torchtune.training

Checkpointing

torchtune offers checkpointers to allow seamless transitioning between checkpoint formats for training and interoperability with the rest of the ecosystem. For a comprehensive overview of checkpointing, please see the checkpointing deep-dive.

FullModelHFCheckpointer	Checkpointer which reads and writes checkpoints in HF's format.
FullModelMetaCheckpointer	Checkpointer which reads and writes checkpoints in Meta's format.
FullModelTorchTuneCheckpointer	Checkpointer which reads and writes checkpoints in a format compatible with torchtune.
ModelType	ModelType is used by the checkpointer to distinguish between different model architectures.
FormattedCheckpointFiles	This class gives a more concise way to represent a list of filenames of the format file_{i}_of_{n_files}.pth.
update_state_dict_for_classifier	Validates the state dict for checkpoint loading for a classifier model.

Reduced Precision

Utilities for working in a reduced precision setting.

get_dtype	Get the torch.dtype corresponding to the given precision string.
set_default_dtype	Context manager to set torch's default dtype.
validate_expected_param_dtype	Validates that all input parameters have the expected dtype.
get_quantizer_mode	Given a quantizer object, returns a string that specifies the type of quantization.

Distributed

Utilities for enabling and working with distributed training.

init_distributed	Initialize process group required for torch.distributed.
is_distributed	Check if all environment variables required to initialize torch.distributed are set and distributed is properly installed.
get_world_size_and_rank	Function that gets the current world size (aka total number of ranks) and rank number of the current process in the default process group.
gather_cpu_state_dict	Converting sharded state dict into a full state dict on CPU Returning non-empty result only on rank0 to avoid peaking CPU memory

Memory Management

Utilities to reduce memory consumption during training.

apply_selective_activation_checkpointing	Utility to setup activation checkpointing and wrap the model for checkpointing.
set_activation_checkpointing	Utility to apply activation checkpointing to the passed-in model.
OptimizerInBackwardWrapper	A bare-bones class meant for checkpoint save and load for optimizers running in backward.
create_optim_in_bwd_wrapper	Create a wrapper for optimizer step running in backward.
register_optim_in_bwd_hooks	Register hooks for optimizer step running in backward.

Schedulers

Utilities to control lr during the training process.

get_cosine_schedule_with_warmup	Create a learning rate schedule that linearly increases the learning rate from 0.0 to lr over num_warmup_steps, then decreases to 0.0 on a cosine schedule over the remaining num_training_steps-num_warmup_steps (assuming num_cycles = 0.5).
get_lr	Full_finetune_distributed and full_finetune_single_device assume all optimizers have the same LR, here to validate whether all the LR are the same and return if True.

Metric Logging

Various logging utilities.

metric_logging.CometLogger	Logger for use w/ Comet (https://www.comet.com/site/).
metric_logging.WandBLogger	Logger for use w/ Weights and Biases application (https://wandb.ai/).
metric_logging.TensorBoardLogger	Logger for use w/ PyTorch's implementation of TensorBoard (https://pytorch.org/docs/stable/tensorboard.html).
metric_logging.StdoutLogger	Logger to standard output.
metric_logging.DiskLogger	Logger to disk.

Performance and Profiling

torchtune provides utilities to profile and debug the memory and performance of your finetuning job.

get_memory_stats	Computes a memory summary for the passed in device.
log_memory_stats	Logs a dict containing memory stats to the logger.
setup_torch_profiler	Sets up profile and returns the profiler config with post-setup updates.

Miscellaneous

get_unmasked_sequence_lengths	Returns the sequence lengths for each batch element, excluding masked tokens.
set_seed	Function that sets seed for pseudo-random number generators across commonly used libraries.