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.

LLM Development: Fine-Tuning, RAG, Agents, and Production Deployment

Using GPT‑4 or Claude 3.5 Sonnet through a public API is not a solution — it's just a tool. When the requirement is to "make it like ChatGPT, but on our data," there is a real engineering challenge behind it: from prompt engineering to training a 70B model on your own infrastructure. End-to-end LLM solution development is a complex stack, and we have been doing it for over 5 years. During this time, we have completed over 20 projects in generative AI: from RAG systems for legal departments to custom support agents. Where exactly your task falls depends on data, latency requirements, budget, and how critical confidentiality is.

A typical situation: the client has already tried ChatGPT, but results are unstable — sometimes accurate, sometimes hallucinating. Or they need integration into a corporate portal while complying with security policies. Let's break down each layer of the stack in detail — from RAG to production deployment.

Why Do RAG Systems Break and How to Fix It?

RAG (Retrieval-Augmented Generation) looks simple: find relevant documents, put them in context, get an answer. In practice, it fails in several places.

Chunking without overlap. Classic mistake: chunk_size=512, overlap=0. If the answer lies across two chunks, retrieval won't find either with sufficient confidence. Solution: overlap 15–25% of chunk_size, or better yet, sentence-aware splitting with spaCy or NLTK instead of naive character splitting.

Poor embedder. text-embedding-ada-002 is good for general use, but on legal or medical texts, specialized models like E5-large-v2, BGE-M3, or fine-tuned sentence-transformers on domain data outperform it. Recall@5 differences can be 15–25%.

No re-ranking. Vector search optimizes for speed, not relevance. A cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2, bge-reranker-large) after initial retrieval improves top-3 accuracy with acceptable latency (+50–150ms). This is often more impactful than improving the embedding model.

Hybrid search. Dense vectors alone work poorly on exact queries: names, SKUs, codes. BM25 (sparse) finds exact matches but misses semantics. Hybrid via RRF (Reciprocal Rank Fusion) is the optimal compromise. Qdrant, Weaviate, and pgvector 0.7+ support hybrid search natively.

When to Fine-Tune Instead of Prompt Engineering?

Prompt engineering solves ~70% of LLM adaptation tasks for a domain. The remaining 30% require fine-tuning. Three indicators: the model ignores a specific output format even with detailed prompting; the task requires deep knowledge of specialized vocabulary (medicine, law); you need to significantly reduce token costs by replacing a large model with a smaller specialized one.

LoRA and QLoRA are the standard for SFT. LoRA adds trainable low-rank matrices to attention layers. A typical configuration for Llama-3 8B: r=64, lora_alpha=128, target_modules=["q_proj","v_proj","k_proj","o_proj"] yields ~0.8% trainable parameters, training on one A100 40GB. QLoRA adds 4-bit quantization (NF4) and allows fine-tuning 70B models on two A100 40GB, though speed drops by half compared to bf16.

DPO instead of RLHF. Direct Preference Optimization requires only (chosen, rejected) pairs, not scalar reward signals. DPOTrainer from the trl library (Hugging Face) implements it in a few dozen lines.

Common mistake. A dataset of 500 examples, 5 epochs, validation loss 0.8 — seems fine. But on test, the model degrades on general instructions. Cause: catastrophic forgetting. Solution: add 10–20% general instruction-following examples (Alpaca, FLAN) to the training set to preserve original capabilities.

How to Choose a Base Model: 8B or 70B?

For most tasks, fine-tuning an 8B model is sufficient. 70B is needed when deep reasoning is required or the 8B baseline does not reach the required quality even after fine-tuning. Inference cost for Llama-3 8B via vLLM on A100 is efficient; the exact cost depends on volume.

What Does PagedAttention Bring to Production?

vLLM is the first choice for serving open-source models. PagedAttention is the key technical innovation: KV-cache is managed like virtual memory in an OS, without fragmentation. This yields 2–4x higher throughput compared to naive HuggingFace Transformers inference. The vLLM documentation confirms that continuous batching and PagedAttention are the standard for high-load LLM services.

Typical numbers on A100 80GB for Llama-3 8B (bf16): 400–600 req/s, P50 latency 200–400ms, P99 latency 600–900ms at concurrency 64. For 70B on two A100 with tensor parallelism: 80–120 req/s, P99 latency 1.5–2.5s. AWQ or GPTQ quantization reduces memory consumption by 2x with quality loss within 1–3%.

Multi-Agent Systems

Agents are LLMs with access to tools: search, code execution, API calls, database interaction. Common patterns:

• ReAct (Reason + Act): the model reasons → chooses a tool → observes the result → reasons again. LangChain and LlamaIndex implement it out of the box.
• Multi-agent orchestration: multiple specialized agents with a coordinator on top. Example: coordinator → researcher (search + summarization) → coder (code generation and execution) → critic (verification). Tools: AutoGen (Microsoft), CrewAI, custom implementation on LangGraph.

In production, agent systems are non-deterministic. Essential: guardrails, step limits, logging of each step, human-in-the-loop for critical actions.

How We Work: Stages, Timeline, Deliverables

What Is Included

We deliver:

• Technical documentation (model card, configs, deployment instructions)
• Access to infrastructure (code repository, trained weights)
• 1 month of post-deployment support (consultations, bug fixes)
• Customer team training (2–3 sessions on system operation)

Timeline: basic RAG prototype — 1–2 weeks. Fine-tuning with customer data — 3–6 weeks (including data preparation). Production system with monitoring and retraining — 2–4 months. Cost is calculated individually based on data volume, model complexity, and infrastructure requirements.

We guarantee the quality of the final model with performance benchmarks and ongoing monitoring. Our engineers have hands‑on experience with dozens of production LLM systems.

Want to evaluate your project? Leave a request — we will prepare a preliminary summary within 1–2 business days. Or get a consultation on choosing the approach: RAG, fine-tuning, or hybrid — we will tell you what works best for you. Contact us to discuss your LLM development needs. Schedule a free consultation today.