Quantizing models without calibration data (no dataset needed)
Need fast quantization (minutes vs hours for GPTQ/AWQ)
Deploying with vLLM or HuggingFace Transformers
Fine-tuning quantized models with LoRA/PEFT
Experimenting with extreme quantization (2-bit, 1-bit)

Key advantages:

No calibration: Quantize any model instantly without sample data
Multiple backends: PyTorch, ATEN, TorchAO, Marlin, BitBlas for optimized inference
Flexible precision: 8/4/3/2/1-bit with configurable group sizes
Framework integration: Native HuggingFace and vLLM support
PEFT compatible: Fine-tune quantized models with LoRA

Use alternatives instead:

AWQ: Need calibration-based accuracy, production serving
GPTQ: Maximum accuracy with calibration data available
bitsandbytes: Simple 8-bit/4-bit without custom backends
llama.cpp/GGUF: CPU inference, Apple Silicon deployment

Quick start

Installation

pip install hqq

# With specific backend
pip install hqq[torch]      # PyTorch backend
pip install hqq[torchao]    # TorchAO int4 backend
pip install hqq[bitblas]    # BitBlas backend
pip install hqq[marlin]     # Marlin backend

Basic quantization

from hqq.core.quantize import BaseQuantizeConfig, HQQLinear
import torch.nn as nn

# Configure quantization
config = BaseQuantizeConfig(
    nbits=4,           # 4-bit quantization
    group_size=64,     # Group size for quantization
    axis=1             # Quantize along output dimension
)

# Quantize a linear layer
linear = nn.Linear(4096, 4096)
hqq_linear = HQQLinear(linear, config)

# Use normally
output = hqq_linear(input_tensor)

Quantize full model with HuggingFace

from transformers import AutoModelForCausalLM, HqqConfig

# Configure HQQ
quantization_config = HqqConfig(
    nbits=4,
    group_size=64,
    axis=1
)

# Load and quantize
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B",
    quantization_config=quantization_config,
    device_map="auto"
)

# Model is quantized and ready to use

Core concepts

Quantization configuration

HQQ uses BaseQuantizeConfig to define quantization parameters:

from hqq.core.quantize import BaseQuantizeConfig

# Standard 4-bit config
config_4bit = BaseQuantizeConfig(
    nbits=4,           # Bits per weight (1-8)
    group_size=64,     # Weights per quantization group
    axis=1             # 0=input dim, 1=output dim
)

# Aggressive 2-bit config
config_2bit = BaseQuantizeConfig(
    nbits=2,
    group_size=16,     # Smaller groups for low-bit
    axis=1
)

# Mixed precision per layer type
layer_configs = {
    "self_attn.q_proj": BaseQuantizeConfig(nbits=4, group_size=64),
    "self_attn.k_proj": BaseQuantizeConfig(nbits=4, group_size=64),
    "self_attn.v_proj": BaseQuantizeConfig(nbits=4, group_size=64),
    "mlp.gate_proj": BaseQuantizeConfig(nbits=2, group_size=32),
    "mlp.up_proj": BaseQuantizeConfig(nbits=2, group_size=32),
    "mlp.down_proj": BaseQuantizeConfig(nbits=4, group_size=64),
}

HQQLinear layer

The core quantized layer that replaces nn.Linear:

from hqq.core.quantize import HQQLinear
import torch

# Create quantized layer
linear = torch.nn.Linear(4096, 4096)
hqq_layer = HQQLinear(linear, config)

# Access quantized weights
W_q = hqq_layer.W_q           # Quantized weights
scale = hqq_layer.scale       # Scale factors
zero = hqq_layer.zero         # Zero points

# Dequantize for inspection
W_dequant = hqq_layer.dequantize()

Backends

HQQ supports multiple inference backends for different hardware:

from hqq.core.quantize import HQQLinear

# Available backends
backends = [
    "pytorch",          # Pure PyTorch (default)
    "pytorch_compile",  # torch.compile optimized
    "aten",            # Custom CUDA kernels
    "torchao_int4",    # TorchAO int4 matmul
    "gemlite",         # GemLite CUDA kernels
    "bitblas",         # BitBlas optimized
    "marlin",          # Marlin 4-bit kernels
]

# Set backend globally
HQQLinear.set_backend("torchao_int4")

# Or per layer
hqq_layer.set_backend("marlin")

Backend selection guide:

Backend	Best For	Requirements
pytorch	Compatibility	Any GPU
pytorch_compile	Moderate speedup	torch>=2.0
aten	Good balance	CUDA GPU
torchao_int4	4-bit inference	torchao installed
marlin	Maximum 4-bit speed	Ampere+ GPU
bitblas	Flexible bit-widths	bitblas installed

HuggingFace integration

Load pre-quantized models

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load HQQ-quantized model from Hub
model = AutoModelForCausalLM.from_pretrained(
    "mobiuslabsgmbh/Llama-3.1-8B-HQQ-4bit",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B")

# Use normally
inputs = tokenizer("Hello, world!", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=50)

Quantize and save

from transformers import AutoModelForCausalLM, HqqConfig

# Quantize
config = HqqConfig(nbits=4, group_size=64)
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B",
    quantization_config=config,
    device_map="auto"
)

# Save quantized model
model.save_pretrained("./llama-8b-hqq-4bit")

# Push to Hub
model.push_to_hub("my-org/Llama-3.1-8B-HQQ-4bit")

Mixed precision quantization

from transformers import AutoModelForCausalLM, HqqConfig

# Different precision per layer type
config = HqqConfig(
    nbits=4,
    group_size=64,
    # Attention layers: higher precision
    # MLP layers: lower precision for memory savings
    dynamic_config={
        "attn": {"nbits": 4, "group_size": 64},
        "mlp": {"nbits": 2, "group_size": 32}
    }
)

vLLM integration

Serve HQQ models with vLLM

from vllm import LLM, SamplingParams

# Load HQQ-quantized model
llm = LLM(
    model="mobiuslabsgmbh/Llama-3.1-8B-HQQ-4bit",
    quantization="hqq",
    dtype="float16"
)

# Generate
sampling_params = SamplingParams(temperature=0.7, max_tokens=100)
outputs = llm.generate(["What is machine learning?"], sampling_params)

vLLM with custom HQQ config

from vllm import LLM

llm = LLM(
    model="meta-llama/Llama-3.1-8B",
    quantization="hqq",
    quantization_config={
        "nbits": 4,
        "group_size": 64
    }
)

PEFT/LoRA fine-tuning

Fine-tune quantized models

from transformers import AutoModelForCausalLM, HqqConfig
from peft import LoraConfig, get_peft_model

# Load quantized model
quant_config = HqqConfig(nbits=4, group_size=64)
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B",
    quantization_config=quant_config,
    device_map="auto"
)

# Apply LoRA
lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)

model = get_peft_model(model, lora_config)

# Train normally with Trainer or custom loop

QLoRA-style training

from transformers import TrainingArguments, Trainer

training_args = TrainingArguments(
    output_dir="./hqq-lora-output",
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    learning_rate=2e-4,
    num_train_epochs=3,
    fp16=True,
    logging_steps=10,
    save_strategy="epoch"
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    data_collator=data_collator
)

trainer.train()

Quantization workflows

Workflow 1: Quick model compression

from transformers import AutoModelForCausalLM, AutoTokenizer, HqqConfig

# 1. Configure quantization
config = HqqConfig(nbits=4, group_size=64)

# 2. Load and quantize (no calibration needed!)
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B",
    quantization_config=config,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B")

# 3. Verify quality
prompt = "The capital of France is"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=20)
print(tokenizer.decode(outputs[0]))

# 4. Save
model.save_pretrained("./llama-8b-hqq")
tokenizer.save_pretrained("./llama-8b-hqq")

Workflow 2: Optimize for inference speed

from hqq.core.quantize import HQQLinear
from transformers import AutoModelForCausalLM, HqqConfig

# 1. Quantize with optimal backend
config = HqqConfig(nbits=4, group_size=64)
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B",
    quantization_config=config,
    device_map="auto"
)

# 2. Set fast backend
HQQLinear.set_backend("marlin")  # or "torchao_int4"

# 3. Compile for additional speedup
import torch
model = torch.compile(model)

# 4. Benchmark
import time
inputs = tokenizer("Hello", return_tensors="pt").to(model.device)
start = time.time()
for _ in range(10):
    model.generate(**inputs, max_new_tokens=100)
print(f"Avg time: {(time.time() - start) / 10:.2f}s")

Best practices

Start with 4-bit: Best quality/size tradeoff for most models
Use group_size=64: Good balance; smaller for extreme quantization
Choose backend wisely: Marlin for 4-bit Ampere+, TorchAO for flexibility
Verify quality: Always test generation quality after quantization
Mixed precision: Keep attention at higher precision, compress MLP more
PEFT training: Use LoRA r=16-32 for good fine-tuning results

Common issues

Out of memory during quantization:

# Quantize layer-by-layer
from hqq.models.hf.base import AutoHQQHFModel

model = AutoHQQHFModel.from_pretrained(
    "meta-llama/Llama-3.1-8B",
    quantization_config=config,
    device_map="sequential"  # Load layers sequentially
)

Slow inference:

# Switch to optimized backend
from hqq.core.quantize import HQQLinear
HQQLinear.set_backend("marlin")  # Requires Ampere+ GPU

# Or compile
model = torch.compile(model, mode="reduce-overhead")

Poor quality at 2-bit:

# Use smaller group size
config = BaseQuantizeConfig(
    nbits=2,
    group_size=16,  # Smaller groups help at low bits
    axis=1
)

References

Advanced Usage - Custom backends, mixed precision, optimization
Troubleshooting - Common issues, debugging, benchmarks

Resources

Repository: https://github.com/mobiusml/hqq
Paper: Half-Quadratic Quantization
HuggingFace Models: https://huggingface.co/mobiuslabsgmbh
Version: 0.2.0+
License: Apache 2.0

This skill ships inside the AI Research Skills plugin.

Run this outside your AI tool, using Node.js 22.20.0 or newer. Skills CLI 1.7.2 skips skill selection; prompts cancel installation without an interactive terminal. If choosing Project when asked about installation scope, open your project's folder before running this.

Install checklist

Check selected tools if a picker appears; Claude Code may be unchecked under Additional agents. ↑/↓ moves, Space toggles selections, Enter confirms.
Without a tool picker, check Installing to: for your intended tool. Stop if it names a different tool.
If asked, choose Symlink (Recommended) for shared files or Copy to all agents for separate copies; review the summary before confirming installation.
Install the find-skills skill? defaults to Yes, adding another skill for all projects. Choose No if you don't want that extra installation.

Skills CLI 1.7.2 sends install data to skills.sh by default, including the source, skill names, target tools and file details; GitHub sources are reported only when GitHub confirms the repository is public, while other remote sources can also be reported. Confirmed-public GitHub sources also trigger a security-audit request. Before installing, set the environment variable DO_NOT_TRACK=1 or DISABLE_TELEMETRY=1 to disable both.

npx skills add orchestra-research/ai-research-skills --skill "hqq-quantization" --full-depth

After installing for Claude Code, open it in the installation project if you chose Project; Global makes the skill available across projects on this computer. Once Claude Code loads the skill, its description helps Claude decide when to use it.

If installation created a top-level skills folder that did not exist when your Claude Code session started, run /reload-skills in that session. Run it again after later changes in that folder.

What's inside

Files included with this skill.

View on GitHub