torchft-nightly
Decision gist · record as of 2026-08-14
Yes, with conditions. Install if you are training large distributed models on infrastructure where worker failures are common and you need per-step recovery without full job restart. The active maintenance, recent releases, and zero known vulnerabilities are positive signals. However, verify the BSD 3-Clause license applies to your use case (metadata is unclear), and be aware that medium install friction (Rust compilation, system dependencies) and nightly-only availability mean this is best suited for teams with infrastructure expertise rather than casual users.AI-flagged interpretation of the facts on this page — verify before relying
Before you install
- Requires Rust, protobuf-compiler, and PyTorch 2.7 RC+ or Nightly.
- Lighthouse server must be running for DDP/HSDP coordination (e.g., torchft_lighthouse --min_replicas 1).
- Medium install friction due to compiled Rust components and system dependencies (protobuf-compiler, Rust toolchain).
License · maintenance · safety
(unclear) — License treatment is unclear; the repository indicates BSD 3-Clause licensing, but this is not formally declared in package metadata. Verify licensing terms before use in proprietary projects.
last release 2026-08-14 (0 days) · last repo commit 2026-07-16 · 530 stars
0 known vulnerabilities (OSV.dev, 2026-08-14) · 102,918 downloads/mo, #12,836 on PyPI
Alternatives
Verify before relying
pip install torchft-nightly
from torchft import Manager, DistributedDataParallel, Optimizer, ProcessGroupGloo
manager = Manager(pg=ProcessGroupGloo(), load_state_dict=..., state_dict=...)
m = DistributedDataParallel(manager, model)
optimizer = Optimizer(manager, optim.AdamW(m.parameters()))- Whether BSD 3-Clause license is formally enforced or if there are additional licensing constraints not reflected in metadata.
- Stability guarantees for nightly builds and whether production use is recommended.
- Performance overhead of per-step fault tolerance compared to standard training without recovery.
What it is and what it does
torchft-nightly is a PyTorch library that adds fault tolerance to distributed training by detecting and recovering from worker failures at the training step level, rather than requiring full job restart. It provides coordination primitives, fault-tolerant process groups, and checkpoint transports built on a Lighthouse server that tracks worker health via heartbeating. The library includes out-of-the-box algorithms for Fault Tolerant DDP, Fault Tolerant HSDP (combining FSDP, tensor parallelism, and DDP), LocalSGD, and DiLoCo.
The package depends on torch, opentelemetry-api, opentelemetry-sdk, and opentelemetry-exporter-otlp-proto-http for observability. It is designed for large-scale training scenarios where worker failures are common and restarting the entire job is expensive. Integration requires wrapping your model, optimizer, and process group with torchft's Manager and fault-tolerant wrappers, then running a Lighthouse server to coordinate membership changes across replica groups.
Use it for
- Training large models like Llama 3 70B on multi-node clusters where worker failures are expected and full restart is costly.
- Implementing fault-tolerant DDP training loops that gracefully handle node failures without losing training progress.
- Scaling HSDP training across heterogeneous hardware with automatic recovery from individual worker outages.
- Running parameter server training with reconfigurable process groups that adapt to membership changes.
- Distributed training on preemptible or spot instances where interruptions are frequent and expected.
Worth the install?
AI-flagged interpretation of the facts on this page. Verify before relying on it.
Yes, with conditions.
Install if you are training large distributed models on infrastructure where worker failures are common and you need per-step recovery without full job restart. The active maintenance, recent releases, and zero known vulnerabilities are positive signals. However, verify the BSD 3-Clause license applies to your use case (metadata is unclear), and be aware that medium install friction (Rust compilation, system dependencies) and nightly-only availability mean this is best suited for teams with infrastructure expertise rather than casual users.
Install
torchft-nightly on PyPI
Before you install
Medium install friction due to compiled Rust components and system dependencies (protobuf-compiler, Rust toolchain). Active maintenance with recent releases; nightly builds available. Requires PyTorch 2.7 RC+ or Nightly.
Requires Rust, protobuf-compiler, and PyTorch 2.7 RC+ or Nightly. Lighthouse server must be running for DDP/HSDP coordination (e.g., torchft_lighthouse --min_replicas 1).
License in practice
License treatment is unclear; the repository indicates BSD 3-Clause licensing, but this is not formally declared in package metadata. Verify licensing terms before use in proprietary projects.
Quickstart
pip install torchft-nightly
from torchft import Manager, DistributedDataParallel, Optimizer, ProcessGroupGloo
manager = Manager(pg=ProcessGroupGloo(), load_state_dict=..., state_dict=...)
m = DistributedDataParallel(manager, model)
optimizer = Optimizer(manager, optim.AdamW(m.parameters()))Verify before relying
- Whether BSD 3-Clause license is formally enforced or if there are additional licensing constraints not reflected in metadata.
- Stability guarantees for nightly builds and whether production use is recommended.
- Performance overhead of per-step fault tolerance compared to standard training without recovery.
Package facts
| License | Not declared unclear |
| Python support | Supports the current Python release >=3.8 |
| Install friction | Medium. Platform-specific wheel |
| Runtime dependencies | 4 packagestorchopentelemetry-exporter-otlp-proto-httpopentelemetry-sdkopentelemetry-api |
| Maintenance | Actively maintained 0 days since the last release |
| Last repo commit | |
| First released | |
| Downloads | 102,918 / month, #12,836 on PyPI 30-day window, as of 2026-08-14 |
| Known vulnerabilities | None known OSV.dev, checked 2026-08-14 |
| Classifiers | Programming Language :: Python :: Implementation :: CPythonProgramming Language :: Python :: Implementation :: PyPyProgramming Language :: Rust |
Evidence: torchft_nightly-2026.8.14-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl; torchft_nightly-2026.8.14-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl; torchft_nightly-2026.8.14-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl; torchft_nightly-2026.8.14-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl; torchft_nightly-2026.8.14-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
Tags
Let your AI agent find packages like this
Example. Real query, live index.
You found this page by searching. An agent finds it by wishing: SkillFed indexes 14,416 PyPI packages by what they can do, searchable in plain language.
wish › “pytorch fault tolerance”
- torchft-nightlyProvides per-step fault tolerance for PyTorch distributed training,…
- nvidia-resiliency-extAdds fault tolerance, checkpointing, and performance monitoring to…
- backoff-utilsProvides functions and decorators to apply backoff and retry…
Give your agent the search over MCP, or paste the wish link into any chat.
More Distributed Computing packages
gRPC Python is an HTTP/2-based RPC framework that enables you to define and call remote procedures across network boundaries using protocol buffers for serialization.
Install it if you need RPC communication in a distributed system or are integrating with existing gRPC services.
execnet lets you spawn and communicate with Python interpreters across local processes, remote hosts, and different platforms, using a simple API for task distribution and inter-process messaging.
However, the aging maintenance status (275 days since last release) means you should verify it meets your concurrency and performance needs before committing to a…
Cloudpickle extends Python's standard pickle module to serialize lambda functions, interactively-defined functions and classes, and other constructs that the default pickle cannot handle, making it suitable for cluster computing and remote code execution.
Install it if you need to serialize lambda functions, interactively-defined code, or non-standard Python constructs for cluster computing or distributed execution.
Provides a unified, open()-compatible Python API for streaming large files from remote storage (S3, GCS, Azure, HDFS, SFTP, HTTP) and local filesystems, with transparent compression support.
Install it if you work with large files on cloud storage or remote systems and want to avoid writing boilerplate around multiple SDKs.
Portalocker provides cross-platform file locking with support for exclusive and shared locks, plus Redis-based distributed locks and process-aware PID file locking.
Install it if you need file or process coordination; the optional extras (pywin32, redis) are only required for specific lock types.
Ray is a distributed computing framework that scales Python applications from a single machine to multi-node clusters, providing abstractions for parallel tasks, stateful actors, and shared objects.
See also nvidia-resiliency-ext · torchx · torchtitan · metaflow-torchrun · torchrunx · megatron-fsdp · fairscale · lightning · accelerate · torchmetrics