Accelerate, Three Powerful Sublibraries for PyTorch

Zachary Mueller

Who am I?

  • Zachary Mueller
  • Deep Learning Software Engineer at 🤗
  • API design geek

What is 🤗 Accelerate?

graph LR
    A{"🤗 Accelerate#32;"}
    A --> B["Launching<br>Interface#32;"]
    A --> C["Training Library#32;"]
    A --> D["Big Model<br>Inference#32;"]

A Launching Interface

Can’t I just use python do_the_thing.py?

A Launching Interface

Launching scripts in different environments is complicated:

  • python script.py
  • torchrun --nnodes=1 --nproc_per_node=2 script.py
  • deepspeed --num_gpus=2 script.py

And more!

A Launching Interface

But it doesn’t have to be:

accelerate launch script.py

A single command to launch with DeepSpeed, Fully Sharded Data Parallelism, across single and multi CPUs and GPUs, and to train on TPUs1 too!

A Launching Interface

Generate a device-specific configuration through accelerate config

A Launching Interface

Or don’t. accelerate config doesn’t have to be done!

torchrun --nnodes=1 --nproc_per_node=2 script.py
accelerate launch --multi_gpu --nproc_per_node=2 script.py

A quick default configuration can be made too:

accelerate config default

A Launching Interface

With the notebook_launcher it’s also possible to launch code directly from your Jupyter environment too!

from accelerate import notebook_launcher
notebook_launcher(
    training_loop_function, 
    args, 
    num_processes=2
)
Launching training on 2 GPUs.
epoch 0: 88.12
epoch 1: 91.73
epoch 2: 92.58
epoch 3: 93.90
epoch 4: 94.71

A Training Library

Okay, will accelerate launch make do_the_thing.py use all my GPUs magically?

A Training Library

  • Just showed that its possible using accelerate launch to launch a python script in various distributed environments
  • This does not mean that the script will just “use” that code and still run on the new compute efficiently.
  • Training on different computes often means many lines of code changed for each specific compute.
  • 🤗 accelerate solves this by ensuring the same code can be ran on a CPU or GPU, multiples, and on TPUs!

A Training Library

for batch in dataloader:
    optimizer.zero_grad()
    inputs, targets = batch
    inputs = inputs.to(device)
    targets = targets.to(device)
    outputs = model(inputs)
    loss = loss_function(outputs, targets)
    loss.backward()
    optimizer.step()
    scheduler.step()

A Training Library




# For alignment purposes
for batch in dataloader:
    optimizer.zero_grad()
    inputs, targets = batch
    inputs = inputs.to(device)
    targets = targets.to(device)
    outputs = model(inputs)
    loss = loss_function(outputs, targets)
    loss.backward()
    optimizer.step()
    scheduler.step()
from accelerate import Accelerator
accelerator = Accelerator()
dataloader, model, optimizer scheduler = (
    accelerator.prepare(
        dataloader, model, optimizer, scheduler
    )
)

for batch in dataloader:
    optimizer.zero_grad()
    inputs, targets = batch
    # inputs = inputs.to(device)
    # targets = targets.to(device)
    outputs = model(inputs)
    loss = loss_function(outputs, targets)
    accelerator.backward(loss) # loss.backward()
    optimizer.step()
    scheduler.step()

A Training Library

What all happened in Accelerator.prepare?

  1. Accelerator looked at the configuration
  2. The dataloader was converted into one that can dispatch each batch onto a seperate GPU
  3. The model was wrapped with the appropriate DDP wrapper from either torch.distributed or torch_xla
  4. The optimizer and scheduler were both converted into an AcceleratedOptimizer and AcceleratedScheduler which knows how to handle any distributed scenario

A Training Library, Mixed Precision

🤗 accelerate also supports automatic mixed precision.

Through a single flag to the Accelerator object when calling accelerator.backward() the mixed precision of your choosing (such as bf16 or fp16) will be applied:

from accelerate import Accelerator
accelerator = Accelerator(mixed_precision="fp16")
...
for batch in dataloader:
    optimizer.zero_grad()
    inputs, targets = batch
    outputs = model(inputs)
    loss = loss_function(outputs, targets)
    accelerator.backward(loss)
    optimizer.step()
    scheduler.step()

A Training Library, Gradient Accumulation

Gradient accumulation in distributed setups often need extra care to ensure gradients are aligned when they need to be and the backward pass is computationally efficient.

🤗 accelerate can just easily handle this for you:

from accelerate import Accelerator
accelerator = Accelerator(gradient_accumulation_steps=4)
...
for batch in dataloader:
    with accelerator.accumulate(model):
        optimizer.zero_grad()
        inputs, targets = batch
        outputs = model(inputs)
        loss = loss_function(outputs, targets)
        accelerator.backward(loss)
        optimizer.step()
        scheduler.step()

A Training Library, Gradient Accumulation

ddp_model, dataloader = accelerator.prepare(model, dataloader)

for index, batch in enumerate(dataloader):
    inputs, targets = batch
    if index != (len(dataloader)-1) or (index % 4) != 0:
        # Gradients don't sync
        with accelerator.no_sync(model):
            outputs = ddp_model(inputs)
            loss = loss_func(outputs, targets)
            accelerator.backward(loss)
    else:
        # Gradients finally sync
        outputs = ddp_model(inputs)
        loss = loss_func(outputs)
        accelerator.backward(loss)

Big Model Inference

Stable Diffusion taking the world by storm

Bigger Models == Higher Compute

As more large models were being released, Hugging Face quickly realized there must be a way to continue our decentralization of Machine Learning and have the day-to-day programmer be able to leverage these big models.

Born out of this effort by Sylvain Gugger:

🤗 Accelerate: Big Model Inference.

The Basic Premise

  • In PyTorch, there exists the meta device.

  • Super small footprint to load in huge models quickly by not loading in their weights immediatly.

  • As an input gets passed through each layer, we can load and unload parts of the PyTorch model quickly so that only a small portion of the big model is loaded in at a single time.

  • The end result? Stable Diffusion v1 can be ran on < 800mb of vRAM

The Code

Generally you start with something like so:

import torch

my_model = ModelClass(...)
state_dict = torch.load(checkpoint_file)
my_model.load_state_dict(state_dict)

But this has issues:

  1. The full version of the model is loaded at 3
  2. Another version of the model is loaded into memory at 4

If a 6 billion parameter model is being loaded, each model class has a dictionary of 24GB so 48GB of vRAM is needed

Empty Model Weights

We can fix step 1 by loading in an empty model skeleton at first:

from accelerate import init_empty_weights

with init_empty_weights():
    my_model = ModelClass(...)
state_dict = torch.load(checkpoint_file)
my_model.load_state_dict(state_dict)

This code will not run

It is likely that just calling my_model(x) will fail as not all tensor operations are supported on the meta device.

Sharded Checkpoints - The Concept

The next step is to have “Sharded Checkpoints” saved for your model.

Basically smaller chunks of your model weights stored that can be brought in at any particular time.

This reduces the amount of memory step 2 takes in since we can just load in a “chunk” of the model at a time, then swap it out for a new chunk through PyTorch hooks

Sharded Checkpoints - The Code

from accelerate import init_empty_weights, load_checkpoint_and_dispatch

with init_empty_weights():
    my_model = ModelClass(...)

my_model = load_checkpoint_and_dispatch(
    my_model, "sharded-weights", device_map="auto"
)

device_map="auto" will tell 🤗 Accelerate that it should determine where to put each layer of the model:

  1. Maximum space on the GPU(s)
  2. Maximum space on the CPU(s)
  3. Utilize disk space through memory-mapped tensors

Big Model Inference Put Together

from accelerate import init_empty_weights, load_checkpoint_and_dispatch

with init_empty_weights():
    my_model = ModelClass(...)

my_model = load_checkpoint_and_dispatch(
    my_model, "sharded-weights", device_map="auto"
)
my_model.eval()

for batch in dataloader:
    output = my_model(batch)

Is there an easier way?

The transformers library combined with the Hub makes all this code wrapping much easier for you with the pipeline

import torch
from transformers import pipeline
pipe = pipeline(
    task="text-generation",
    model="EleutherAI/gpt-j-6B",
    device_map="auto",
    torch_dtype=torch.float16
)

text = pipe("This is some generated text, I think")

What about Stable Diffusion?

A demo with diffusers

Some Handy Resources