* feat(train): parallel training engine with FSDP2, HSDP, and DCP checkpoints
Replace the FSDP1 training path with a config-owned parallel-training
engine:
- Topology and runtime configs (--parallelism.*, --accelerator.*):
dp_replicate x dp_shard degrees select single-process, DDP (unchanged
default), FSDP2, or HSDP; mixed precision, first-class gradient
accumulation, and FSDP/DDP tuning knobs are mirrored as plain
dataclasses that build the accelerate objects at runtime, so every
run is reproducible from its train_config.json alone. Accelerate env
vars are guarded against configuring the engine behind the config
system's back.
- Declarative policy surface: policies declare FSDP2 wrap units
(_fsdp_wrap_modules) and non-forward entry points
(_fsdp_forward_methods); a shared engine resolves them around
accelerator.prepare(). Context-parallel fields are reserved and
validated to 1.
- Checkpoints: selectable --checkpoint_format (safetensors | dcp |
safetensors_dcp); the sharded optimizer channel is always DCP;
two-phase resume (step+RNG before prepare, DCP model/optimizer after)
reshards across GPU-topology changes; lerobot-convert-dcp merges DCP
shards into a distributable model.safetensors offline.
- Publishing: PreTrainedPolicy.push_model_to_hub is replaced by the
free publish_trained_model (model + processors + card + train config,
all-ranks gather with main-rank writes);
PreTrainedPolicy._save_pretrained gathers state dicts internally,
removing the state_dict= threading from save_pretrained.
- lerobot_train is restructured around the engine: optimizer built
before the single prepare() call, deferred weight load on DCP
resumes, collective save_checkpoint with no call-site rank branches,
dp-world-size-based sample accounting.
Breaking changes: FSDP checkpoints from lerobot <= 0.6.x are not
resumable (weights stay loadable via from_pretrained; pin
lerobot==0.6.x to finish old runs); the `accelerate launch
--config_file` yaml flow is superseded by the config flags; training
autocast is owned exclusively by --accelerator.mixed_precision
(policy.dtype only casts parameters).
Also fixes: reward-model hub publishing crash (TypeError on extra
kwargs).
Verified by ~200 new CPU tests (config round-trips, checkpoint
round-trips per format, two-phase resume, publisher contracts,
converter equivalence, accelerate canaries), a 5-test 4-GPU suite
(FSDP2 save/resume bit-exactness, HSDP/DDP loss parity,
changed-topology resume, all-ranks save_pretrained, grad-accum
equivalence), and end-to-end ACT (1/4/8 GPUs) + FastWAM 6B
(FSDP2 + HSDP) training runs.
* feat(train): FSDP checkpoint saving
* adding docs for FSDP
* adding a test for the fsdp checkpoint path
* cleanup
* fixing final upload to hub
* refactored initial implementation to use torch fsdp api and adding new tests
* Enhance training and logging functionality with accelerator support
- Added support for multi-GPU training by introducing an `accelerator` parameter in training functions.
- Updated `update_policy` to handle gradient updates based on the presence of an accelerator.
- Modified logging to prevent duplicate messages in non-main processes.
- Enhanced `set_seed` and `get_safe_torch_device` functions to accommodate accelerator usage.
- Updated `MetricsTracker` to account for the number of processes when calculating metrics.
- Introduced a new feature in `pyproject.toml` for the `accelerate` library dependency.
* Initialize logging in training script for both main and non-main processes
- Added `init_logging` calls to ensure proper logging setup when using the accelerator and in standard training mode.
- This change enhances the clarity and consistency of logging during training sessions.
* add docs and only push model once
* Place logging under accelerate and update docs
* fix pre commit
* only log in main process
* main logging
* try with local rank
* add tests
* change runner
* fix test
* dont push to hub in multi gpu tests
* pre download dataset in tests
* small fixes
* fix path optimizer state
* update docs, and small improvements in train
* simplify accelerate main process detection
* small improvements in train
* fix OOM bug
* change accelerate detection
* add some debugging
* always use accelerate
* cleanup update method
* cleanup
* fix bug
* scale lr decay if we reduce steps
* cleanup logging
* fix formatting
* encorperate feedback pr
* add min memory to cpu tests
* use accelerate to determin logging
* fix precommit and fix tests
* chore: minor details
---------
Co-authored-by: AdilZouitine <adilzouitinegm@gmail.com>
Co-authored-by: Steven Palma <steven.palma@huggingface.co>
Haoming Song