torchao

torchao is a PyTorch architecture optimization library with support for custom high performance data types, quantization, and sparsity. It is composable with native PyTorch features such as torch.compile for even faster inference and training.

Install torchao with the following command.

# Updating 🤗 Transformers to the latest version, as the example script below uses the new auto compilation
pip install --upgrade torch torchao transformers

torchao supports many quantization types for different data types (int4, float8, weight only, etc.). Starting with version 0.10.0, torchao provides enhanced flexibility through the AOBaseConfig API, allowing for more customized quantization configurations. And full access to the techniques offered in the torchao library.

You can manually choose the quantization types and settings or automatically select the quantization types.

manual

automatic

Create a TorchAoConfig and specify the quantization type and group_size of the weights to quantize. Set the cache_implementation to "static" to automatically torch.compile the forward method.

Run the quantized model on a CPU by changing device_map to "cpu" and layout to Int4CPULayout(). This is only available in torchao 0.8.0+.

In torchao 0.10.0+, you can use the more flexible AOBaseConfig approach instead of string identifiers:

import torch
from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
from torchao.quantization import Int4WeightOnlyConfig

# Using AOBaseConfig instance (torchao >= 0.10.0)
quant_config = Int4WeightOnlyConfig(group_size=128)
quantization_config = TorchAoConfig(quant_type=quant_config)

# Load and quantize the model
quantized_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Meta-Llama-3-8B",
    torch_dtype="auto",
    device_map="auto",
    quantization_config=quantization_config
)

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B")
input_text = "What are we having for dinner?"
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")

# auto-compile the quantized model with `cache_implementation="static"` to get speed up
output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
print(tokenizer.decode(output[0], skip_special_tokens=True))

Available Quantization Schemes

TorchAO provides a variety of quantization configurations:

Int4WeightOnlyConfig
Int8WeightOnlyConfig
Int8DynamicActivationInt8WeightConfig
Float8WeightOnlyConfig

Each configuration can be further customized with parameters such as group_size, scheme, and layout to optimize for specific hardware and model architectures.

For a complete list of available configurations, see our quantization API documentation.

⚠️ DEPRECATION WARNING

Starting with version 0.10.0, the string-based API for quantization configuration (e.g., TorchAoConfig("int4_weight_only", group_size=128)) is deprecated and will be removed in a future release.

Please use the new AOBaseConfig-based approach instead:
# Old way (deprecated)
quantization_config = TorchAoConfig("int4_weight_only", group_size=128)

# New way (recommended)
from torchao.quantization import Int4WeightOnlyConfig
quant_config = Int4WeightOnlyConfig(group_size=128)
quantization_config = TorchAoConfig(quant_type=quant_config)
The new API offers greater flexibility, better type safety, and access to the full range of features available in torchao.

Migration Guide

Here’s how to migrate from common string identifiers to their AOBaseConfig equivalents:

Old String API New AOBaseConfig API

"int4_weight_only" Int4WeightOnlyConfig()

"int8_weight_only" Int8WeightOnlyConfig()

"int8_dynamic_activation_int8_weight" Int8DynamicActivationInt8WeightConfig()

All configuration objects accept parameters for customization (e.g., group_size, scheme, layout).

Old String API	New `AOBaseConfig` API
`"int4_weight_only"`	`Int4WeightOnlyConfig()`
`"int8_weight_only"`	`Int8WeightOnlyConfig()`
`"int8_dynamic_activation_int8_weight"`	`Int8DynamicActivationInt8WeightConfig()`

Below is the API for for torchao < 0.9.0

import torch
from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer

quantization_config = TorchAoConfig("int4_weight_only", group_size=128)
quantized_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Meta-Llama-3-8B",
    torch_dtype="auto",
    device_map="auto",
    quantization_config=quantization_config
)

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B")
input_text = "What are we having for dinner?"
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")

# auto-compile the quantized model with `cache_implementation="static"` to get speed up
output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
print(tokenizer.decode(output[0], skip_special_tokens=True))

Run the code below to benchmark the quantized models performance.

from torch._inductor.utils import do_bench_using_profiling
from typing import Callable

def benchmark_fn(func: Callable, *args, **kwargs) -> float:
    """Thin wrapper around do_bench_using_profiling"""
    no_args = lambda: func(*args, **kwargs)
    time = do_bench_using_profiling(no_args)
    return time * 1e3

MAX_NEW_TOKENS = 1000
print("int4wo-128 model:", benchmark_fn(quantized_model.generate, **input_ids, max_new_tokens=MAX_NEW_TOKENS, cache_implementation="static"))

bf16_model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", torch_dtype=torch.bfloat16)
output = bf16_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static") # auto-compile
print("bf16 model:", benchmark_fn(bf16_model.generate, **input_ids, max_new_tokens=MAX_NEW_TOKENS, cache_implementation="static"))

For best performance, you can use recommended settings by calling torchao.quantization.utils.recommended_inductor_config_setter()

Serialization

torchao implements torch.Tensor subclasses for maximum flexibility in supporting new quantized torch.Tensor formats. Safetensors serialization and deserialization does not work with torchao.

To avoid arbitrary user code execution, torchao sets weights_only=True in torch.load to ensure only tensors are loaded. Any known user functions can be whitelisted with add_safe_globals.

# don't serialize model with Safetensors
output_dir = "llama3-8b-int4wo-128"
quantized_model.save_pretrained("llama3-8b-int4wo-128", safe_serialization=False)

Resources

For a better sense of expected performance, view the benchmarks for various models with CUDA and XPU backends.

Refer to Other Available Quantization Techniques for more examples and documentation.

< > Update on GitHub