This document explains the TorchScript serialization format, and the anatomy
of a call to torch::jit::save()
or torch::jit::load()
.
- Overview
code/
: How code is serialized- How data is serialized
constants.pkl
: Constants in codetorch:jit::load()
__getstate__
and__setstate__
- Appendix:
CompilationUnit
and code object ownership
A serialized model (call it model.pt
) is a ZIP archive containing many
files. If you want to manually crack it open, you can call unzip
on it to
inspect the file structure directly:
$ unzip model.pt
Archive: model.pt
extracting ...
$ tree model/
├── code/
│ ├── __torch__.py
│ ├── __torch__.py.debug_pkl
│ ├── foo/
│ │ ├── bar.py
│ │ ├── bar.py.debug_pkl
├── data.pkl
├── constants.pkl
└── data/
├── 0
└── 1
You'll notice that there are .py
and .pkl
files in this archive. That's
because our serialization format tries to mimic Python's. All "code-like"
information (methods, modules, classes, functions) are stored as
human-readable .py
containing valid Python syntax, and all "data-like"
information (attributes, objects, etc.) are pickled using a subset of
Python's pickle protocol.
A model is really a top-level module with some submodules, parameters, and so
on depending on what the author needs. So, data.pkl
contains the pickled
top-level module. Deserializing the model is as simple as calling
unpickle()
on data.pkl
, which will restore the module object state and
load its associated code on demand.
Some things to keep in mind while working on the serialization code. These may help make technical decisions on which approach to take when making a change.
Do what Python does. When it comes to the serialized format, it's much simpler in the long-run to be consistent with whatever Python does. A good rule of thumb is: if I tried to interact with serialized artifacts using Python, would it work? i.e., all serialized code should be valid Python, and all pickled objects should be depickle-able by Python.
Being consistent with Python means our format is more debuggable (you can always crack it open and poke at it from Python) and leads to fewer surprises for developers familiar with Python but not familiar with TorchScript.
Human readable. In addition to being valid Python, serialized code should attempt to be readable Python. We should try to preserve the variable names that authors wrote, appropriately inline short expressions, and so on. This helps with debugging the serialized code.
No jitter. If we do:
m = MyModule()
m.save("foo.pt")
m_loaded = torch.load("foo.pt")
m_loaded.save("foo2.pt")
m_loaded2 = torch.load("foo2.pt")
We want the property that m_loaded
and m_loaded2
are identical. This
"no-jitter" property is useful in catching bugs in the serialization process,
and generally is desirable for debugging (models won't drift depending on how
many times you saved/loaded them).
Initial load should be fast. Calling load()
should be effectively
instantaneous to a human. Anything that takes a long time (reading in tensor
data, for example) should be done lazily.
At a high level, code serialization means:
- Transforming
ClassType
s andFunction
s (called "code objects") into Python source code. - Placing the source code in the model ZIP archive.
PythonPrint
is the function that takes as input a ClassType
or Function
("code object") and outputs Python source code. ScriptModule
s are
implemented as class types, so their methods and attributes will get
serialized as well.
PythonPrint
works by walking a Graph
(the IR representation of either a
ClassType
's method or raw Function
) and emitting Python code that
corresponds to it. The rules for emitting Python code are mostly
straightforward and uninteresting. There are some extra pieces of information
that PythonPrint
tracks, however:
Class dependencies. While walking the graph, PythonPrint
keeps track of
what classes are used in the graph and adds them to a list of classes that
the current code object depends on. For example, if we are printing a
Module
, it will depend on its submodules, as well as any classes used in
its methods or attributes.
Uses of tensor constants. Most constants are inlined as literals, like
strings or ints. But since tensors are potentially very large, when
PythonPrint
encounters a constant tensor it will emit a reference to a
global CONSTANTS
table (like foo = CONSTANTS.c0
).
When importing, the importer will know how to resolve this reference into an
actual tensor by looking it up in the tensor table. So CONSTANTS.c0
means
"this is the 0th
tensor in the tensor tuple in constants.pkl
." See
the constants section for more info.
Original source range records. To aid debugging, PythonPrint
remembers
the "original" (user-written) location of the source code it's emitting. That
way, when the user is debugging a model they loaded, they will see
diagnostics that point to the code that they actually wrote, rather than the
code that PythonPrint
emitted.
The original source range records are pickled and saved in a corresponding
.debug_pkl
file with the same name as the code. You can think of this
.debug_pkl
file as a map between source ranges in the serialized code and
the original user-written code.
Module information. Modules are special in a few ways. First are
Parameter
s: some module attributes are actually Parameter
s, which have
special properties (see the torch.nn
documentation for exact
details). We track which attributes are parameters by emitting a special
assignment in the class body, like:
class MyModule(Module):
__parameters__ = ["foo", "bar", ]
foo : Tensor
bar : Tensor
attribute_but_not_param : Tensor
Another special thing with modules is that they are typically constructed in
Python, and we do not compile the __init__()
method. So in order to ensure
they are statically typed, PythonPrint
must enumerate a module's attributes
(as you can see above), because it can't rely on compiling __init__()
to
infer the attributes.
A final special thing is that some modules (like nn.Sequential
) have
attributes that are not valid Python identifiers. We can't write
# wrong!
class MyModule(Module):
0 : ASubmodule
1 : BSubmodule
because this is not valid Python syntax (even though it is legal in Python to
have attributes with those names!). So we use a trick where we write directly
to the __annotations__
dict:
class MyModule(Module):
__annotations__ = []
__annotations__["0"] = ASubmodule
__annotations__["1"] = ASubmodule
Once all code objects have been PythonPrint
ed into source strings, we have
to figure out where to actually put this source. Explaining this necessitates
an introduction to CompilationUnit
and QualifiedName
. See the appendix on
CompilationUnit
for more info.
CompilationUnit
: this is the owning container for all code objects
associated with a given model. When we load, we load all the code objects to
a single CompilationUnit
.
QualifiedName
: this is the fully qualified name for a code object. It is
similar to qualified names in Python, and looks like "foo.bar.baz"
. Each
code object has a unique QualifiedName
within a CompilationUnit
.
The exporter uses the QualifiedName
of a code object to determine its
location in the code/
folder. The way it does so is similar to how Python
does it; for example, the class Baz
with a QualifiedName
"foo.bar.Baz"
will be placed in code/foo/bar.py
under the name Baz
.
Classes at the root of the hierarchy are given the qualified name __torch__
as a prefix, just so that they can go in __torch__.py
. (Why not __main__
?
Because pickle has weird special rules about things that live in __main__
).
That's about it; there's some additional logic to make sure that within a file, we place the classes in reverse-dependency order so that we compile the "leaf" dependencies before things that depend on them.
A model is really a top-level ScriptModule
with any number of submodules,
parameters, attributes, and so on. We implement a subset of the Pickle format
necessary for pickling a module object.
pickle
's format was chosen due to:
- user friendliness - the attributes file can be loaded in Python with
pickle
- size limits - formats such as Protobuf empose size limits on total message size, whereas pickle limits are on individual values (e.g. strings cannot be longer than 4 GB)
- standard format -
pickle
is a standard Python module with a reasonably simple format. The format is a program to be consumed by a stack machine that is detailed in Python's pickletools.py
- built-in memoization - for shared reference types (e.g. Tensor, string, lists, dicts)
- self describing - a separate definition file is not needed to understand the pickled data
- eager mode save -
torch.save()
already produces apickle
archive, so doing the same with attributes avoids introducing yet another format
All data is written into the data.pkl
file with the exception of tensors
(see the tensor section below).
"Data" means all parts of the module object state, like attributes,
submodules, etc.
PyTorch functions defined in torch/jit/_pickle.py are used to mark special data types, such as this tensor table index or specialized lists.
During export a list of all the tensors in a model is created. Tensors can come from either module parameters or attributes of Tensor type.
Tensors are treated differently from other data (which is pickled using the standard pickling process) for a few reasons:
- Tensors regularly exceed the
pickle
file size limit. - We'd like to be able to
mmap
Tensors directly. - We'd like to maintain compatibility with regular
PyTorch
's serialization format
The pickle
format enforces a separation between data and code, which the
TorchScript serialization process represents by having code/
and
data.pkl + tensors/
.
However, TorchScript inlines constants (i.e. prim::Constant
nodes) directly
into code/
. This poses a problem for tensor constants, which are not easily
representable in string form.
We can't put tensor constants in data.pkl
, because the source code must be
loaded before data.pkl
, and so putting the tensor constants there would
create a cyclic loading dependency.
We solve this problem by creating a separate pickle
file called
constants.pkl
, which holds all tensor constants referenced in code. The
load order will be explained in the next section.
The load process has the following steps:
- Unpickle
constants.pkl
, which produces a tuple of all tensor constants referenced in code. - Unpickle
data.pkl
into the top-levelModule
and return it.
The unpickling process consists of a single call to unpickle the module
object contained in data.pkl
. The Unpickler
is given a callback that lets it
resolve any qualified names it encounters into ClassType
s. This is done by
resolving the qualified name to the appropriate file in code/
, then
compiling that file and returning the appropriate ClassType
.
This is why it's important to give code objects unique qualified names in the
CompilationUnit
. That way, every class that Unpickler
encounters has a
deterministic location in code/
where it is stored.
Unpickler
is also responsible for resolving references to tensors into
actual at::Tensor
s. This is done by looking up offsets in the tensor table
during the unpickling process, (soon to be replaced with the same pickling
strategy as all other data).
Like in Python's pickle
, users can customize the pickling behavior of their
class or module by implementing __getstate__()
and __setstate__()
methods. For basic usage, refer to the relevant Python
docs.
Calls to __getstate__
and __setstate__
are handled transparently by
Pickler
and Unpickler
, so the serialization process shouldn't worry about
it too much.
One thing worth calling out is that the compiler implements a few special
type inference behaviors to cheat the fact that users currently cannot type
annotate Module
s.
__getstate__
and __setstate__
do not require type annotations. For
__getstate__
, the compiler can fully infer the return based on what
attributes the user is returning. Then, __setstate__
simply looks up the
return type of __getstate__
and uses that as its input type.
For example:
class M(torch.nn.Module):
def __init__(self):
self.a = torch.rand(2, 3)
self.b = torch.nn.Linear(10, 10)
def __getstate__(self):
# Compiler infers that this is a tuple of (Tensor, Linear)
return (self.a, self.b)
def __setstate__(self, state):
# Don't need to annotate this, we know what type `state` is!
self.a = state[0]
self.b = state[1]
CompilationUnit
performs two functions:
- It is the owner (in a C++ sense) for all code objects.
- It forms a namespace in which code objects must have unique names.
A CompilationUnit
is created whenever torch::jit::load()
is invoked, to
place the newly deserialized code objects in. In Python, there is a single
global CompilationUnit
that holds all code objects defined in Python.
There are a few different entities that participate in the ownership model:
CompilationUnit
: A container that owns code objects and gives them name.
Every code object has a unique qualified name within the CompilationUnit.
There are two kinds of code objects: Function
s and ClassType
s.
Function
: A Graph
with an associated executor. The Graph
may own
ClassType
s, since some Value
s hold a shared_ptr
to their type (for
now). The Graph
may also weakly reference other Function
s through
function calls.
ClassType
: A definition of a type. This could refer to a user-defined
TorchScript class, or a ScriptModule
. Owns other its attribute types
(including other ClassTypes). Weakly references the class’s methods
(Function
s).
Object
: An instance of a particular class. Own the CompilationUnit
that owns its ClassType
. This is to ensure that if the user passes the
object around in C++, all its code will stay around and methods will be
invokable.
Module
: A view over a ClassType
and the Object
that holds its state.
Also responsible for turning unqualified names (e.g. forward()
) into
qualified ones for lookup in the owning CompilationUnit
(e.g.
__torch__.MyModule.forward
). Owns the Object
, which transitively owns the
CompilationUnit
.
Method
: A tuple of (Module, Function)
.
CompilationUnit
maintains a namespace in which all code objects
(ClassType
s and Function
s) are uniquely named. These names don't have any
particular meaning, except that they uniquely identify a code object during
serialization and deserialization. The basic naming scheme is:
- Everything starts in the
__torch__
namespace. - Classes are named parallel to Python’s module namespacing: so class
Bar
infoo.py
would become__torch__.foo.Bar
. - Methods are attached to the module’s namespace. So
Bar.forward()
would be__torch__.foo.Bar.forward
.
There are some caveats:
Some CompilationUnit
s have no prefix: For testing and other internal
purposes, occasionally it’s useful to have no prefixes on names. In this
case, everything is just a bare name inside the CompilationUnit
. Users
cannot construct `CompilationUnits that look like this.
Name mangling: In Python, we can construct code objects that have the same qualified name. There are two cases where this happens:
- For
ScriptModule
s, since everyScriptModule
is a singleton class in the JIT, a user that is constructing multipleScriptModule
s will create multiple correspondingClassType
s with identical names. - Nesting functions will also cause qualified name clashes, due to
limitations in Python. In these cases, we mangle the names of the code
objects before they are placed in the global Python
CompilationUnit
.
The rules for mangling are simple. Say we have a qualified name __torch__.foo.Bar
:
__torch__.foo.Bar # first time, unchanged
__torch__.foo.__torch_mangle_0.Bar # second time, when we request a mangle
__torch__.foo.__torch_mangle_1.Bar # and so on
Notice that we mangle the namespace before Bar
. This is so that when we
pretty-print code, the unqualified name (Bar
) is unchanged. This is a
useful property so that things like trace-checking are oblivious to the
mangling.