CKPT file format: contents

.ckpt has no single format. A PyTorch Lightning checkpoint is one file written by torch.save, so its bytes are those of a PyTorch archive; what matters is the dictionary inside. A TensorFlow checkpoint is a set of files sharing a prefix such as model.ckpt. This page follows Lightning’s checkpoint_connector.py (dump_checkpoint) and TensorFlow’s tensor_bundle.h, tensor_bundle.cc and tensor_bundle.proto, all read 2026-11-12.

PyTorch Lightning

The pickled object is a dictionary. Lightning always writes the first five keys; the next ones are left out when the checkpoint is saved with weights_only.

Key	Value
`epoch`	The training epoch. The code notes it is kept for compatibility and not used for restoring.
`global_step`	The global training step, likewise.
`pytorch-lightning_version`	The Lightning version that wrote the file.
`state_dict`	The model’s `state_dict`: parameter names mapped to tensors.
`loops`	State of `fit_loop`, `validate_loop`, `test_loop` and `predict_loop`.
`callbacks`	Each callback’s state (not with `weights_only`).
`optimizer_states`	A list with each optimiser’s `state_dict` (not with `weights_only`).
`lr_schedulers`	A list with each learning-rate scheduler’s `state_dict` (not with `weights_only`).
The precision plugin’s class name	Its state, when it has any (not with `weights_only`).
`hyper_parameters`	The model’s and data module’s saved hyperparameters, as a dictionary (or an OmegaConf container).
`hparams_name`, `hparams_type`	The name the hyperparameters were saved under, and the container type when OmegaConf is used.
The data module’s class name	Its `state_dict`, when it returns one.
Anything else	The model’s `on_save_checkpoint` hook may add its own keys.

TensorFlow checkpoint bundle

A tensor bundle is a set of immutable files with a common prefix:

ckpt.index
ckpt.data-00000-of-00020
ckpt.data-00001-of-00020
…
ckpt.data-00019-of-00020

The .index file is a string-to-string immutable table (TensorFlow’s table::Table). Each key is a tensor name and each value a serialised BundleEntryProto.
The key that is the empty string, always the first in the table, holds a BundleHeaderProto.
The .data-NNNNN-of-NNNNN shards hold the raw tensor bytes. The writer can pad each tensor to a set data alignment.

Message	Field	Meaning
`BundleHeaderProto`	1 `num_shards`	Number of data files.
	2 `endianness`	0 little (the default), 1 big. It affects only the tensor bytes, not the protobufs, and a bundle can only be read on a platform of the same byte order.
	3 `version`	A `VersionDef`; the writer sets producer 1 and minimum consumer 0.
`BundleEntryProto`	1 `dtype`, 2 `shape`	The tensor’s data type and shape.
	3 `shard_id`	Which data file holds the bytes.
	4 `offset`	Start of the bytes in that file.
	5 `size`	Number of bytes.
	6 `crc32c`	CRC-32C of the tensor bytes.
	7 `slices`	Present only for a partitioned tensor: then fields 1 and 2 describe the full tensor, fields 3 to 6 are ignored, and each slice has its own entry, keyed by the slice name.

.CKPT File Format – Contents

PyTorch Lightning

TensorFlow checkpoint bundle