Saving and loading tensors#

torch.save() andtorch.load() let you easily save and load tensors:

>>>t=torch.tensor([1.,2.])>>>torch.save(t,'tensor.pt')>>>torch.load('tensor.pt')tensor([1., 2.])

By convention, PyTorch files are typically written with a ‘.pt’ or ‘.pth’ extension.

torch.save() andtorch.load() use Python’s pickle by default,so you can also save multiple tensors as part of Python objects like tuples,lists, and dicts:

>>>d={'a':torch.tensor([1.,2.]),'b':torch.tensor([3.,4.])}>>>torch.save(d,'tensor_dict.pt')>>>torch.load('tensor_dict.pt'){'a': tensor([1., 2.]), 'b': tensor([3., 4.])}

Custom data structures that include PyTorch tensors can also be saved if thedata structure is pickle-able.

Saving and loading tensors preserves views#

Saving tensors preserves their view relationships:

>>>numbers=torch.arange(1,10)>>>evens=numbers[1::2]>>>torch.save([numbers,evens],'tensors.pt')>>>loaded_numbers,loaded_evens=torch.load('tensors.pt')>>>loaded_evens*=2>>>loaded_numberstensor([ 1,  4,  3,  8,  5, 12,  7, 16,  9])

Behind the scenes, these tensors share the same “storage.” SeeTensor Views for moreon views and storage.

When PyTorch saves tensors it saves their storage objects and tensormetadata separately. This is an implementation detail that may change in thefuture, but it typically saves space and lets PyTorch easilyreconstruct the view relationships between the loaded tensors. In the abovesnippet, for example, only a single storage is written to ‘tensors.pt’.

In some cases, however, saving the current storage objects may be unnecessaryand create prohibitively large files. In the following snippet a storage muchlarger than the saved tensor is written to a file:

>>>large=torch.arange(1,1000)>>>small=large[0:5]>>>torch.save(small,'small.pt')>>>loaded_small=torch.load('small.pt')>>>loaded_small.storage().size()999

Instead of saving only the five values in thesmall tensor to ‘small.pt,’the 999 values in the storage it shares withlarge were saved and loaded.

When saving tensors with fewer elements than their storage objects, the size ofthe saved file can be reduced by first cloning the tensors. Cloning a tensorproduces a new tensor with a new storage object containing only the valuesin the tensor:

>>>large=torch.arange(1,1000)>>>small=large[0:5]>>>torch.save(small.clone(),'small.pt')# saves a clone of small>>>loaded_small=torch.load('small.pt')>>>loaded_small.storage().size()5

Since the cloned tensors are independent of each other, however, they havenone of the view relationships the original tensors did. If both file size andview relationships are important when saving tensors smaller than theirstorage objects, then care must be taken to construct new tensors that minimizethe size of their storage objects but still have the desired view relationshipsbefore saving.

Saving and loading torch.nn.Modules#

See also:Tutorial: Saving and loading modules

In PyTorch, a module’s state is frequently serialized using a ‘state dict.’A module’s state dict contains all of its parameters and persistent buffers:

>>>bn=torch.nn.BatchNorm1d(3,track_running_stats=True)>>>list(bn.named_parameters())[('weight', Parameter containing: tensor([1., 1., 1.], requires_grad=True)), ('bias', Parameter containing: tensor([0., 0., 0.], requires_grad=True))]>>>list(bn.named_buffers())[('running_mean', tensor([0., 0., 0.])), ('running_var', tensor([1., 1., 1.])), ('num_batches_tracked', tensor(0))]>>>bn.state_dict()OrderedDict([('weight', tensor([1., 1., 1.])),             ('bias', tensor([0., 0., 0.])),             ('running_mean', tensor([0., 0., 0.])),             ('running_var', tensor([1., 1., 1.])),             ('num_batches_tracked', tensor(0))])

Instead of saving a module directly, for compatibility reasons it is recommendedto instead save only its state dict. Python modules even have a function,load_state_dict(), to restore their states from a state dict:

>>>torch.save(bn.state_dict(),'bn.pt')>>>bn_state_dict=torch.load('bn.pt')>>>new_bn=torch.nn.BatchNorm1d(3,track_running_stats=True)>>>new_bn.load_state_dict(bn_state_dict)<All keys matched successfully>

Note that the state dict is first loaded from its file withtorch.load()and the state then restored withload_state_dict().

Even custom modules and modules containing other modules have state dicts andcan use this pattern:

# A module with two linear layers>>>classMyModule(torch.nn.Module):def__init__(self):super().__init__()self.l0=torch.nn.Linear(4,2)self.l1=torch.nn.Linear(2,1)defforward(self,input):out0=self.l0(input)out0_relu=torch.nn.functional.relu(out0)returnself.l1(out0_relu)>>>m=MyModule()>>>m.state_dict()OrderedDict([('l0.weight',tensor([[0.1400,0.4563,-0.0271,-0.4406],[-0.3289,0.2827,0.4588,0.2031]])),('l0.bias',tensor([0.0300,-0.1316])),('l1.weight',tensor([[0.6533,0.3413]])),('l1.bias',tensor([-0.1112]))])>>>torch.save(m.state_dict(),'mymodule.pt')>>>m_state_dict=torch.load('mymodule.pt')>>>new_m=MyModule()>>>new_m.load_state_dict(m_state_dict)<Allkeysmatchedsuccessfully>

Serialized file format fortorch.save#

Since PyTorch 1.6.0,torch.save defaults to returning an uncompressed ZIP64archive unless the user sets_use_new_zipfile_serialization=False.

In this archive, the files are ordered as such

checkpoint.pth├── data.pkl├── byteorder  # added in PyTorch 2.1.0├── data/│   ├── 0│   ├── 1│   ├── 2│   └── …└── version

The entries are as follows:

• data.pkl is the result of pickling the object passed totorch.saveexcludingtorch.Storage objects that it contains

• byteorder contains a string with thesys.byteorder when saving (“little” or “big”)

• data/ contains all the storages in the object, where each storage is a separate file

• version contains a version number at save time that can be used at load time

When saving, PyTorch will ensure that the local file header of each file is paddedto an offset that is a multiple of 64 bytes, ensuring that the offset of each fileis 64-byte aligned.

Layout Control#

Themmap argument intorch.load() allows for lazy loading of tensor storages.

In addition, there are some advanced features that allow for more fine-grainedcontrol and manipulation of atorch.save checkpoint.

Thetorch.serialization.skip_data context manager enables

• Saving a checkpoint withtorch.save that includes empty space for data bytesto be written later.

• Loading a checkpoint withtorch.load and filling in the data bytes of tensors later.

To inspect tensor metadata in atorch.save checkpoint without allocating memory for storagedata, usetorch.load within theFakeTensorMode context manager. On top of skipping loadingstorage data similar toskip_data above, it additionally tags storages with their offset withinthe checkpoint, enabling direct checkpoint manipulation.

importtorch.nnasnnfromtorch._subclasses.fake_tensorimportFakeTensorModem=nn.Linear(10,10)torch.save(m.state_dict(),"checkpoint.pt")withFakeTensorMode()asmode:fake_sd=torch.load("checkpoint.pt")fork,vinfake_sd.items():print(f"key={k}, dtype={v.dtype}, shape={v.shape}, stride={v.stride()}, storage_offset={v.storage_offset()}")# offset of the storage in the checkpointprint(f"key={k}, checkpoint_offset={v.untyped_storage()._checkpoint_offset}")

For more information,this tutorialoffers a comprehensive example of using these features to manipulate a checkpoint.

torch.load withweights_only=True#

Starting in version 2.6,torch.load will useweights_only=True if thepickle_moduleargument is not passed.

_pickle.UnpicklingError: Weights only load failed. This file can still be loaded,to do so you have two options, do those steps only if you trust the source of the checkpoint.    1. Re-running `torch.load` with `weights_only` set to `False` will likely succeed,        but it can result in arbitrary code execution. Do it only if you got the file from a trusted source.    2. Alternatively, to load with `weights_only=True` please check the recommended       steps in the following error message.       WeightsUnpickler error: Unsupported global: GLOBAL {__module__}.{__name__} was not an allowed global by       default. Please use `torch.serialization.add_safe_globals([{__name__}])` or the       `torch.serialization.safe_globals([{__name__}])` context manager to allowlist this global       if you trust this class/function.

WeightsUnpickler error: Can only build Tensor, Parameter, OrderedDict or types allowlisted via `add_safe_globals`,but got <class 'numpy.dtype[float32]'>

WeightsUnpickler error: Can only build Tensor, Parameter, OrderedDict or types allowlisted via `add_safe_globals`,but got <class 'numpy.dtypes.Float32DType'>

Utility functions#

The following utility functions are related to serialization:

torch.serialization.register_package(priority,tagger,deserializer)[source]#

Registers callables for tagging and deserializing storage objects with an associated priority.Tagging associates a device with a storage object at save time while deserializing moves astorage object to an appropriate device at load time.tagger anddeserializerare run in the order given by theirpriority until a tagger/deserializer returns avalue that is notNone.

To override the deserialization behavior for a device in the global registry, one can register atagger with a higher priority than the existing tagger.

This function can also be used to register a tagger and deserializer for new devices.

Parameters:

• priority (int) – Indicates the priority associated with the tagger and deserializer, where a lowervalue indicates higher priority.

• tagger (Callable[[Storage |TypedStorage |UntypedStorage],str |None]) – Callable that takes in a storage object and returns its tagged device as a stringor None.

• deserializer (Callable[[Storage |TypedStorage |UntypedStorage,str],Storage |TypedStorage |UntypedStorage |None]) – Callable that takes in storage object and a device string and returns a storageobject on the appropriate device or None.

Returns:

None

Example

>>>defipu_tag(obj):>>>ifobj.device.type=='ipu':>>>return'ipu'>>>defipu_deserialize(obj,location):>>>iflocation.startswith('ipu'):>>>ipu=getattr(torch,"ipu",None)>>>assertipuisnotNone,"IPU device module is not loaded">>>asserttorch.ipu.is_available(),"ipu is not available">>>returnobj.ipu(location)>>>torch.serialization.register_package(11,ipu_tag,ipu_deserialize)

torch.serialization.get_crc32_options()[source]#

Get whethertorch.save() computes and writes crc32 for each record.

Defaults toTrue.

Return type:

bool

torch.serialization.set_crc32_options(compute_crc32)[source]#

Set whethertorch.save() computes and writes crc32 for each record.

Note

Setting this toFalse may make unzipping of thetorch.save outputfail or warn due to corrupted CRC32. Howevertorch.load will beable to load the file.

Parameters:

compute_crc32 (bool) – set crc32 computation flag

torch.serialization.get_default_load_endianness()[source]#

Get fallback byte order for loading files

If byteorder mark is not present in saved checkpoint,this byte order is used as fallback.By default, it’s “native” byte order.

Returns:

Optional[LoadEndianness]

Return type:

default_load_endian

torch.serialization.set_default_load_endianness(endianness)[source]#

Set fallback byte order for loading files

If byteorder mark is not present in saved checkpoint,this byte order is used as fallback.By default, it’s “native” byte order.

Parameters:

endianness – the new fallback byte order

torch.serialization.get_default_mmap_options()[source]#

Get default mmap options fortorch.load() withmmap=True.

Defaults tommap.MAP_PRIVATE.

Returns:

int

Return type:

default_mmap_options

torch.serialization.set_default_mmap_options(flags)[source]#

Context manager or function to set default mmap options fortorch.load() withmmap=True to flags.

For now, only eithermmap.MAP_PRIVATE ormmap.MAP_SHARED are supported.Please open an issue if you need any other option to be added here.

Note

This feature is currently not supported for Windows.

Parameters:

flags (int) –mmap.MAP_PRIVATE ormmap.MAP_SHARED

torch.serialization.add_safe_globals(safe_globals)[source]#

Marks the given globals as safe forweights_only load. For example, functionsadded to this list can be called during unpickling, classes could be instantiatedand have state set.

Each item in the list can either be a function/class or a tuple of the form(function/class, string) where string is the full path of the function/class.

Within the serialized format, each function is identified with its fullpath as{__module__}.{__qualname__}. When calling this API, you can provide thisfull path that should match the one in the checkpoint otherwise the default{fn.__module__}.{fn.__qualname__} will be used.

Parameters:

safe_globals (List[Union[Callable,Tuple[Callable,str]]]) – list of globals to mark as safe

Example

>>>importtempfile>>>classMyTensor(torch.Tensor):...pass>>>t=MyTensor(torch.randn(2,3))>>>withtempfile.NamedTemporaryFile()asf:...torch.save(t,f.name)# Running `torch.load(f.name, weights_only=True)` will fail with# Unsupported global: GLOBAL __main__.MyTensor was not an allowed global by default.# Check the code and make sure MyTensor is safe to be used when loaded from an arbitrary checkpoint....torch.serialization.add_safe_globals([MyTensor])...torch.load(f.name,weights_only=True)# MyTensor([[-0.5024, -1.8152, -0.5455],#          [-0.8234,  2.0500, -0.3657]])

torch.serialization.clear_safe_globals()[source]#

Clears the list of globals that are safe forweights_only load.

torch.serialization.get_safe_globals()[source]#

Returns the list of user-added globals that are safe forweights_only load.

Return type:

list[Callable |tuple[Callable,str]]

torch.serialization.get_unsafe_globals_in_checkpoint(f)[source]#

Returns a list of strings of functions/classes in atorch.save object that are not safe forweights_only.

For a given function or classf, the corresponding string will be of the form{f.__module__}.{f.__name__}.

This function will return any GLOBALs in the checkpoint that are not in the set marked safeforweights_only (either viaadd_safe_globals() orsafe_globals context orallowlisted bytorch by default).

Note

This function will statically disassemble the pickle file in the checkpoint.The implication is any classes dynamically pushed onto the stack during unpicklingwill not be included in the output.

Parameters:

f (str |PathLike[str]|IO[bytes]) – File-like object or string containing the checkpoint object saved viatorch.save

Returns:

A list of strings of pickle GLOBALs in the checkpoint that are not allowlisted forweights_only.

Return type:

list[str]

classtorch.serialization.safe_globals(safe_globals)[source]#

Context-manager that adds certain globals as safe forweights_only load.

Parameters:

safe_globals (list[Callable |tuple[Callable,str]]) – List of globals for weights_only load.

Example

>>>importtempfile>>>classMyTensor(torch.Tensor):...pass>>>t=MyTensor(torch.randn(2,3))>>>withtempfile.NamedTemporaryFile()asf:...torch.save(t,f.name)# Running `torch.load(f.name, weights_only=True)` will fail with# Unsupported global: GLOBAL __main__.MyTensor was not an allowed global by default.# Check the code and make sure MyTensor is safe to be used when loaded from an arbitrary checkpoint....withtorch.serialization.safe_globals([MyTensor]):...torch.load(f.name,weights_only=True)# MyTensor([[-0.5024, -1.8152, -0.5455],#          [-0.8234,  2.0500, -0.3657]])>>>asserttorch.serialization.get_safe_globals()==[]

classtorch.serialization.skip_data(materialize_fake_tensors=False)[source]#

Context-manager that skips writing/reading storage bytes fortorch.save /torch.load calls.

For the save path, storages will still be saved, but the space that their bytes would usually be written towill be empty space. The storage bytes can then be populated in a separate pass.

For the load path, tensors will be loaded per the checkpoint but their storages will not be populated with data.

Warning

Theskip_data context manager is an early prototype and is subject to change.

Parameters:

materialize_fake_tensors (bool) – Whether to materialize FakeTensors during save. This is a no-op for the load path.

Example

>>>importtempfile>>>t=torch.randn(2,3)>>>withtempfile.NamedTemporaryFile()asf:...withtorch.serialization.skip_data():...torch.save(t,f.name)...torch.load(f.name,weights_only=True)tensor([[0., 0., 0.],        [0., 0., 0.]])

Config#

torch.utils.serialization.config provides a global config that can control the behavior oftorch.save andtorch.load.

torch.utils.serialization.config.save contains options that control the behavior oftorch.save.

• compute_crc32: whether to compute and write the zip file checksum (Default :True).Seeset_crc32_options().

• use_pinned_memory_for_d2h: for storages that are on an accelerator when passed totorch.save, whether tomove storage to pinned memory or pageable memory on CPU withintorch.save. (Default:False (i.e. pageable))

• storage_alignment: alignment of storages in the checkpoint duringtorch.save in bytes. (Default64)

torch.utils.serialization.config.load contains options that control the behavior oftorch.load.

• mmap: See the documentation formmap argument intorch.load().This config will set the behavior ofmmap fortorch.load if it is notalready explicitly passed to thetorch.load call (Default :False).

• endianness: Seeset_default_load_endianness().(Default :torch.serialization.LoadEndianness.NATIVE)

• mmap_flags: Seeset_default_mmap_options.(Default :MAP_PRIVATE)

• calculate_storage_offsets: If this config is set toTrue, offsets for storages will becalculated rather than read via random reads when usingtorch.load(mmap=True). This minimizesrandom reads, which can be helpful when the file is being loaded over a network. (Default :False)