Quantization

Quantization techniques focus on representing data with less information while also trying to not lose too much accuracy. This often means converting a data type to represent the same information with fewer bits. For example, if your model weights are stored as 32-bit floating points and they’re quantized to 16-bit floating points, this halves the model size which makes it easier to store and reduces memory-usage. Lower precision can also speedup inference because it takes less time to perform calculations with fewer bits.

Transformers supports several quantization schemes to help you run inference with large language models (LLMs) and finetune adapters on quantized models. This guide will show you how to use Activation-aware Weight Quantization (AWQ), AutoGPTQ, and bitsandbytes.

AWQ

Activation-aware Weight Quantization (AWQ) doesn’t quantize all the weights in a model, and instead, it preserves a small percentage of weights that are important for LLM performance. This significantly reduces quantization loss such that you can run models in 4-bit precision without experiencing any performance degradation.

There are several libraries for quantizing models with the AWQ algorithm, such as llm-awq, autoawq or optimum-intel. Transformers supports loading models quantized with the llm-awq and autoawq libraries. This guide will show you how to load models quantized with autoawq, but the processs is similar for llm-awq quantized models.

Make sure you have autoawq installed:

pip install autoawq

AWQ-quantized models can be identified by checking the quantization_config attribute in the model’s config.json file:

{
  "_name_or_path": "/workspace/process/huggingfaceh4_zephyr-7b-alpha/source",
  "architectures": [
    "MistralForCausalLM"
  ],
  ...
  ...
  ...
  "quantization_config": {
    "quant_method": "awq",
    "zero_point": true,
    "group_size": 128,
    "bits": 4,
    "version": "gemm"
  }
}

A quantized model is loaded with the from_pretrained() method. If you loaded your model on the CPU, make sure to move it to a GPU device first. Use the device_map parameter to specify where to place the model:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "TheBloke/zephyr-7B-alpha-AWQ"
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="cuda:0")

Loading an AWQ-quantized model automatically sets other weights to fp16 by default for performance reasons. If you want to load these other weights in a different format, use the torch_dtype parameter:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "TheBloke/zephyr-7B-alpha-AWQ"
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float32)

AWQ quantization can also be combined with FlashAttention-2 to further accelerate inference:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("TheBloke/zephyr-7B-alpha-AWQ", attn_implementation="flash_attention_2", device_map="cuda:0")

Benchmarks

We performed some speed, throughput and latency benchmarks using optimum-benchmark library.

Note at that time of writing this documentation section, the available quantization methods were: awq, gptq and bitsandbytes.

The benchmark was run on a NVIDIA-A100 instance and the model used was TheBloke/Mistral-7B-v0.1-AWQ for the AWQ model, TheBloke/Mistral-7B-v0.1-GPTQ for the GPTQ model. We also benchmarked it against bitsandbytes quantization methods and native float16 model. Some results are shown below:

You can find the full results together with packages versions in this link.

From the results it appears that AWQ quantization method is the fastest quantization method for inference, text generation and among the lowest peak memory for text generation. However, AWQ seems to have the largest forward latency per batch size.

Make use of fused modules

You can benefit from fused modules by passing an AwqConfig with fuse_modules=True and your expected maximum sequence length for generation to fuse_max_seq_len. For architectures that do not support do_fuse=True, you can still fuse the modules, however you need to pass a custom fusing_mapping to AwqConfig(). Let’s dive into these specific usecases.

Note that you cannot combine fusing modules and other optimization techniques such as Flash Attention 2.

Fusing modules for supported architectures

Currently we support out of the box AWQ module fusing for llama and mistral.

To enable this feature for supported architectures simply create an AwqConfig and pass the arguments fuse_max_seq_len and do_fuse=True.

For example to enable module fusing for the model TheBloke/Mistral-7B-OpenOrca-AWQ, run:

import torch
from transformers import AwqConfig, AutoModelForCausalLM

model_id = "TheBloke/Mistral-7B-OpenOrca-AWQ"

quantization_config = AwqConfig(
    bits=4,
    fuse_max_seq_len=512,
    do_fuse=True,
)

model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=quantization_config).to(0)

Note that you need to define fuse_max_seq_len to AwqConfig. That total sequence length should include the context length and the expected generation length. You can set it to a large value to be on the safe zone.

You can also apply module fusing for other architectures that are not supported.

Fusing modules for unsupported architectures

For architectures that do not support out of the box module fusing, you can pass a custom fusing mapping; simply pass a dictionnary modules_to_fuse to AwqConfig, let’s take an example with the Yi model:

import torch
from transformers import AwqConfig, AutoModelForCausalLM

model_id = "TheBloke/Yi-34B-AWQ"

quantization_config = AwqConfig(
    bits=4,
    fuse_max_seq_len=512,
    modules_to_fuse={
        "attention": ["q_proj", "k_proj", "v_proj", "o_proj"],
        "layernorm": ["ln1", "ln2", "norm"],
        "mlp": ["gate_proj", "up_proj", "down_proj"],
        "use_alibi": False,
        "num_attention_heads": 56,
        "num_key_value_heads": 8,
        "hidden_size": 7168
    }
)

model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=quantization_config).to(0)

The parameter modules_to_fuse needs to have the following respective fields:

• "attention": The names of the attention layers to fuse - in the order: query, key, value and output projection layer. In case you don’t want to fuse the attention layers you can pass an empty list.
• "layernorm": The names of all the layernorm layers you want to replace with a custom fused layer norm. In case you don’t want to fuse these layers you can also pass an empty list.
• "mlp": The names of the MLP layers you want to fuse into a single MLP layer in the order: (gate (dense layer post-attention) / up / down layers).
• "use_alibi": If you model uses alibi positional embedding
• "num_attention_heads": The number of attention heads
• "num_key_value_heads": This is the number of key value heads that should be used to implement Grouped Query Attention. If num_key_value_heads=num_attention_heads, the model will use Multi Head Attention (MHA), if num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used.
• "hidden_size": Dimension of the hidden representations.

Benchmarks

We benchmarked the model with and without fused modules first using only batch_size=1 on the TheBloke/Mistral-7B-OpenOrca-AWQ model and below are the results:

unfused case

fused case

We also performed benchmarks with optimum-benchmark library. And below are the results:

AutoGPTQ

The AutoGPTQ library implements the GPTQ algorithm, a post-training quantization technique where each row of the weight matrix is quantized independently to find a version of the weights that minimizes the error. These weights are quantized to int4, but they’re restored to fp16 on the fly during inference. This can save your memory-usage by 4x because the int4 weights are dequantized in a fused kernel rather than a GPU’s global memory, and you can also expect a speedup in inference because using a lower bitwidth takes less time to communicate.

Before you begin, make sure the following libraries are installed:

pip install auto-gptq
pip install git+https://github.com/huggingface/optimum.git
pip install git+https://github.com/huggingface/transformers.git
pip install --upgrade accelerate

To quantize a model (currently only supported for text models), you need to create a GPTQConfig class and set the number of bits to quantize to, a dataset to calibrate the weights for quantization, and a tokenizer to prepare the dataset.

from transformers import AutoModelForCausalLM, AutoTokenizer, GPTQConfig

model_id = "facebook/opt-125m"
tokenizer = AutoTokenizer.from_pretrained(model_id)
gptq_config = GPTQConfig(bits=4, dataset="c4", tokenizer=tokenizer)

You could also pass your own dataset as a list of strings, but it is highly recommended to use the same dataset from the GPTQ paper.

dataset = ["auto-gptq is an easy-to-use model quantization library with user-friendly apis, based on GPTQ algorithm."]
gptq_config = GPTQConfig(bits=4, dataset=dataset, tokenizer=tokenizer)

Load a model to quantize and pass the gptq_config to the from_pretrained() method. Set device_map="auto" to automatically offload the model to a CPU to help fit the model in memory, and allow the model modules to be moved between the CPU and GPU for quantization.

quantized_model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", quantization_config=gptq_config)

If you’re running out of memory because a dataset is too large, disk offloading is not supported. If this is the case, try passing the max_memory parameter to allocate the amount of memory to use on your device (GPU and CPU):

quantized_model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", max_memory={0: "30GiB", 1: "46GiB", "cpu": "30GiB"}, quantization_config=gptq_config)

Once your model is quantized, you can push the model and tokenizer to the Hub where it can be easily shared and accessed. Use the push_to_hub() method to save the GPTQConfig:

quantized_model.push_to_hub("opt-125m-gptq")
tokenizer.push_to_hub("opt-125m-gptq")

You could also save your quantized model locally with the save_pretrained() method. If the model was quantized with the device_map parameter, make sure to move the entire model to a GPU or CPU before saving it. For example, to save the model on a CPU:

quantized_model.save_pretrained("opt-125m-gptq")
tokenizer.save_pretrained("opt-125m-gptq")

# if quantized with device_map set
quantized_model.to("cpu")
quantized_model.save_pretrained("opt-125m-gptq")

Reload a quantized model with the from_pretrained() method, and set device_map="auto" to automatically distribute the model on all available GPUs to load the model faster without using more memory than needed.

from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("{your_username}/opt-125m-gptq", device_map="auto")

ExLlama

ExLlama is a Python/C++/CUDA implementation of the Llama model that is designed for faster inference with 4-bit GPTQ weights (check out these benchmarks). The ExLlama kernel is activated by default when you create a GPTQConfig object. To boost inference speed even further, use the ExLlamaV2 kernels by configuring the exllama_config parameter:

import torch
from transformers import AutoModelForCausalLM, GPTQConfig

gptq_config = GPTQConfig(bits=4, exllama_config={"version":2})
model = AutoModelForCausalLM.from_pretrained("{your_username}/opt-125m-gptq", device_map="auto", quantization_config=gptq_config)

The ExLlama kernels are only supported when the entire model is on the GPU. If you’re doing inference on a CPU with AutoGPTQ (version > 0.4.2), then you’ll need to disable the ExLlama kernel. This overwrites the attributes related to the ExLlama kernels in the quantization config of the config.json file.

import torch
from transformers import AutoModelForCausalLM, GPTQConfig
gptq_config = GPTQConfig(bits=4, use_exllama=False)
model = AutoModelForCausalLM.from_pretrained("{your_username}/opt-125m-gptq", device_map="cpu", quantization_config=gptq_config)

bitsandbytes

bitsandbytes is the easiest option for quantizing a model to 8 and 4-bit. 8-bit quantization multiplies outliers in fp16 with non-outliers in int8, converts the non-outlier values back to fp16, and then adds them together to return the weights in fp16. This reduces the degradative effect outlier values have on a model’s performance. 4-bit quantization compresses a model even further, and it is commonly used with QLoRA to finetune quantized LLMs.

To use bitsandbytes, make sure you have the following libraries installed:

pip install transformers accelerate bitsandbytes>0.37.0

Now you can quantize a model with the load_in_8bit or load_in_4bit parameters in the from_pretrained() method. This works for any model in any modality, as long as it supports loading with Accelerate and contains torch.nn.Linear layers.

from transformers import AutoModelForCausalLM

model_8bit = AutoModelForCausalLM.from_pretrained("bigscience/bloom-1b7", device_map="auto", load_in_8bit=True)

import torch
from transformers import AutoModelForCausalLM

model_8bit = AutoModelForCausalLM.from_pretrained("facebook/opt-350m", load_in_8bit=True, torch_dtype=torch.float32)
model_8bit.model.decoder.layers[-1].final_layer_norm.weight.dtype

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("bigscience/bloom-560m", device_map="auto", load_in_8bit=True)
tokenizer = AutoTokenizer.from_pretrained("bigscience/bloom-560m")

model.push_to_hub("bloom-560m-8bit")

You can check your memory footprint with the get_memory_footprint method:

print(model.get_memory_footprint())

Quantized models can be loaded from the from_pretrained() method without needing to specify the load_in_8bit or load_in_4bit parameters:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("{your_username}/bloom-560m-8bit", device_map="auto")

8-bit

This section explores some of the specific features of 8-bit models, such as offloading, outlier thresholds, skipping module conversion, and finetuning.

Offloading

8-bit models can offload weights between the CPU and GPU to support fitting very large models into memory. The weights dispatched to the CPU are actually stored in float32, and aren’t converted to 8-bit. For example, to enable offloading for the bigscience/bloom-1b7 model, start by creating a BitsAndBytesConfig:

from transformers import AutoModelForCausalLM, BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(llm_int8_enable_fp32_cpu_offload=True)

Design a custom device map to fit everything on your GPU except for the lm_head, which you’ll dispatch to the CPU:

device_map = {
    "transformer.word_embeddings": 0,
    "transformer.word_embeddings_layernorm": 0,
    "lm_head": "cpu",
    "transformer.h": 0,
    "transformer.ln_f": 0,
}

Now load your model with the custom device_map and quantization_config:

model_8bit = AutoModelForCausalLM.from_pretrained(
    "bigscience/bloom-1b7",
    device_map=device_map,
    quantization_config=quantization_config,
)

Outlier threshold

An “outlier” is a hidden state value greater than a certain threshold, and these values are computed in fp16. While the values are usually normally distributed ([-3.5, 3.5]), this distribution can be very different for large models ([-60, 6] or [6, 60]). 8-bit quantization works well for values ~5, but beyond that, there is a significant performance penalty. A good default threshold value is 6, but a lower threshold may be needed for more unstable models (small models or finetuning).

To find the best threshold for your model, we recommend experimenting with the llm_int8_threshold parameter in BitsAndBytesConfig:

from transformers import AutoModelForCausalLM, BitsAndBytesConfig

model_id = "bigscience/bloom-1b7"

quantization_config = BitsAndBytesConfig(
    llm_int8_threshold=10,
)

model_8bit = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map=device_map,
    quantization_config=quantization_config,
)

Skip module conversion

For some models, like Jukebox, you don’t need to quantize every module to 8-bit which can actually cause instability. With Jukebox, there are several lm_head modules that should be skipped using the llm_int8_skip_modules parameter in BitsAndBytesConfig:

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

model_id = "bigscience/bloom-1b7"

quantization_config = BitsAndBytesConfig(
    llm_int8_skip_modules=["lm_head"],
)

model_8bit = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    quantization_config=quantization_config,
)

Finetuning

With the PEFT library, you can finetune large models like flan-t5-large and facebook/opt-6.7b with 8-bit quantization. You don’t need to pass the device_map parameter for training because it’ll automatically load your model on a GPU. However, you can still customize the device map with the device_map parameter if you want to (device_map="auto" should only be used for inference).

4-bit

This section explores some of the specific features of 4-bit models, such as changing the compute data type, using the Normal Float 4 (NF4) data type, and using nested quantization.

Compute data type

To speedup computation, you can change the data type from float32 (the default value) to bf16 using the bnb_4bit_compute_dtype parameter in BitsAndBytesConfig:

import torch
from transformers import BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16)

Normal Float 4 (NF4)

NF4 is a 4-bit data type from the QLoRA paper, adapted for weights initialized from a normal distribution. You should use NF4 for training 4-bit base models. This can be configured with the bnb_4bit_quant_type parameter in the BitsAndBytesConfig:

from transformers import BitsAndBytesConfig

nf4_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
)

model_nf4 = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=nf4_config)

For inference, the bnb_4bit_quant_type does not have a huge impact on performance. However, to remain consistent with the model weights, you should use the bnb_4bit_compute_dtype and torch_dtype values.

Nested quantization

Nested quantization is a technique that can save additional memory at no additional performance cost. This feature performs a second quantization of the already quantized weights to save an addition 0.4 bits/parameter. For example, with nested quantization, you can finetune a Llama-13b model on a 16GB NVIDIA T4 GPU with a sequence length of 1024, a batch size of 1, and enabling gradient accumulation with 4 steps.

from transformers import BitsAndBytesConfig

double_quant_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
)

model_double_quant = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-13b", quantization_config=double_quant_config)

Optimum

The Optimum library supports quantization for Intel, Furiosa, ONNX Runtime, GPTQ, and lower-level PyTorch quantization functions. Consider using Optimum for quantization if you’re using specific and optimized hardware like Intel CPUs, Furiosa NPUs or a model accelerator like ONNX Runtime.

Benchmarks

To compare the speed, throughput, and latency of each quantization scheme, check the following benchmarks obtained from the optimum-benchmark library. The benchmark was run on a NVIDIA A1000 for the TheBloke/Mistral-7B-v0.1-AWQ and TheBloke/Mistral-7B-v0.1-GPTQ models. These were also tested against the bitsandbytes quantization methods as well as a native fp16 model.

The benchmarks indicate AWQ quantization is the fastest for inference, text generation, and has the lowest peak memory for text generation. However, AWQ has the largest forward latency per batch size. For a more detailed discussion about the pros and cons of each quantization method, read the Overview of natively supported quantization schemes in 🤗 Transformers blog post.