meta-pytorch/torchft
A PyTorch library providing per-step fault tolerance primitives for distributed training jobs.
Not currently ranked — collecting fresh signals.
star history
torchft implements fault tolerance techniques for large-scale PyTorch training including Fault Tolerant DDP, HSDP, LocalSGD, and DiLoCo. It provides coordination primitives, fault tolerant ProcessGroup implementations, and checkpoint transports that allow training jobs to recover from errors without restarting entirely. The library targets distributed training scenarios where node failures can disrupt long-running training runs.
Frequently asked
- What is meta-pytorch/torchft?
- A PyTorch library providing per-step fault tolerance primitives for distributed training jobs.
- Is torchft open source?
- Yes — meta-pytorch/torchft is an open-source project tracked on heatdrop.
- What language is torchft written in?
- meta-pytorch/torchft is primarily written in Python.
- How popular is torchft?
- meta-pytorch/torchft has 526 stars on GitHub.
- Where can I find torchft?
- meta-pytorch/torchft is on GitHub at https://github.com/meta-pytorch/torchft.