Note: This repository is currently under heavy development - if you have suggestions on the API or use-cases you'd like to be covered, please open an github issue

The torchao package allows you to quantize and prune your models using native PyTorch.

The repo hosts both

1. lower precision dtypes such as nf4, uint4
2. Quantization algorithms such as dynamic quant, smoothquant
3. Sparsity algorithms such as Wanda

Our kernels have has been used to achieve SOTA inference performance on

1. Image segmentation modelss with sam-fast
2. Language models with gpt-fast
3. Diffusion models with sd-fast

Note: this library makes liberal use of several new features in pytorch, its recommended to use it with the current pytorch nightly if you want full feature coverage. If not, the subclass APIs may not work, though the module swap api's will still work.

1. From PyPI:

pip install torchao

2. From Source:

git clone https://github.com/pytorch-labs/ao
cd ao
pip install -e .

Typically quantization algorithms will have different schemes for how the activation and weights are quantized so A16W8 for instance means the activations are quantized to 16 bits wheras the weights are quantized to 8 bits. Trying out different quantization schemes in torchao is generally a 1 line change.

import torch
from torchao.quantization import quant_api

# Fuse the int8*int8 -> int32 matmul and subsequent mul op avoiding materialization of the int32 intermediary tensor
torch._inductor.config.force_fuse_int_mm_with_mul = True

# Plug in your model and example input
model = torch.nn.Sequential(torch.nn.Linear(32, 64)).cuda().to(torch.bfloat16)
input = torch.randn(32,32, dtype=torch.bfloat16, device='cuda')

# convert linear modules to quantized linear modules
quant_api.change_linear_weights_to_int8_dqtensors(model)

# compile the model to improve performance
model = torch.compile(model, mode='max-autotune')
model(input)

quant_api.change_linear_weights_to_int8_woqtensors(model)

This technique works best when the torch._inductor.config.use_mixed_mm option is enabled. This avoids dequantizing the weight tensor before the matmul, instead fusing the dequantization into the matmul, thereby avoiding materialization of a large floating point weight tensor.

quant_api.change_linear_weights_to_int4_woqtensors(model)

Note: The quantization error incurred by applying int4 quantization to your model can be fairly significant, so using external techniques like GPTQ may be necessary to obtain a usable model.

We've also implemented a version of smoothquant with the same GEMM format as above. Due to requiring calibration, the API is more complicated.

Example

import torch
from torchao.quantization.smoothquant import swap_linear_with_smooth_fq_linear, smooth_fq_linear_to_inference

# Fuse the int8*int8 -> int32 matmul and subsequent mul op avoiding materialization of the int32 intermediary tensor
torch._inductor.config.force_fuse_int_mm_with_mul = True

# plug in your model
model = get_model()

# convert linear modules to smoothquant
# linear module in calibration mode
swap_linear_with_smooth_fq_linear(model)

# Create a data loader for calibration
calibration_data = get_calibration_data()
calibration_dataset = MyDataset(calibration_data)
calibration_loader = DataLoader(calibration_dataset, batch_size=32, shuffle=True)

# Calibrate the model
model.train()
for batch in calibration_loader:
    inputs = batch
    model(inputs)

# set it to inference mode
smooth_fq_linear_to_inference(model)

# compile the model to improve performance
model = torch.compile(model, mode='max-autotune')
model(input)

1. While these techniques are designed to improve model performance, in some cases the opposite can occur. This is because quantization adds additional overhead to the model that is hopefully made up for by faster matmuls (dynamic quantization) or loading weights faster (weight-only quantization). If your matmuls are small enough or your non-quantized perf isn't bottlenecked by weight load time, these techniques may reduce performance.
2. Use the PyTorch nightlies so you can leverage tensor subclasses which is preferred over older module swap based methods because it doesn't modify the graph and is generally more composable and flexible.

torchao is released under the BSD 3 license.