Open source framework for simulated distributed training methods. Instead of training with multiple ranks, we simulate the distributed training process by running multiple nodes on a single machine.
- CPU
- CUDA
- MPS (CPU-bound for copy operations, see here)
Install with core dependencies only:
pip install exogymOptional feature flags allowed are:
wandb,gpt,demo,examples,all,devFor example, pip install exogym[demo]
To install for development:
git clone https://github.com/exo-explore/gym.git exogym
cd exogym
pip install -e ".[dev]"MNIST comparison of DDP, DiLoCo, and SPARTA:
python run/mnist.pyNanoGPT Shakespeare DiLoCo:
python run/nanogpt_diloco.py --dataset shakespearefrom exogym import LocalTrainer
from exogym.strategy import DiLoCoStrategy
train_dataset, val_dataset = ...
model = ... # model.forward() expects a batch, and returns a scalar loss
trainer = LocalTrainer(model, train_dataset, val_dataset)
# Strategy for optimization & communication
strategy = DiLoCoStrategy(
inner_optim='adam',
H=100
)
trainer.fit(
strategy=strategy,
num_nodes=4,
device='mps'
)Trainer: Builds simulation environment.Trainerwill spawn multipleTrainNodeinstances, connect them together, and starts the training run.TrainNode: A single node (rank) running its own training loop. At each train step, instead of callingoptim.step(), it callsstrategy.step().Strategy: Abstract class for an optimization strategy, which both defines how the nodes communicate with each other and how model weights are updated. Typically, a gradient strategy will include an optimizer as well as a communication step. Sometimes (eg. DeMo), the optimizer step is comingled with the communication.
EXO Gym uses pytorch multiprocessing to a subprocess per-node, which are able to communicate with each other using regular operations such as all_reduce.
The model is expected in a form that takes a batch (the same format as dataset outputs), and returns a scalar loss over the entire batch. This ensures the model is agnostic to the format of the data (eg. masked LM training doesn't have a clear x/y split).
Recall that when we call trainer.fit(), dataset object is passed to every subprocess, and a DistributedSampler will be used to select indices per-node. If the dataset is entirely loaded into memory, this memory will be duplicated per-node - be careful not to run out of memory! If the dataset is larger, it should be lazily loaded.