⚠️ WARNING: This is a prototype for PyTorch fault tolerance and may have bugs or breaking changes as this is actively under development. This is a public repo to encourage collaboration and contributions are welcome. There currently are no plans to make this a stable component of PyTorch Distributed and may be abandonded at any time if better approachs arise.
This repository implements techniques for doing a per-step fault tolerance so you can keep training if errors occur without interrupting the entire training job.
This is based off of the large scale training techniques presented at PyTorch Conference 2024.
torchft is designed to allow for fault tolerance when using training with replicated weights such as in DDP or HSDP (FSDP with DDP).
torchft implements a lighthouse server that coordinates across the different replica groups and then a per replica group manager and fault tolerance library that can be used in a standard PyTorch training loop.
This allows for membership changes at the training step granularity which can greatly improve efficiency by avoiding stop the world training on errors.
Before proceeding, ensure you have Rust installed on your system. Note that the Rust versions available in many conda environments may be outdated. To install the latest version of Rust, we recommend downloading it directly from the official website as shown in the below command:
$ curl --proto '=https' --tlsv1.2 https://sh.rustup.rs -sSf | sh$ pip install .This uses pyo3+maturin to build the package, you'll need maturin installed.
If the installation command fails to invoke cargo update due to an inability to fetch the manifest, it may be caused by the proxy, proxySSLCert, and proxySSLKey settings in your .gitconfig file affecting the cargo command. To resolve this issue, try temporarily removing these fields from your .gitconfig before running the installation command.
To install in editable mode w/ the Rust extensions you can use the normal pip install command:
$ pip install -e .The lighthouse is used for fault tolerance across replicated workers (DDP/FSDP) when using synchronous training.
You can start a lighthouse server by running:
$ RUST_BACKTRACE=1 torchft_lighthouse --min_replicas 1 --quorum_tick_ms 100 --join_timeout_ms 1000See train_ddp.py for the full example.
Invoke with:
$ TORCHFT_MANAGER_PORT=29512 TORCHFT_LIGHTHOUSE=http://localhost:29510 torchrun --master_port 29501 --nnodes 1 --nproc_per_node 1 train.pytrain.py:
from torchft import Manager, DistributedDataParallel, Optimizer, ProcessGroupGloo
manager = Manager(
pg=ProcessGroupGloo(),
load_state_dict=...,
state_dict=...,
)
m = nn.Linear(2, 3)
m = DistributedDataParallel(manager, m)
optimizer = Optimizer(manager, optim.AdamW(m.parameters()))
for i in range(1000):
batch = torch.rand(2, 2, device=device)
optimizer.zero_grad()
out = m(batch)
loss = out.sum()
loss.backward()
optimizer.step()torchft has a fault tolerant parameter server implementation built on it's reconfigurable ProcessGroups. This does not require/use a Lighthouse server.
See parameter_server_test.py for an example.
We welcome PRs! See the CONTRIBUTING file.
torchft is BSD 3-Clause licensed. See LICENSE for more details.