Ravex

Transparent checkpoint and resume for PyTorch training. Your training script does not change — not one line, not one import.

pip install ravex

Ravex has a small Rust core — the reshard planner and the replica transport — so wheels are built per interpreter for Linux x86_64. On any other platform pip falls back to the source distribution, which builds if you have a Rust toolchain. Through 0.0.5 Ravex was pure Python and installed anywhere; what that bought and what it cost is in the CHANGELOG, along with everything else that changed between versions.

Its default checkpoint engine, Moonclip, ships wheels for the same platform — so pip install "ravex[moonclip]" is a Linux thing, and where it is unavailable Ravex falls back to torch_save on its own.

Drop a ravex.yaml next to your code and run what you always ran:

python train.py

The run now checkpoints itself. If the process dies — spot instance reclaimed, node rebooted, OOM killer, power cut — running the same command again picks up where it stopped: same weights, same optimizer moments, same LR schedule, same RNG state, same position in the dataset.

Why this is not just torch.save

Saving the weights is the easy part. What makes a resumed run continue rather than merely restart from a good place is everything around them:

Restored	Why it matters
Model weights and buffers	the obvious part
Optimizer state	Adam moments; without them the first steps after resume are wrong
LR scheduler	resume at the wrong LR and the loss curve visibly kinks
AMP GradScaler	its loss scale is tuned state, not a constant
RNG: torch, CUDA, Python, NumPy	dropout masks and augmentations replay identically
Dataset position	you continue with batch 4001, not batch 1

Ravex's test suite asserts the strong form of this: a run killed at step 20 and resumed produces losses that are bit-identical, step by step, to the run that was never interrupted, and the same final weights.

How you use it

One decorator on the function that trains:

import ravex

@ravex.train_loop(preemption_handler=True)
def train():
    model = build_model()
    optimizer = torch.optim.Adam(model.parameters())
    for batch in loader:
        ...

train()

Inside that function nothing changes. Ravex notices the model, the optimizer, the LR scheduler, the AMP scaler and the dataloader as they are built, counts every optimizer.step(), and writes a checkpoint on the cadence you configured. Rerun the script and it picks up where it stopped.

If "on its own" is a guess — two models and only one being trained, an optimizer built before the function was entered — name it:

ravex.track(model, optimizer)

What the decorator buys, beyond being readable. It is a boundary. Ravex knows where the loop begins and ends, so the teardown is a finally rather than an atexit hook hoping to run, and it holds the function before the model exists — which is what resuming onto a different number of ranks needs and could never have.

What it replaced. Through 0.0.5 pip install ravex also put a one-line .pth file in site-packages that ran in every Python process in the environment and attached Ravex to any of them that had a ravex.yaml above the working directory. It is gone. It was genuinely zero-code-changes, and it existed so a platform could turn on checkpointing for code it did not own — a real thing to give up, given up on purpose. What it cost was that Ravex had to guess where the loop began, and the guessing is what limited it: with HuggingFace Trainer and Lightning, dataset and RNG replay could not be made exact, and there was no defined moment of exit to write a final sharded checkpoint at. Both were filed as accepted limitations. Neither has to be, now.

Configuration

ravex.yaml, anywhere at or above the working directory:

checkpoint_every: 500        # optimizer steps between checkpoints
backend: moonclip            # moonclip | torch_save
storage:
  type: local                # local | s3 | r2
  path: ./checkpoints
keep_last: 5
max_steps: null              # optional hard stop, see below
sharded_checkpoints: gather  # gather | per_rank, for FSDP — see below

Every option also reads from RAVEX_* environment variables, which win over the file — so a scheduler can override a config committed to the repository:

RAVEX_CHECKPOINT_EVERY=100 RAVEX_STORAGE_TYPE=r2 RAVEX_STORAGE_BUCKET=runs python train.py

Credentials are never read from the config file. Set RAVEX_S3_ACCESS_KEY / RAVEX_S3_SECRET_KEY, or the usual AWS_* pair.

Full reference: docs/configuration.md.

Backends

moonclip (default) — the Moonclip engine: per-tensor delta tracking, so unchanged weights cost zero I/O; zstd compression; background writes; direct S3/R2 sync.

torch_save — one .pt file per checkpoint, written on a background thread. Used automatically when Moonclip is not installed. Correct, just larger and slower.

How it works

Ravex patches five things in PyTorch and nothing in your code:

• nn.Module.__init__ and .train() — to notice your models
• Optimizer.__init__ — to attach a step hook to every optimizer
• DataLoader.__init__ and .__iter__ — to track the dataset position and to find the one moment where a resume can be applied

The step counter advances once per optimizer.step(), so gradient accumulation needs no special handling. Checkpoints are collected at the top of an iteration, never inside one: mid-iteration the LR scheduler has not stepped yet, and a checkpoint taken there resumes with a stale learning rate.

Collection runs on the training thread — it has to, to be consistent with the step that just finished — and copies the state; the write itself happens in the background. What the loop pays for is the copy, not the I/O.

More detail: docs/how-it-works.md.

Safety

Ravex is designed to be un-noticeable when it works and harmless when it does not:

• every hook is wrapped; if one raises, your call still returns normally
• if a checkpoint fails, Ravex disables itself and logs it — training continues
• nothing is ever written to stdout; logs go to log_file, or to stderr at WARNING and above
• installing the package starts no checkpointing anywhere, and neither does importing it. Ravex does nothing at all until a decorated function is called, which is a much easier promise to check than the one this used to make: it is visible in the source of the thing being checkpointed
• the patches go on at that call and come off when it returns, including when it returns by raising. A process that finishes training is a process with an unpatched PyTorch in it
• RAVEX_ENABLED=0 turns it off for a single run without touching the code

Status and limits

Alpha. Works with plain PyTorch loops, and with anything built on them, since the hooks are on PyTorch itself.

With a framework driving the loop

HuggingFace Trainer and Lightning are covered by their own tests, and the result deserves to be stated precisely rather than as "it works":

• State restoration is exact. Model, optimizer, LR scheduler and step count all come back. With the per-step randomness removed, a killed run resumes into a loss sequence identical to the uninterrupted one.
• Replay is not. With shuffling and dropout on, the resumed run continues correctly from the checkpointed state but sees a different draw. Both frameworks iterate the dataloader on their own schedule and consume the global RNG around the loop, so the epoch-start snapshot no longer lines up.

Plain loops, DDP and FSDP are bit-exact with randomness on. This is a framework-interaction limit, not a general one, and it costs you a different shuffle from the resume point onwards — not a wrong model.

Verified: plain loops, gradient accumulation, LR schedulers, AMP loss-scale state, num_workers > 0, DDP, and FSDP. A killed torchrun job resumes on every rank with bit-identical losses, sharded or not, and the checkpoint it leaves behind loads into a plain single-process model afterwards.

Sharded models

With FSDP each rank holds a slice of every parameter, so state_dict() returns a fragment. Two ways to turn that into a checkpoint, picked with sharded_checkpoints:

gather (default) rebuilds the whole state on rank 0, which writes it. The checkpoint is then independent of the topology that produced it — eight GPUs in, one out — and it does not scale: rank 0 has to hold the entire model and optimizer in host memory, and it is the rank that then does the writing.

per_rank has every rank write its own shard into its own store, <storage.path>/rank_<n>. Nothing is gathered, so nothing is bounded by one rank's memory, and on a 1.48B model collecting the state went from 15.6 s to 1.5 s. What you give up is the resharding: those shards are cut for one topology, so the checkpoint resumes at the same world size and starts clean at any other. Needs FSDP2 — under FSDP1 Ravex degrades to gather and says so.

Either way, collecting is a collective: every rank participates, and there is no final checkpoint at exit for a sharded model. Shutdown is where ranks stop being in lockstep, and a collective nobody else joins hangs. Losing the last few steps is bounded; a hang is not. Set checkpoint_every accordingly.

Numbers and the FSDP1 details: docs/configuration.md.

More than one machine

Ranks, directories and collectives all cross machines unchanged. What does not is per_rank on local storage: each machine writes only its own ranks' shards to its own disk, so no machine holds a whole checkpoint. It resumes only if every machine is handed the same ranks again — no launcher promises that — and not at all if a machine is lost. Ravex probes the storage at activation and says which case you are in rather than letting you find out at the first resume.

Two ways out, and they are not equivalent:

• A bucket (storage.type: s3). Checkpoints leave the machines on their own, and since 0.0.4 a rank that comes up with an empty disk pulls its store back. Before that the remote was push-only — a backup you could not resume from.
• replicate_every, when there is no bucket and no shared filesystem. Each rank copies its store to a peer on another machine every N checkpoints. Survives losing any one machine, at a cost of at most N checkpoints of progress.

Keeping the data after the run ends is yours unless a remote is configured — and for a sharded model the newest checkpoint is the last periodic one, since there is none at exit.

Known limits today:

• IterableDataset: no index sampler exists, so the stream position cannot be replayed. Everything else is still restored.
• Your loop's bounds: a resumed script runs its own for epoch in range(N) again from the top; it has no idea 3000 steps already happened. Set max_steps and Ravex ends the run at the right step regardless of how many times the process restarted.

Under AMP, note that an overflowing gradient makes scaler.step() skip the optimizer. Ravex counts optimizer steps, not loop iterations, so a skipped iteration does not advance the counter — which is the right unit, since nothing about the model changed, but it does mean the step count and the number of batches you fed differ.

The GPU paths — AMP with real fp16 overflow, the CUDA RNG, FSDP1, NCCL — are covered by integration/test_cuda.py, which skips without a GPU. They were last verified on 8× RTX 5060 Ti with torch 2.12/cu130.

Project layout

Everything under ravex is private except ravex itself: the public surface is train_loop(), track(), a handful of small helpers (checkpoint, flush, step, status, is_active, deactivate), __version__ and the ravex command. Every name below that starts with an underscore is free to move.

ravex/
├── __init__.py       Public surface, and nothing else
├── _cli.py           ravex status
├── _core.pyi         Types for the compiled Rust core
├── _config.py        Defaults < ravex.yaml < RAVEX_*
├── _patches.py       The five monkey patches on PyTorch
├── _registry.py      What is being trained, held by weakref
├── _runtime.py       One per process; what the patches call into
├── _resume.py        Best-effort restore
├── _backends.py      moonclip | torch_save
├── _sampler.py       Dataset position
├── _frameworks.py    HF Trainer / Lightning / Accelerate detection
├── _dist/            More than one GPU, more than one machine
│   ├── collectives.py    gather vs per_rank; the SIGTERM channel
│   ├── reshard.py        8 shards onto 4 ranks — re-exports the Rust core
│   ├── identity.py       Who wrote this store, as part of which run
│   ├── replication.py    Each rank copies its store to a peer — framing from the Rust core
│   └── elastic.py        Membership changes without a restart
└── _interop/         Checkpoints somebody else wrote
    ├── foreign.py        What is this directory? Layout first, fields second
    ├── zero.py           DeepSpeed ZeRO stages 1–3
    ├── dcp.py            torch.distributed.checkpoint — FSDP, Megatron-core
    ├── convert.py        Into the shape the resume path already consumes
    └── resume.py         When to act on all that, and when to decline

The two subpackages are groups, not layers: neither re-exports anything, and callers import the submodule they want inside the function that wants it. That used to be a startup-cost rule — the .pth ran in every interpreter on the machine, so an import that did not have to happen was paid for by every python -c on the box. The .pth is gone and the discipline is kept for the weaker reason that still holds: _dist.collectives pulls in torch.distributed, and a single-process run should load neither subpackage.

The import graph is a DAG with _runtime as its only hub; there are no cycles, and nothing in _interop is imported by anything outside it except _runtime.

And around them

Path
tests/	821 unit tests, in-process, no GPU and no container. Named for what they cover: test_dist_*, test_interop_*
integration/	What only exists across a real process boundary — a resume from an empty interpreter, torchrun, a SIGKILL that runs no finally. Linux, in Docker
integration/scripts/	The training scripts those tests kill and restart
integration/multinode/	One container per rank, for questions --nproc_per_node cannot ask (README)
integration/two-machines/	The rented-box harness: two real hosts, real network
docs/	configuration.md, how-it-works.md
src/	The Rust core: reshard.rs is the planner, transport.rs frames and moves a store, python.rs is the only file that knows an interpreter exists
.forgejo/workflows/	checks.yml on branches; ci.yml on main adds the moonclip backend and both integration jobs

As of 0.1.0 that is about 13k lines across 26 Python modules, plus 2.7k of Rust, against 12k lines of tests.

Development

pip install -e ".[dev]"
pytest

That editable install compiles the Rust core into the source tree, so it needs a toolchain: rust-toolchain.toml names the version and rustup will fetch it. The half of the engine that has no Python in it has its own tests, and they are the faster gate — no interpreter, no torch, milliseconds:

cargo test --no-default-features

The unit suite runs in-process. The parts that only exist across a real process boundary — a resume starting from an empty interpreter, config discovery from the working directory, torchrun, and a SIGKILL that runs no finally — live in integration/ and need Linux:

docker build -f integration/Dockerfile -t ravex-integration .
docker run --rm ravex-integration

Licence

Apache 2.0. See LICENSE.