Skip to content
Research Agent Skillsby Kalaris Labs
Install
ML Trainingv1.0.0AdaptedMIT

quantizing-models-bitsandbytes

Quantizes LLMs to 8-bit or 4-bit for 50-75% memory reduction with minimal accuracy loss. Use when GPU memory is limited, need to fit larger models, or want faster inference. Supports INT8, NF4, FP4 formats, QLoRA training, and 8-bit optimizers. Works with HuggingFace Transformers.

Install this skill

Works with Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode, and 70+ other harnesses.

skills.sh CLI (Recommended)
npx skills add KalarisLabs/research-agent-skills --skill quantizing-models-bitsandbytes
uvx (Python tool runner)
uvx research-agent-skills install quantizing-models-bitsandbytes

Documentation & Instructions

bitsandbytes - LLM Quantization

Quick start

bitsandbytes reduces LLM memory by 50% (8-bit) or 75% (4-bit) with <1% accuracy loss.

Installation:

pip install bitsandbytes transformers accelerate

8-bit quantization (50% memory reduction):

from transformers import AutoModelForCausalLM, BitsAndBytesConfig

config = BitsAndBytesConfig(load_in_8bit=True)
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-2-7b-hf",
    quantization_config=config,
    device_map="auto"
)

# Memory: 14GB → 7GB

4-bit quantization (75% memory reduction):

config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16
)
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-2-7b-hf",
    quantization_config=config,
    device_map="auto"
)

# Memory: 14GB → 3.5GB

Common workflows

Workflow 1: Load large model in limited GPU memory

Copy this checklist:

Quantization Loading:
- [ ] Step 1: Calculate memory requirements
- [ ] Step 2: Choose quantization level (4-bit or 8-bit)
- [ ] Step 3: Configure quantization
- [ ] Step 4: Load and verify model

Step 1: Calculate memory requirements

Estimate model memory:

FP16 memory (GB) = Parameters × 2 bytes / 1e9
INT8 memory (GB) = Parameters × 1 byte / 1e9
INT4 memory (GB) = Parameters × 0.5 bytes / 1e9

Example (Llama 2 7B):
FP16: 7B × 2 / 1e9 = 14 GB
INT8: 7B × 1 / 1e9 = 7 GB
INT4: 7B × 0.5 / 1e9 = 3.5 GB

Step 2: Choose quantization level

GPU VRAM Model Size Recommended
8 GB 3B 4-bit
12 GB 7B 4-bit
16 GB 7B 8-bit or 4-bit
24 GB 13B 8-bit or 70B 4-bit
40+ GB 70B 8-bit

Step 3: Configure quantization

For 8-bit (better accuracy):

from transformers import BitsAndBytesConfig
import torch

config = BitsAndBytesConfig(
    load_in_8bit=True,
    llm_int8_threshold=6.0,  # Outlier threshold
    llm_int8_has_fp16_weight=False
)

For 4-bit (maximum memory savings):

config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16,  # Compute in FP16
    bnb_4bit_quant_type="nf4",  # NormalFloat4 (recommended)
    bnb_4bit_use_double_quant=True  # Nested quantization
)

Step 4: Load and verify model

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-2-13b-hf",
    quantization_config=config,
    device_map="auto",  # Automatic device placement
    torch_dtype=torch.float16
)

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-13b-hf")

# Test inference
inputs = tokenizer("Hello, how are you?", return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_length=50)
print(tokenizer.decode(outputs[0]))

# Check memory
import torch
print(f"Memory allocated: {torch.cuda.memory_allocated()/1e9:.2f}GB")

Workflow 2: Fine-tune with QLoRA (4-bit training)

QLoRA enables fine-tuning large models on consumer GPUs.

Copy this checklist:

QLoRA Fine-tuning:
- [ ] Step 1: Install dependencies
- [ ] Step 2: Configure 4-bit base model
- [ ] Step 3: Add LoRA adapters
- [ ] Step 4: Train with standard Trainer

Step 1: Install dependencies

pip install bitsandbytes transformers peft accelerate datasets

Step 2: Configure 4-bit base model

from transformers import AutoModelForCausalLM, BitsAndBytesConfig
import torch

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True
)

model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-2-7b-hf",
    quantization_config=bnb_config,
    device_map="auto"
)

Step 3: Add LoRA adapters

from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training

# Prepare model for training
model = prepare_model_for_kbit_training(model)

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

# Add LoRA adapters
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# Output: trainable params: 4.2M || all params: 6.7B || trainable%: 0.06%

Step 4: Train with standard Trainer

from transformers import Trainer, TrainingArguments

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

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

trainer.train()

# Save LoRA adapters (only ~20MB)
model.save_pretrained("./qlora-adapters")

Workflow 3: 8-bit optimizer for memory-efficient training

Use 8-bit Adam/AdamW to reduce optimizer memory by 75%.

8-bit Optimizer Setup:
- [ ] Step 1: Replace standard optimizer
- [ ] Step 2: Configure training
- [ ] Step 3: Monitor memory savings

Step 1: Replace standard optimizer

import bitsandbytes as bnb
from transformers import Trainer, TrainingArguments

# Instead of torch.optim.AdamW
model = AutoModelForCausalLM.from_pretrained("model-name")

training_args = TrainingArguments(
    output_dir="./output",
    per_device_train_batch_size=8,
    optim="paged_adamw_8bit",  # 8-bit optimizer
    learning_rate=5e-5
)

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

trainer.train()

Manual optimizer usage:

import bitsandbytes as bnb

optimizer = bnb.optim.AdamW8bit(
    model.parameters(),
    lr=1e-4,
    betas=(0.9, 0.999),
    eps=1e-8
)

# Training loop
for batch in dataloader:
    loss = model(**batch).loss
    loss.backward()
    optimizer.step()
    optimizer.zero_grad()

Step 2: Configure training

Compare memory:

Standard AdamW optimizer memory = model_params × 8 bytes (states)
8-bit AdamW memory = model_params × 2 bytes
Savings = 75% optimizer memory

Example (Llama 2 7B):
Standard: 7B × 8 = 56 GB
8-bit: 7B × 2 = 14 GB
Savings: 42 GB

Step 3: Monitor memory savings

import torch

before = torch.cuda.memory_allocated()

# Training step
optimizer.step()

after = torch.cuda.memory_allocated()
print(f"Memory used: {(after-before)/1e9:.2f}GB")

When to use vs alternatives

Use bitsandbytes when:

  • GPU memory limited (need to fit larger model)
  • Training with QLoRA (fine-tune 70B on single GPU)
  • Inference only (50-75% memory reduction)
  • Using HuggingFace Transformers
  • Acceptable 0-2% accuracy degradation

Use alternatives instead:

  • GPTQ/AWQ: Production serving (faster inference than bitsandbytes)
  • GGUF: CPU inference (llama.cpp)
  • FP8: H100 GPUs (hardware FP8 faster)
  • Full precision: Accuracy critical, memory not constrained

Common issues

Issue: CUDA error during loading

Install matching CUDA version:

# Check CUDA version
nvcc --version

# Install matching bitsandbytes
pip install bitsandbytes --no-cache-dir

Issue: Model loading slow

Use CPU offload for large models:

model = AutoModelForCausalLM.from_pretrained(
    "model-name",
    quantization_config=config,
    device_map="auto",
    max_memory={0: "20GB", "cpu": "30GB"}  # Offload to CPU
)

Issue: Lower accuracy than expected

Try 8-bit instead of 4-bit:

config = BitsAndBytesConfig(load_in_8bit=True)
# 8-bit has <0.5% accuracy loss vs 1-2% for 4-bit

Or use NF4 with double quantization:

config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",  # Better than fp4
    bnb_4bit_use_double_quant=True  # Extra accuracy
)

Issue: OOM even with 4-bit

Enable CPU offload:

model = AutoModelForCausalLM.from_pretrained(
    "model-name",
    quantization_config=config,
    device_map="auto",
    offload_folder="offload",  # Disk offload
    offload_state_dict=True
)

Advanced topics

QLoRA training guide: See references/qlora-training.md for complete fine-tuning workflows, hyperparameter tuning, and multi-GPU training.

Quantization formats: See references/quantization-formats.md for INT8, NF4, FP4 comparison, double quantization, and custom quantization configs.

Memory optimization: See references/memory-optimization.md for CPU offloading strategies, gradient checkpointing, and memory profiling.

Hardware requirements

  • GPU: NVIDIA with compute capability 7.0+ (Turing, Ampere, Hopper)
  • VRAM: Depends on model and quantization
    • 4-bit Llama 2 7B: 4GB
    • 4-bit Llama 2 13B: 8GB
    • 4-bit Llama 2 70B: 24GB
  • CUDA: 11.1+ (12.0+ recommended)
  • PyTorch: 2.0+

Supported platforms: NVIDIA GPUs (primary), AMD ROCm, Intel GPUs (experimental)

Resources

Agent operating procedure

  1. Check the environment. Check GPU type, memory and driver/CUDA versions (nvidia-smi), framework versions, and dataset location and size.
  2. Pin down the inputs. Confirm formats, identifiers and parameters from the data or the user. Ask rather than guess any value that changes the result.
  3. Run a small version first. Do a smoke run: tiny model or subset, few steps, and confirm loss decreases and checkpoints save.
  4. Execute the full task using the instructions and references above.
  5. Validate the result. Track metrics on held-out data, compare against a baseline, and record seeds, configs and hardware.
  6. Report. State what was run (versions, commands, parameters), what was checked, and what is still uncertain.
If this happens Do this
CUDA out-of-memory Reduce batch size, enable gradient accumulation/checkpointing or mixed precision, or shard the model.
Loss is NaN or diverges Lower the learning rate, check data for invalid values, and enable gradient clipping.
A function, flag or endpoint in these instructions is missing in the installed version Check the installed version's own documentation (help(), --help, official docs), adapt, and tell the user. Never invent an API.
A required input, identifier or parameter is ambiguous Ask the user, or state the assumption explicitly before running.

Integrity rules

  • Never fabricate results, parameters, identifiers, citations or statistics. If something cannot be run or verified, say so plainly.
  • Never claim training results without logs; estimate compute cost before launching large jobs.
  • Treat version-specific details here as possibly outdated: confirm them against the official documentation for the installed version.
  • Ask before actions that cost money, consume shared GPUs or cloud quota, touch personal or patient data, or cannot be undone.

Related skills

  • awq-quantization: Activation-aware weight quantization for 4-bit LLM compression with 3x speedup and minimal accuracy loss.
  • gptq: Quantizes LLMs to 4-bit (also 3-bit) with GPTQ using group-wise quantization (group size 128 by default), via AutoGPTQ and transformers.
  • hqq-quantization: Half-Quadratic Quantization for LLMs without calibration data.

Trigger Prompts & Evaluation Corpus

These evaluation prompts test that the agent correctly selects quantizing-models-bitsandbytes for relevant tasks and leaves it out for near misses.

Should trigger (3)

  • should trigger

    “how do i load llama 2 13b in 4-bit on my 12GB 3060?”

    expects Gives a BitsAndBytesConfig with load_in_4bit, NF4 quant type, and bfloat16/float16 compute dtype, loaded through from_pretrained with device_map. Checks that the estimated memory fits in 12GB (about 8GB for 13B in 4-bit) and suggests CPU offload or a smaller model if it doesn't.

  • should trigger

    “I'm fine-tuning Mistral-7B on my instruction dataset `train.jsonl` with a single 24GB RTX 4090. Full-precision LoRA hits `torch.cuda.OutOfMemoryError` at step 0 with batch size 8. Can you set it up so it fits? I'd like to use the HF Trainer and keep the loss comparable to what I got on the A100.”

    expects Sets up QLoRA: a 4-bit NF4 base model with double quantization, prepare_model_for_kbit_training, and peft LoRA adapters. Uses a paged 8-bit optimizer, gradient checkpointing, a smaller per-device batch size with gradient accumulation, and a smoke run before the full run. Checks GPU and library versions first. Says plainly that no results have been measured and that loss may differ slightly from the A100 run.

  • should trigger

    “My lab only has a couple of older 16GB GPUs, and I need to run a 7B language model for inference and also cut the optimizer state memory while training a small adapter. What's the least invasive way to shrink memory without changing my HuggingFace code much?”

    expects Recommends bitsandbytes quantization through the Transformers integration: 8-bit or 4-bit loading via a quantization config, plus an 8-bit Adam/AdamW optimizer. Explains the memory and accuracy trade-offs, checks that the GPU's compute capability is supported, and recommends validating accuracy against a baseline.

Near misses (2)

  • near miss

    “I want to quantize Llama-3-8B to 4-bit with AutoGPTQ using group size 128 and serve it with the Marlin kernel. Walk me through calibration and the quantization script.”

    expects Recognizes this as a GPTQ task (calibration data, group size, Marlin kernel) and does not use bitsandbytes on-the-fly NF4 loading. Uses the gptq skill.

  • near miss

    “I want to quantize a 70B instruction-tuned model to 4-bit without any calibration data, so I can serve it with vLLM. What's the fastest approach?”

    expects Recognizes the request is for calibration-free quantization for vLLM serving, which fits Half-Quadratic Quantization. Points to the hqq-quantization skill rather than bitsandbytes.