Skip to content
forked from pytorch/ao

The torchao repository contains api's and workflows for quantization and pruning gpu models.

License

Notifications You must be signed in to change notification settings

yiliu30/torchao-fork

 
 

Repository files navigation

torchao: PyTorch Architecture Optimization

Introduction | Inference | Training | Composability | Custom Kernels | Alpha Features | Installation | Integrations | Videos | License

Introduction

torchao: PyTorch library for custom data types & optimizations. Quantize and sparsify weights, gradients, optimizers & activations for inference and training.

From the team that brought you the fast series

  • 9.5x speedups for Image segmentation models with sam-fast
  • 10x speedups for Language models with gpt-fast
  • 3x speedup for Diffusion models with sd-fast

torchao just works with torch.compile() and FSDP2 over most PyTorch models on Huggingface out of the box.

Inference

Post Training Quantization

Quantizing and Sparsifying your models is a 1 liner that should work on any model with an nn.Linear including your favorite HuggingFace model. You can find a more comprehensive usage instructions here, sparsity here and a HuggingFace inference example here

For inference, we have the option of

  1. Quantize only the weights: works best for memory bound models
  2. Quantize the weights and activations: works best for compute bound models
  3. Quantize the activations and weights and sparsify the weight
from torchao.quantization.quant_api import (
    quantize_,
    int8_dynamic_activation_int8_weight,
    int4_weight_only,
    int8_weight_only
)
quantize_(m, int4_weight_only())

For gpt-fast int4_weight_only() is the best option at bs=1 as it 2x the tok/s and reduces the VRAM requirements by about 65% over a torch.compiled baseline.

If you don't have enough VRAM to quantize your entire model on GPU and you find CPU quantization to be too slow then you can use the device argument like so quantize_(model, int8_weight_only(), device="cuda") which will send and quantize each layer individually to your GPU.

If you see slowdowns with any of these techniques or you're unsure which option to use, consider using autoquant which will automatically profile layers and pick the best way to quantize each layer.

model = torchao.autoquant(torch.compile(model, mode='max-autotune'))

We also provide a developer facing API so you can implement your own quantization algorithms so please use the excellent HQQ algorithm as a motivating example.

KV Cache Quantization

We've added kv cache quantization and other features in order to enable long context length (and necessarily memory efficient) inference.

In practice these features alongside int4 weight only quantization allow us to reduce peak memory by ~55%, meaning we can Llama3.1-8B inference with a 130k context length with only 18.9 GB of peak memory. More details can be found here

Quantization Aware Training

Post-training quantization can result in a fast and compact model, but may also lead to accuracy degradation. We recommend exploring Quantization Aware Training (QAT) to overcome this limitation. In collaboration with Torchtune, we've developed a QAT recipe that demonstrates significant accuracy improvements over traditional PTQ, recovering 96% of the accuracy degradation on hellaswag and 68% of the perplexity degradation on wikitext for Llama3 compared to post-training quantization (PTQ). And we've provided a full recipe here

from torchao.quantization.prototype.qat import Int8DynActInt4WeightQATQuantizer

qat_quantizer = Int8DynActInt4WeightQATQuantizer()

# Insert "fake quantize" operations into linear layers.
# These operations simulate quantization numerics
model = qat_quantizer.prepare(model)

# Run Training...

# Convert fake quantize to actual quantize operations
model = qat_quantizer.convert(model)

Training

Float8

torchao.float8 implements training recipes with the scaled float8 dtypes, as laid out in https://arxiv.org/abs/2209.05433.

With torch.compile on, current results show throughput speedups of up to 1.5x on 128 H100 GPU LLaMa 3 70B pretraining jobs (details)

from torchao.float8 import convert_to_float8_training
convert_to_float8_training(m, module_filter_fn=...)

And for an end-to-minimal training recipe of pretraining with float8, you can check out torchtitan

Sparse Training

We've added support for semi-structured 2:4 sparsity with 6% end-to-end speedups on ViT-L. Full blog here

The code change is a 1 liner with the full example available here

swap_linear_with_semi_sparse_linear(model, {"seq.0": SemiSparseLinear})

Memory-efficient optimizers

ADAM takes 2x as much memory as the model params so we can quantize the optimizer state to either 8 or 4 bit effectively reducing the optimizer VRAM requirements by 2x or 4x respectively over an fp16 baseline

from torchao.prototype.low_bit_optim import AdamW8bit, AdamW4bit, AdamWFp8
optim = AdamW8bit(model.parameters()) # replace with Adam4bit and AdamFp8 for the 4 / fp8 versions

In practice, we are a tiny bit slower than expertly written kernels but the implementations for these optimizers were written in a few hundred lines of PyTorch code and compiled so please use them or copy-paste them for your quantized optimizers. Benchmarks here

We also have support for single GPU CPU offloading where both the gradients (same size as weights) and the optimizers will be efficiently sent to the CPU. This alone can reduce your VRAM requirements by 60%

optim = CPUOffloadOptimizer(model.parameters(), torch.optim.AdamW, fused=True)
optim.load_state_dict(ckpt["optim"])

Composability

  1. torch.compile: A key design principle for us is composability as in any new dtype or layout we provide needs to work with our compiler. It shouldn't matter if the kernels are written in pure PyTorch, CUDA, C++, or Triton - things should just work! So we write the dtype, layout, or bit packing logic in pure PyTorch and code-generate efficient kernels.
  2. FSDP2: Historically most quantization has been done for inference, there is now a thriving area of research combining distributed algorithms and quantization.

The best example we have combining the composability of lower bit dtype with compile and fsdp is NF4 which we used to implement the QLoRA algorithm. So if you're doing research at the intersection of this area we'd love to hear from you.

Custom Kernels

We've added support for authoring and releasing custom ops that do not graph break with torch.compile() so if you love writing kernels but hate packaging them so they work all operating systems and cuda versions, we'd love to accept contributions for your custom ops. We have a few examples you can follow

  1. fp6 for 2x faster inference over fp16 with an easy to use API quantize_(model, fp6_llm_weight_only())
  2. 2:4 Sparse Marlin GEMM 2x speedups for FP16xINT4 kernels even at batch sizes up to 256
  3. int4 tinygemm unpacker which makes it easier to switch quantized backends for inference

If you believe there's other CUDA kernels we should be taking a closer look at please leave a comment on this issue

Alpha features

Things we're excited about but need more time to cook in the oven

  1. MX training and inference support with tensors using the OCP MX spec data types, which can be described as groupwise scaled float8/float6/float4/int8, with the scales being constrained to powers of two. This work is prototype as the hardware support is not available yet.
  2. Int8 Quantized Training: We're trying out full int8 training. This is easy to use with quantize_(model, int8_weight_only_quantized_training()). This work is prototype as the memory benchmarks are not compelling yet.
  3. IntX: We've managed to support all the ints by doing some clever bitpacking in pure PyTorch and then compiling it. This work is prototype as unfortunately without some more investment in either the compiler or low-bit kernels, int4 is more compelling than any smaller dtype
  4. Bitnet: Mostly this is very cool to people on the team. This is prototype because how useful these kernels are is highly dependent on better hardware and kernel support.

Installation

torchao makes liberal use of several new features in Pytorch, it's recommended to use it with the current nightly or latest stable version of PyTorch.

Stable release from Pypi which will default to CUDA 12.1

pip install torchao

Stable Release from the PyTorch index

pip install torchao --extra-index-url https://download.pytorch.org/whl/cu121 # full options are cpu/cu118/cu121/cu124

Nightly Release

pip install --pre torchao --index-url https://download.pytorch.org/whl/nightly/cu121 # full options are cpu/cu118/cu121/cu124

For most developers you probably want to skip building custom C++/CUDA extensions for faster iteration

USE_CPP=0 pip install -e .

Integrations

We're also fortunate to be integrated into some of the leading open-source libraries including

  1. Hugging Face transformers with a builtin inference backend and low bit optimizers
  2. Hugging Face diffusers best practices with torch.compile and torchao standalone repo
  3. Mobius HQQ backend leveraged our int4 kernels to get 195 tok/s on a 4090

Videos

License

torchao is released under the BSD 3 license.

About

The torchao repository contains api's and workflows for quantization and pruning gpu models.

Resources

License

Code of conduct

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 85.1%
  • C++ 9.8%
  • Cuda 4.0%
  • Shell 0.7%
  • CMake 0.4%
  • Objective-C++ 0.0%