LLM Quantization Formats: INT8, INT4, GPTQ, AWQ, GGUF

A 70B model in fp16 weighs 140 GB — it doesn't fit on two RTX 3090s. LLM quantization is the only way to shrink it to 35 GB with minimal quality loss. Over 5 years, we have helped more than 100 projects optimize inference, cutting hardware costs by up to 75%.

Why LLM Quantization Is Critical for Deployment

Insufficient VRAM is the main bottleneck when deploying large language models. A 70B model in fp16 doesn't fit into a single consumer GPU, and two RTX 3090s offer 48 GB — after quantization to INT4, there is headroom for batch processing. Inference speed jumps from 50 to over 200 tok/s, and GPU rental costs (e.g., 8×A100) drop 4x — 2×L40 suffices. Hardware savings are the key driver: quantizing a 70B model to INT4 allows deployment on two RTX 3090s instead of eight A100s, cutting capital expenditure by 4x. Reducing GPU rental costs when moving from fp16 to INT4 can reach 75% due to fewer accelerators needed.

Comparison of Quantization Formats

Each format requires a calibration dataset (128–512 examples) representative of the model's tasks. Incorrect calibration degrades quality — we tailor it to each project.

Which Quantization Format to Choose?

GPTQ: Post‑Training Quantization with Error Correction

GPTQ quantizes layer by layer, minimizing error on a small calibration dataset:

from transformers import AutoModelForCausalLM, GPTQConfig

gptq_config = GPTQConfig(
    bits=4,
    dataset="c4",
    desc_act=True,
    group_size=128,
    damp_percent=0.1,
)

model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Meta-Llama-3.1-8B-Instruct",
    quantization_config=gptq_config,
    device_map="auto"
)

model.save_pretrained("./llama3-8b-gptq-int4")

Calibration takes 30–120 minutes on CPU or GPU. As shown in GPTQ, this method delivers quality close to fp16 at 4x compression.

AWQ: Activation‑Aware Weight Quantization

AWQ identifies "important" weights based on activations and protects them from aggressive quantization:

from awq import AutoAWQForCausalLM
from transformers import AutoTokenizer

model = AutoAWQForCausalLM.from_pretrained("meta-llama/Meta-Llama-3.1-8B-Instruct")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3.1-8B-Instruct")

quant_config = {
    "zero_point": True,
    "q_group_size": 128,
    "w_bit": 4,
    "version": "GEMM"
}

model.quantize(tokenizer, quant_config=quant_config)
model.save_quantized("./llama3-8b-awq")

AWQ yields ~0.5–1% improvement in perplexity on reasoning tasks over GPTQ (see AWQ).

GGUF: Universal Format for llama.cpp

GGUF is designed for deployment via llama.cpp, supporting CPU inference and partial GPU offloading:

# Convert HuggingFace model to GGUF
python convert_hf_to_gguf.py \
  --model meta-llama/Meta-Llama-3.1-8B-Instruct \
  --outtype f16 \
  --outfile llama3-8b-f16.gguf

# Quantize to Q4_K_M (recommended)
./quantize llama3-8b-f16.gguf llama3-8b-q4km.gguf Q4_K_M

GGUF quantization variants (from best quality to smallest size):

• Q8_0: 8-bit, ~8.5GB for 8B model, excellent quality
• Q6_K: 6-bit, ~6.1GB, high quality
• Q5_K_M: 5-bit mixed, ~5.1GB, good quality
• Q4_K_M: 4-bit mixed, ~4.1GB, recommended for most tasks
• Q3_K_M: 3-bit, ~3.2GB, noticeable degradation

Step-by-Step Format Selection Algorithm

1. Determine your hardware: which GPU, VRAM capacity, is CPU inference acceptable?
2. Measure baseline: latency and throughput on fp16/bf16.
3. Select 2-3 candidates: for NVIDIA GPUs — AWQ and GPTQ; for CPU/hybrid — GGUF.
4. Perform quantization and test on your data: perplexity, task metrics, P95 latency.
5. Compare and pick the optimum. If the difference is unnoticeable, choose the format with the best support (AWQ or GGUF).

Practical Example: Deployment on 2×RTX 3090

Task: deploy a fine-tuned Llama 3.1 8B on a server with 2×RTX 3090 (48 GB total VRAM) for 50 concurrent users.

Requirements: P95 latency < 3s, throughput > 100 tok/s.

Choice: AWQ INT4 — fits on a single 3090 24GB with headroom, throughput 280 tok/s exceeds requirements, quality minimally degraded.

Inference of Quantized Model via vLLM

from vllm import LLM, SamplingParams

# AWQ model
llm = LLM(
    model="./llama3-8b-awq",
    quantization="awq",
    dtype="auto",
    gpu_memory_utilization=0.85,
)

# GPTQ model
llm = LLM(
    model="./llama3-8b-gptq-int4",
    quantization="gptq",
    dtype="auto",
)

outputs = llm.generate(["Hello, how are you?"], SamplingParams(max_tokens=200))

When Quantization Is Ineffective?

If the model already works with acceptable response time and isn't VRAM-bound, quantization is overkill. It's also unsuitable for tasks where every tenth of a percent in quality is critical (medical, legal). In such cases, we keep fp16 or bf16 but sacrifice speed.

What's Included in the Work and Timelines

• Analysis of model and hardware, selection of 2–3 formats for testing
• Quantization (GPTQ/AWQ/GGUF) with calibration on your data
• Integration via vLLM, llama.cpp, or Triton Inference Server
• Testing: P50/P95/P99 latency, throughput, quality (perplexity + task metrics)
• Documentation for deployment and operation
• Training your team to work with the quantized model

Estimated timelines:

• GPTQ/AWQ for an 8B model: 1–3 hours. For 70B: 6–18 hours.
• GGUF conversion: 15–60 minutes.
• Testing and format selection: 1–3 days.
• Total: 2–5 days turnkey.

We will assess your project in one day — contact us to select the optimal quantization format. Order a model audit and receive a quantization recommendation. Experience: over 5 years and 100+ successful cases.