Accelerate Comandos

Accelerate é uma biblioteca HuggingFace que permite que código PyTorch execute em qualquer configuração distribuída (multi-GPU, TPU, precisão mista) com alterações mínimas de código. Gerencia automaticamente posicionamento de dispositivos, sincronização de gradientes e precisão mista.

Instalação

# Install from PyPI
pip install accelerate

# Install with DeepSpeed support
pip install accelerate[deepspeed]

# Install with testing tools
pip install accelerate[testing]

# Verify installation
accelerate env

Configuração

# Interactive configuration wizard
accelerate config

# Generate default config
accelerate config default

# Config for multi-GPU
accelerate config --config_file multi_gpu.yaml

# View current config
accelerate env

# Estimate memory requirements
accelerate estimate-memory meta-llama/Llama-3.1-8B --dtypes float16 int8

Config File Exemplos

# multi_gpu.yaml - Multi-GPU on single machine
compute_environment: LOCAL_MACHINE
distributed_type: MULTI_GPU
num_machines: 1
num_processes: 4
mixed_precision: bf16
use_cpu: false

# deepspeed.yaml - DeepSpeed ZeRO-3
compute_environment: LOCAL_MACHINE
distributed_type: DEEPSPEED
deepspeed_config:
  zero_stage: 3
  offload_optimizer_device: cpu
  offload_param_device: cpu
  zero3_save_16bit_model: true
num_machines: 1
num_processes: 4
mixed_precision: bf16

# fsdp.yaml - Fully Sharded Data Parallel
compute_environment: LOCAL_MACHINE
distributed_type: FSDP
fsdp_config:
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_backward_prefetch_policy: BACKWARD_PRE
  fsdp_sharding_strategy: FULL_SHARD
  fsdp_state_dict_type: SHARDED_STATE_DICT
num_machines: 1
num_processes: 4
mixed_precision: bf16

Launch Comandos

# Launch with config file
accelerate launch --config_file multi_gpu.yaml train.py

# Launch with inline args (no config file)
accelerate launch --multi_gpu --num_processes=4 --mixed_precision=bf16 train.py

# Single GPU
accelerate launch --num_processes=1 train.py

# Multi-node
accelerate launch \
  --num_machines=2 \
  --num_processes=8 \
  --machine_rank=0 \
  --main_process_ip=10.0.0.1 \
  --main_process_port=29500 \
  train.py

# With DeepSpeed
accelerate launch --use_deepspeed \
  --deepspeed_config_file ds_config.json \
  --zero_stage 3 \
  train.py

# TPU training
accelerate launch --tpu --num_processes=8 train.py

Accelerator Class

from accelerate import Accelerator

accelerator = Accelerator(
    mixed_precision="bf16",
    gradient_accumulation_steps=4,
    log_with="wandb",
)

# Prepare model, optimizer, dataloader, scheduler
model, optimizer, train_dataloader, scheduler = accelerator.prepare(
    model, optimizer, train_dataloader, scheduler
)

# Training loop
for epoch in range(num_epochs):
    model.train()
    for batch in train_dataloader:
        with accelerator.accumulate(model):
            outputs = model(**batch)
            loss = outputs.loss
            accelerator.backward(loss)
            optimizer.step()
            scheduler.step()
            optimizer.zero_grad()

        # Log metrics (only on main process)
        if accelerator.is_main_process:
            accelerator.log({"loss": loss.item()})

# Save model (handles distributed state)
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
accelerator.save_model(unwrapped_model, "./output")

Gradient Accumulation

accelerator = Accelerator(gradient_accumulation_steps=8)

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

for batch in dataloader:
    # accumulate() handles gradient sync and scaling
    with accelerator.accumulate(model):
        loss = model(**batch).loss
        accelerator.backward(loss)
        optimizer.step()
        optimizer.zero_grad()

Mixed Precision Training

# Via Accelerator
accelerator = Accelerator(mixed_precision="bf16")

# Custom autocast blocks
with accelerator.autocast():
    output = model(input_data)

Checkpointing

# Save full training state (model + optimizer + scheduler + RNG)
accelerator.save_state("./checkpoint-1000")

# Load training state to resume
accelerator.load_state("./checkpoint-1000")

# Save just the model weights
accelerator.wait_for_everyone()
unwrapped = accelerator.unwrap_model(model)
accelerator.save_model(unwrapped, "./model-final")

# Push to HuggingFace Hub
unwrapped.push_to_hub("my-org/my-model")

Notebook Launcher

from accelerate import notebook_launcher

def training_function():
    accelerator = Accelerator()
    model = MyModel()
    optimizer = torch.optim.AdamW(model.parameters(), lr=2e-5)
    model, optimizer = accelerator.prepare(model, optimizer)
    # ... training loop ...

# Launch multi-GPU training from a Jupyter notebook
notebook_launcher(training_function, num_processes=4)

Distributed Utilities

from accelerate import Accelerator

accelerator = Accelerator()

# Check process info
print(accelerator.process_index)     # Current process rank
print(accelerator.num_processes)     # Total number of processes
print(accelerator.is_main_process)   # True on rank 0
print(accelerator.device)            # Device for this process

# Gather tensors from all processes
gathered = accelerator.gather(local_tensor)

# Run only on main process
if accelerator.is_main_process:
    print("Training complete!")

# Synchronize all processes
accelerator.wait_for_everyone()

# Print only on main process
accelerator.print("This prints once across all processes")

Big Model Loading

from accelerate import init_empty_weights, load_checkpoint_and_dispatch

# Load a model that doesn't fit in single GPU memory
with init_empty_weights():
    model = AutoModelForCausalLM.from_config(config)

# Dispatch across GPUs automatically
model = load_checkpoint_and_dispatch(
    model,
    checkpoint="./model-dir",
    device_map="auto",
    no_split_module_classes=["LlamaDecoderLayer"],
    dtype=torch.float16,
)

Common Opções

Flag	Descrição
--multi_gpu	Multi-GPU on single machine
--num_processes N	Number of GPUs/processes
--mixed_precision bf16	BF16 mixed precision
--use_deepspeed	Enable DeepSpeed
--use_fsdp	Enable FSDP
--tpu	TPU training
--num_machines N	Multi-node count
--gradient_accumulation_steps N	Gradient accumulation