Démarrage rapide 🤗 Transformateurs

In this federated learning tutorial we will learn how to train a large language model (LLM) on the IMDB dataset using Flower and the 🤗 Hugging Face Transformers library. It is recommended to create a virtual environment and run everything within a virtualenv.

Let’s use flwr new to create a complete Flower+🤗 Hugging Face project. It will generate all the files needed to run, by default with the Flower Simulation Engine, a federation of 10 nodes using FedAvg The dataset will be partitioned using Flower Datasets’s IidPartitioner.

Now that we have a rough idea of what this example is about, let’s get started. First, install Flower in your new environment:

# In a new Python environment
$ pip install flwr

Then, run the command below. You will be prompted to select one of the available templates (choose HuggingFace), give a name to your project, and type in your developer name:

$ flwr new

After running it you’ll notice a new directory with your project name has been created. It should have the following structure:

<your-project-name>
├── <your-project-name>
│   ├── __init__.py
│   ├── client_app.py   # Defines your ClientApp
│   ├── server_app.py   # Defines your ServerApp
│   └── task.py         # Defines your model, training and data loading
├── pyproject.toml      # Project metadata like dependencies and configs
└── README.md

If you haven’t yet installed the project and its dependencies, you can do so by:

# From the directory where your pyproject.toml is
$ pip install -e .

To run the project, do:

# Run with default arguments
$ flwr run .

With default arguments you will see an output like this one:

Loading project configuration...
Success
INFO :      Starting FedAvg strategy:
INFO :          ├── Number of rounds: 3
INFO :          ├── ArrayRecord (16.74 MB)
INFO :          ├── ConfigRecord (train): (empty!)
INFO :          ├── ConfigRecord (evaluate): (empty!)
INFO :          ├──> Sampling:
INFO :          │       ├──Fraction: train (0.50) | evaluate ( 1.00)
INFO :          │       ├──Minimum nodes: train (2) | evaluate (2)
INFO :          │       └──Minimum available nodes: 2
INFO :          └──> Keys in records:
INFO :                  ├── Weighted by: 'num-examples'
INFO :                  ├── ArrayRecord key: 'arrays'
INFO :                  └── ConfigRecord key: 'config'
INFO :
INFO :
INFO :      [ROUND 1/3]
INFO :      configure_train: Sampled 5 nodes (out of 10)
INFO :      aggregate_train: Received 5 results and 0 failures
INFO :          └──> Aggregated MetricRecord: {'train_loss': 0.6974}
INFO :      configure_evaluate: Sampled 10 nodes (out of 10)
INFO :      aggregate_evaluate: Received 10 results and 0 failures
INFO :          └──> Aggregated MetricRecord: {'val_loss': 0.0223, 'val_accuracy': 0.5024}
INFO :
INFO :      [ROUND 2/3]
INFO :      configure_train: Sampled 5 nodes (out of 10)
INFO :      aggregate_train: Received 5 results and 0 failures
INFO :          └──> Aggregated MetricRecord: {'train_loss': 0.7019}
INFO :      configure_evaluate: Sampled 10 nodes (out of 10)
INFO :      aggregate_evaluate: Received 10 results and 0 failures
INFO :          └──> Aggregated MetricRecord: {'val_loss': 0.0221, 'val_accuracy': 0.5176}
INFO :
INFO :      [ROUND 3/3]
INFO :      configure_train: Sampled 5 nodes (out of 10)
INFO :      aggregate_train: Received 5 results and 0 failures
INFO :          └──> Aggregated MetricRecord: {'train_loss': 0.6845}
INFO :      configure_evaluate: Sampled 10 nodes (out of 10)
INFO :      aggregate_evaluate: Received 10 results and 0 failures
INFO :          └──> Aggregated MetricRecord: {'val_loss': 0.0221, 'val_accuracy': 0.5042}
INFO :
INFO :      Strategy execution finished in 151.02s
INFO :
INFO :      Final results:
INFO :
INFO :          Global Arrays:
INFO :                  ArrayRecord (16.737 MB)
INFO :
INFO :          Aggregated ClientApp-side Train Metrics:
INFO :          { 1: {'train_loss': '6.9738e-01'},
INFO :            2: {'train_loss': '7.0191e-01'},
INFO :            3: {'train_loss': '6.8449e-01'}}
INFO :
INFO :          Aggregated ClientApp-side Evaluate Metrics:
INFO :          { 1: {'val_accuracy': '5.0240e-01', 'val_loss': '2.2265e-02'},
INFO :            2: {'val_accuracy': '5.1760e-01', 'val_loss': '2.2134e-02'},
INFO :            3: {'val_accuracy': '5.0420e-01', 'val_loss': '2.2124e-02'}}
INFO :
INFO :          ServerApp-side Evaluate Metrics:
INFO :          {}
INFO :

Saving final model to disk...

You can also run the project with GPU as follows:

# Run with default arguments
$ flwr run . localhost-gpu

This will use the default arguments where each ClientApp will use 4 CPUs and at most 4 ClientApps will run in a given GPU.

You can also override the parameters defined in the [tool.flwr.app.config] section in pyproject.toml like this:

# Override some arguments
$ flwr run . --run-config "num-server-rounds=5 fraction-train=0.2"

What follows is an explanation of each component in the project you just created: dataset partition, the model, defining the ClientApp and defining the ServerApp.

The Data

This tutorial uses Flower Datasets to easily download and partition the IMDB dataset. In this example you’ll make use of the IidPartitioner to generate num_partitions partitions. You can choose other partitioners available in Flower Datasets. To tokenize the text, we will also load the tokenizer from the pre-trained Transformer model that we’ll use during training - more on that in the next section. Each ClientApp will call this function to create dataloaders with the data that correspond to their data partition.

partitioner = IidPartitioner(num_partitions=num_partitions)
fds = FederatedDataset(
    dataset="stanfordnlp/imdb",
    partitioners={"train": partitioner},
)
partition = fds.load_partition(partition_id)
# Divide data: 80% train, 20% test
partition_train_test = partition.train_test_split(test_size=0.2, seed=42)

tokenizer = AutoTokenizer.from_pretrained(model_name)


def tokenize_function(examples):
    return tokenizer(
        examples["text"], truncation=True, add_special_tokens=True, max_length=512
    )


partition_train_test = partition_train_test.map(tokenize_function, batched=True)
partition_train_test = partition_train_test.remove_columns("text")
partition_train_test = partition_train_test.rename_column("label", "labels")

data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
trainloader = DataLoader(
    partition_train_test["train"],
    shuffle=True,
    batch_size=32,
    collate_fn=data_collator,
)

testloader = DataLoader(
    partition_train_test["test"], batch_size=32, collate_fn=data_collator
)

The Model

We will leverage 🤗 Hugging Face to federate the training of language models over multiple clients using Flower. More specifically, we will fine-tune a pre-trained Transformer model (bert-tiny) for sequence classification over the dataset of IMDB ratings. The end goal is to detect if a movie rating is positive or negative. If you have access to larger GPUs, feel free to use larger models!

net = AutoModelForSequenceClassification.from_pretrained(
    model_name, num_labels=num_labels
)

Note that here, model_name is a string that will be loaded from the Context in the ClientApp and ServerApp.

In addition to loading the pretrained model weights and architecture, we also include two utility functions to perform both training (i.e. train()) and evaluation (i.e. test()) using the above model. These functions should look fairly familiar if you have some prior experience with PyTorch. Note these functions do not have anything specific to Flower. That being said, the training function will normally be called, as we’ll see later, from a Flower client passing its own data. In summary, your clients can use standard training/testing functions to perform local training or evaluation:

def train(net, trainloader, epochs, device):
    optimizer = AdamW(net.parameters(), lr=5e-5)
    net.train()
    for _ in range(epochs):
        for batch in trainloader:
            batch = {k: v.to(device) for k, v in batch.items()}
            outputs = net(**batch)
            loss = outputs.loss
            loss.backward()
            optimizer.step()
            optimizer.zero_grad()


def test(net, testloader, device):
    metric = load_metric("accuracy")
    loss = 0
    net.eval()
    for batch in testloader:
        batch = {k: v.to(device) for k, v in batch.items()}
        with torch.no_grad():
            outputs = net(**batch)
        logits = outputs.logits
        loss += outputs.loss.item()
        predictions = torch.argmax(logits, dim=-1)
        metric.add_batch(predictions=predictions, references=batch["labels"])
    loss /= len(testloader.dataset)
    accuracy = metric.compute()["accuracy"]
    return loss, accuracy

The ClientApp

The main changes we have to make to use 🤗 Hugging Face with Flower have to do with converting the ArrayRecord received in the Message into a PyTorch state_dict and vice versa when generating the reply Message from the ClientApp. We can make use of the built-in methods in the ArrayRecord to make these conversions:

# Load the model
net = AutoModelForSequenceClassification.from_pretrained(
    model_name, num_labels=num_labels
)

# Extract ArrayRecord from Message and convert to PyTorch state_dict,
state_dict = msg.content["arrays"].to_torch_state_dict()
# Load state_dict into the model
net.load_state_dict(state_dict)

# ... do some training

# Convert state_dict back into an ArrayRecord
model_record = ArrayRecord(net.state_dict())

The rest of the functionality is directly inspired by the centralized case. The ClientApp comes with three core methods (train, evaluate, and query) that we can implement for different purposes. For example: train to train the received model using the local data; evaluate to assess its performance of the received model on a validation set; and query to retrieve information about the node executing the ClientApp. In this tutorial we will only make use of train and evaluate.

Let’s see how the train method can be implemented. It receives as input arguments a Message from the ServerApp. By default it carries:

• an ArrayRecord with the arrays of the model to federate. By default they can be retrieved with key "arrays" when accessing the message content.

• a ConfigRecord with the configuration sent from the ServerApp. By default it can be retrieved with key "config" when accessing the message content.

The train method also receives the Context, giving access to configs for your run and node. The run config hyperparameters are defined in the pyproject.toml of your Flower App. The node config can only be set when running Flower with the Deployment Runtime and is not directly configurable during simulations.

# Flower ClientApp
app = ClientApp()


@app.train()
def train(msg: Message, context: Context):
    """Train the model on local data."""

    # Get this client's dataset partition
    partition_id = context.node_config["partition-id"]
    num_partitions = context.node_config["num-partitions"]
    model_name = context.run_config["model-name"]
    trainloader, _ = load_data(partition_id, num_partitions, model_name)

    # Load model
    num_labels = context.run_config["num-labels"]
    net = AutoModelForSequenceClassification.from_pretrained(
        model_name, num_labels=num_labels
    )

    # Initialize it with the received weights
    net.load_state_dict(msg.content["arrays"].to_torch_state_dict())
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    net.to(device)

    # Train the model on local data
    train_loss = train_fn(
        net,
        trainloader,
        context.run_config["local-steps"],
        device,
    )

    # Construct and return reply Message
    model_record = ArrayRecord(net.state_dict())
    metrics = {
        "train_loss": train_loss,
        "num-examples": len(trainloader.dataset),
    }
    metric_record = MetricRecord(metrics)
    content = RecordDict({"arrays": model_record, "metrics": metric_record})
    return Message(content=content, reply_to=msg)

The @app.evaluate() method would be near identical with two exceptions: (1) the model is not locally trained, instead it is used to evaluate its performance on the locally held-out validation set; (2) including the model in the reply Message is no longer needed because it is not locally modified.

The ServerApp

To construct a ServerApp we define its @app.main() method. This method receive as input arguments:

• a Grid object that will be used to interface with the nodes running the ClientApp to involve them in a round of train/evaluate/query or other.

• a Context object that provides access to the run configuration.

In this example we use the FedAvg and configure it with a specific value of fraction_train which is read from the run config. You can find the default value defined in the pyproject.toml. Then, the execution of the strategy is launched when invoking its start method. To it we pass:

• the Grid object.

• an ArrayRecord carrying a randomly initialized model that will serve as the global model to be federated.

• a ConfigRecord with the training hyperparameters to be sent to the clients. The strategy will also insert the current round number in this config before sending it to the participating nodes.

• the num_rounds parameter specifying how many rounds of FedAvg to perform.

# Create ServerApp
app = ServerApp()


@app.main()
def main(grid: Grid, context: Context) -> None:
    """Main entry point for the ServerApp."""

    # Read from config
    num_rounds = context.run_config["num-server-rounds"]
    fraction_train = context.run_config["fraction-train"]

    # Initialize global model
    model_name = context.run_config["model-name"]
    num_labels = context.run_config["num-labels"]
    net = AutoModelForSequenceClassification.from_pretrained(
        model_name, num_labels=num_labels
    )
    arrays = ArrayRecord(net.state_dict())

    # Initialize FedAvg strategy
    strategy = FedAvg(fraction_train=fraction_train)

    # Start strategy, run FedAvg for `num_rounds`
    result = strategy.start(
        grid=grid,
        initial_arrays=arrays,
        num_rounds=num_rounds,
    )

    # Save final model to disk
    print("\nSaving final model to disk...")
    state_dict = result.arrays.to_torch_state_dict()
    torch.save(state_dict, "final_model.pt")

Note the start method of the strategy returns a result object. This object contains all the relevant information about the FL process, including the final model weights as an ArrayRecord, and federated training and evaluation metrics as MetricRecords. You can easily log the metrics using Python’s pprint and save the global model state_dict using torch.save.

Congratulations! You’ve successfully built and run your first federated learning system for an LLM.

Note

Check the source code of the extended version of this tutorial in examples/quickstart-huggingface in the Flower GitHub repository. For a comprehensive example of a federated fine-tuning of an LLM with Flower, refer to the FlowerTune LLM example in the Flower GitHub repository.