Shortcuts

TorchScript Language Reference

This reference manual describes the syntax and core semantics of the TorchScript language. TorchScript is a statically typed subset of the Python language. This document explains the supported features of Python in TorchScript and also how the language diverges from regular Python. Any features of Python that are not mentioned in this reference manual are not part of TorchScript. TorchScript focuses specifically on the features of Python that are needed to represent neural network models in PyTorch.

Terminology

This document uses the following terminologies:

Pattern

Notes

::=

Indicates that the given symbol is defined as.

" "

Represents real keywords and delimiters that are part of the syntax.

A | B

Indicates either A or B.

( )

Indicates grouping.

[]

Indicates optional.

A+

Indicates a regular expression where term A is repeated at least once.

A*

Indicates a regular expression where term A is repeated zero or more times.

Type System

TorchScript is a statically typed subset of Python. The largest difference between TorchScript and the full Python language is that TorchScript only supports a small set of types that are needed to express neural net models.

TorchScript Types

The TorchScript type system consists of TSType and TSModuleType as defined below.

TSAllType ::= TSType | TSModuleType
TSType    ::= TSMetaType | TSPrimitiveType | TSStructuralType | TSNominalType

TSType represents the majority of TorchScript types that are composable and that can be used in TorchScript type annotations. TSType refers to any of the following:

  • Meta Types, e.g., Any

  • Primitive Types, e.g., int, float, and str

  • Structural Types, e.g., Optional[int] or List[MyClass]

  • Nominal Types (Python classes), e.g., MyClass (user-defined), torch.tensor (built-in)

TSModuleType represents torch.nn.Module and its subclasses. It is treated differently from TSType because its type schema is inferred partly from the object instance and partly from the class definition. As such, instances of a TSModuleType may not follow the same static type schema. TSModuleType cannot be used as a TorchScript type annotation or be composed with TSType for type safety considerations.

Meta Types

Meta types are so abstract that they are more like type constraints than concrete types. Currently TorchScript defines one meta-type, Any, that represents any TorchScript type.

Any Type

The Any type represents any TorchScript type. Any specifies no type constraints, thus there is no type-checking on Any. As such it can be bound to any Python or TorchScript data types (e.g., int, TorchScript tuple, or an arbitrary Python class that is not scripted).

TSMetaType ::= "Any"

Where:

  • Any is the Python class name from the typing module. Therefore, to use the Any type, you must import it from typing (e.g., from typing import Any).

  • Since Any can represent any TorchScript type, the set of operators that are allowed to operate on values of this type on Any is limited.

Operators Supported for Any Type

  • Assignment to data of Any type.

  • Binding to parameter or return of Any type.

  • x is, x is not where x is of Any type.

  • isinstance(x, Type) where x is of Any type.

  • Data of Any type is printable.

  • Data of List[Any] type may be sortable if the data is a list of values of the same type T and that T supports comparison operators.

Compared to Python

Any is the least constrained type in the TorchScript type system. In that sense, it is quite similar to the Object class in Python. However, Any only supports a subset of the operators and methods that are supported by Object.

Design Notes

When we script a PyTorch module, we may encounter data that is not involved in the execution of the script. Nevertheless, it has to be described by a type schema. It is not only cumbersome to describe static types for unused data (in the context of the script), but also may lead to unnecessary scripting failures. Any is introduced to describe the type of the data where precise static types are not necessary for compilation.

Example 1

This example illustrates how Any can be used to allow the second element of the tuple parameter to be of any type. This is possible because x[1] is not involved in any computation that requires knowing its precise type.

import torch

from typing import Tuple
from typing import Any

@torch.jit.export
def inc_first_element(x: Tuple[int, Any]):
    return (x[0]+1, x[1])

m = torch.jit.script(inc_first_element)
print(m((1,2.0)))
print(m((1,(100,200))))

The example above produces the following output:

(2, 2.0)
(2, (100, 200))

The second element of the tuple is of Any type, thus can bind to multiple types. For example, (1, 2.0) binds a float type to Any as in Tuple[int, Any], whereas (1, (100, 200)) binds a tuple to Any in the second invocation.

Example 2

This example illustrates how we can use isinstance to dynamically check the type of the data that is annotated as Any type:

import torch
from typing import Any

def f(a:Any):
    print(a)
    return (isinstance(a, torch.Tensor))

ones = torch.ones([2])
m = torch.jit.script(f)
print(m(ones))

The example above produces the following output:

 1
 1
[ CPUFloatType{2} ]
True

Primitive Types

Primitive TorchScript types are types that represent a single type of value and go with a single pre-defined type name.

TSPrimitiveType ::= "int" | "float" | "double" | "complex" | "bool" | "str" | "None"

Structural Types

Structural types are types that are structurally defined without a user-defined name (unlike nominal types), such as Future[int]. Structural types are composable with any TSType.

TSStructuralType ::=  TSTuple | TSNamedTuple | TSList | TSDict |
                    TSOptional | TSFuture | TSRRef

TSTuple          ::= "Tuple" "[" (TSType ",")* TSType "]"
TSNamedTuple     ::= "namedtuple" "(" (TSType ",")* TSType ")"
TSList           ::= "List" "[" TSType "]"
TSOptional       ::= "Optional" "[" TSType "]"
TSFuture         ::= "Future" "[" TSType "]"
TSRRef           ::= "RRef" "[" TSType "]"
TSDict           ::= "Dict" "[" KeyType "," TSType "]"
KeyType          ::= "str" | "int" | "float" | "bool" | TensorType | "Any"

Where:

  • Tuple, List, Optional, Union, Future, Dict represent Python type class names that are defined in the module typing. To use these type names, you must import them from typing (e.g., from typing import Tuple).

  • namedtuple represents the Python class collections.namedtuple or typing.NamedTuple.

  • Future and RRef represent the Python classes torch.futures and torch.distributed.rpc.

Compared to Python

Apart from being composable with TorchScript types, these TorchScript structural types often support a common subset of the operators and methods of their Python counterparts.

Example 1

This example uses typing.NamedTuple syntax to define a tuple:

import torch
from typing import NamedTuple
from typing import Tuple

class MyTuple(NamedTuple):
    first: int
    second: int

def inc(x: MyTuple) -> Tuple[int, int]:
    return (x.first+1, x.second+1)

t = MyTuple(first=1, second=2)
scripted_inc = torch.jit.script(inc)
print("TorchScript:", scripted_inc(t))

The example above produces the following output:

TorchScript: (2, 3)

Example 2

This example uses collections.namedtuple syntax to define a tuple:

import torch
from typing import NamedTuple
from typing import Tuple
from collections import namedtuple

_AnnotatedNamedTuple = NamedTuple('_NamedTupleAnnotated', [('first', int), ('second', int)])
_UnannotatedNamedTuple = namedtuple('_NamedTupleAnnotated', ['first', 'second'])

def inc(x: _AnnotatedNamedTuple) -> Tuple[int, int]:
    return (x.first+1, x.second+1)

m = torch.jit.script(inc)
print(inc(_UnannotatedNamedTuple(1,2)))

The example above produces the following output:

(2, 3)

Example 3

This example illustrates a common mistake of annotating structural types, i.e., not importing the composite type classes from the typing module:

import torch

# ERROR: Tuple not recognized because not imported from typing
@torch.jit.export
def inc(x: Tuple[int, int]):
    return (x[0]+1, x[1]+1)

m = torch.jit.script(inc)
print(m((1,2)))

Running the above code yields the following scripting error:

File "test-tuple.py", line 5, in <module>
    def inc(x: Tuple[int, int]):
NameError: name 'Tuple' is not defined

The remedy is to add the line from typing import Tuple to the beginning of the code.

Nominal Types

Nominal TorchScript types are Python classes. These types are called nominal because they are declared with a custom name and are compared using class names. Nominal classes are further classified into the following categories:

TSNominalType ::= TSBuiltinClasses | TSCustomClass | TSEnum

Among them, TSCustomClass and TSEnum must be compilable to TorchScript Intermediate Representation (IR). This is enforced by the type-checker.

Built-in Class

Built-in nominal types are Python classes whose semantics are built into the TorchScript system (e.g., tensor types). TorchScript defines the semantics of these built-in nominal types, and often supports only a subset of the methods or attributes of its Python class definition.

TSBuiltinClass ::= TSTensor | "torch.device" | "torch.Stream" | "torch.dtype" |
                   "torch.nn.ModuleList" | "torch.nn.ModuleDict" | ...
TSTensor       ::= "torch.Tensor" | "common.SubTensor" | "common.SubWithTorchFunction" |
                   "torch.nn.parameter.Parameter" | and subclasses of torch.Tensor

Special Note on torch.nn.ModuleList and torch.nn.ModuleDict

Although torch.nn.ModuleList and torch.nn.ModuleDict are defined as a list and dictionary in Python, they behave more like tuples in TorchScript:

  • In TorchScript, instances of torch.nn.ModuleList or torch.nn.ModuleDict are immutable.

  • Code that iterates over torch.nn.ModuleList or torch.nn.ModuleDict is completely unrolled so that elements of torch.nn.ModuleList or keys of torch.nn.ModuleDict can be of different subclasses of torch.nn.Module.

Example

The following example highlights the use of a few built-in Torchscript classes (torch.*):

import torch

@torch.jit.script
class A:
    def __init__(self):
        self.x = torch.rand(3)

    def f(self, y: torch.device):
        return self.x.to(device=y)

def g():
    a = A()
    return a.f(torch.device("cpu"))

script_g = torch.jit.script(g)
print(script_g.graph)

Custom Class

Unlike built-in classes, semantics of custom classes are user-defined and the entire class definition must be compilable to TorchScript IR and subject to TorchScript type-checking rules.

TSClassDef ::= [ "@torch.jit.script" ]
                 "class" ClassName [ "(object)" ]  ":"
                    MethodDefinition |
                [ "@torch.jit.ignore" ] | [ "@torch.jit.unused" ]
                    MethodDefinition

Where:

  • Classes must be new-style classes. Python 3 supports only new-style classes. In Python 2.x, a new-style class is specified by subclassing from the object.

  • Instance data attributes are statically typed, and instance attributes must be declared by assignments inside the __init__() method.

  • Method overloading is not supported (i.e., you cannot have multiple methods with the same method name).

  • MethodDefinition must be compilable to TorchScript IR and adhere to TorchScript’s type-checking rules, (i.e., all methods must be valid TorchScript functions and class attribute definitions must be valid TorchScript statements).

  • torch.jit.ignore and torch.jit.unused can be used to ignore the method or function that is not fully torchscriptable or should be ignored by the compiler.

Compared to Python

TorchScript custom classes are quite limited compared to their Python counterpart. Torchscript custom classes:

  • Do not support class attributes.

  • Do not support subclassing except for subclassing an interface type or object.

  • Do not support method overloading.

  • Must initialize all its instance attributes in __init__(); this is because TorchScript constructs a static schema of the class by inferring attribute types in __init__().

  • Must contain only methods that satisfy TorchScript type-checking rules and are compilable to TorchScript IRs.

Example 1

Python classes can be used in TorchScript if they are annotated with @torch.jit.script, similar to how a TorchScript function would be declared:

@torch.jit.script
class MyClass:
    def __init__(self, x: int):
        self.x = x

    def inc(self, val: int):
        self.x += val

Example 2

A TorchScript custom class type must “declare” all its instance attributes by assignments in __init__(). If an instance attribute is not defined in __init__() but accessed in other methods of the class, the class cannot be compiled as a TorchScript class, as shown in the following example:

import torch

@torch.jit.script
class foo:
    def __init__(self):
        self.y = 1

# ERROR: self.x is not defined in __init__
def assign_x(self):
    self.x = torch.rand(2, 3)

The class will fail to compile and issue the following error:

RuntimeError:
Tried to set nonexistent attribute: x. Did you forget to initialize it in __init__()?:
def assign_x(self):
    self.x = torch.rand(2, 3)
    ~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE

Example 3

In this example, a TorchScript custom class defines a class variable name, which is not allowed:

import torch

@torch.jit.script
class MyClass(object):
    name = "MyClass"
    def __init__(self, x: int):
        self.x = x

def fn(a: MyClass):
    return a.name

It leads to the following compile-time error:

RuntimeError:
'__torch__.MyClass' object has no attribute or method 'name'. Did you forget to initialize an attribute in __init__()?:
    File "test-class2.py", line 10
def fn(a: MyClass):
    return a.name
        ~~~~~~ <--- HERE

Enum Type

Like custom classes, semantics of the enum type are user-defined and the entire class definition must be compilable to TorchScript IR and adhere to TorchScript type-checking rules.

TSEnumDef ::= "class" Identifier "(enum.Enum | TSEnumType)" ":"
               ( MemberIdentifier "=" Value )+
               ( MethodDefinition )*

Where:

  • Value must be a TorchScript literal of type int, float, or str, and must be of the same TorchScript type.

  • TSEnumType is the name of a TorchScript enumerated type. Similar to Python enum, TorchScript allows restricted Enum subclassing, that is, subclassing an enumerated is allowed only if it does not define any members.

Compared to Python

  • TorchScript supports only enum.Enum. It does not support other variations such as enum.IntEnum, enum.Flag, enum.IntFlag, and enum.auto.

  • Values of TorchScript enum members must be of the same type and can only be int, float, or str types, whereas Python enum members can be of any type.

  • Enums containing methods are ignored in TorchScript.

Example 1

The following example defines the class Color as an Enum type:

import torch
from enum import Enum

class Color(Enum):
    RED = 1
    GREEN = 2

def enum_fn(x: Color, y: Color) -> bool:
    if x == Color.RED:
        return True
    return x == y

m = torch.jit.script(enum_fn)

print("Eager: ", enum_fn(Color.RED, Color.GREEN))
print("TorchScript: ", m(Color.RED, Color.GREEN))

Example 2

The following example shows the case of restricted enum subclassing, where BaseColor does not define any member, thus can be subclassed by Color:

import torch
from enum import Enum

class BaseColor(Enum):
    def foo(self):
        pass

class Color(BaseColor):
    RED = 1
    GREEN = 2

def enum_fn(x: Color, y: Color) -> bool:
    if x == Color.RED:
        return True
    return x == y

m = torch.jit.script(enum_fn)

print("TorchScript: ", m(Color.RED, Color.GREEN))
print("Eager: ", enum_fn(Color.RED, Color.GREEN))

TorchScript Module Class

TSModuleType is a special class type that is inferred from object instances that are created outside TorchScript. TSModuleType is named by the Python class of the object instance. The __init__() method of the Python class is not considered a TorchScript method, so it does not have to comply with TorchScript’s type-checking rules.

The type schema of a module instance class is constructed directly from an instance object (created outside the scope of TorchScript) rather than inferred from __init__() like custom classes. It is possible that two objects of the same instance class type follow two different type schemas.

In this sense, TSModuleType is not really a static type. Therefore, for type safety considerations, TSModuleType cannot be used in a TorchScript type annotation or be composed with TSType.

Module Instance Class

TorchScript module type represents the type schema of a user-defined PyTorch module instance. When scripting a PyTorch module, the module object is always created outside TorchScript (i.e., passed in as parameter to forward). The Python module class is treated as a module instance class, so the __init__() method of the Python module class is not subject to the type-checking rules of TorchScript.

TSModuleType ::= "class" Identifier "(torch.nn.Module)" ":"
                    ClassBodyDefinition

Where:

  • forward() and other methods decorated with @torch.jit.export must be compilable to TorchScript IR and subject to TorchScript’s type-checking rules.

Unlike custom classes, only the forward method and other methods decorated with @torch.jit.export of the module type need to be compilable. Most notably, __init__() is not considered a TorchScript method. Consequently, module type constructors cannot be invoked within the scope of TorchScript. Instead, TorchScript module objects are always constructed outside and passed into torch.jit.script(ModuleObj).

Example 1

This example illustrates a few features of module types:

  • The TestModule instance is created outside the scope of TorchScript (i.e., before invoking torch.jit.script).

  • __init__() is not considered a TorchScript method, therefore, it does not have to be annotated and can contain arbitrary Python code. In addition, the __init__() method of an instance class cannot be invoked in TorchScript code. Because TestModule instances are instantiated in Python, in this example, TestModule(2.0) and TestModule(2) create two instances with different types for its data attributes. self.x is of type float for TestModule(2.0), whereas self.y is of type int for TestModule(2.0).

  • TorchScript automatically compiles other methods (e.g., mul()) invoked by methods annotated via @torch.jit.export or forward() methods.

  • Entry-points to a TorchScript program are either forward() of a module type, functions annotated as torch.jit.script, or methods annotated as torch.jit.export.

import torch

class TestModule(torch.nn.Module):
    def __init__(self, v):
        super().__init__()
        self.x = v

    def forward(self, inc: int):
        return self.x + inc

m = torch.jit.script(TestModule(1))
print(f"First instance: {m(3)}")

m = torch.jit.script(TestModule(torch.ones([5])))
print(f"Second instance: {m(3)}")

The example above produces the following output:

First instance: 4
Second instance: tensor([4., 4., 4., 4., 4.])

Example 2

The following example shows an incorrect usage of module type. Specifically, this example invokes the constructor of TestModule inside the scope of TorchScript:

import torch

class TestModule(torch.nn.Module):
    def __init__(self, v):
        super().__init__()
        self.x = v

    def forward(self, x: int):
        return self.x + x

class MyModel:
    def __init__(self, v: int):
        self.val = v

    @torch.jit.export
    def doSomething(self, val: int) -> int:
        # error: should not invoke the constructor of module type
        myModel = TestModule(self.val)
        return myModel(val)

# m = torch.jit.script(MyModel(2)) # Results in below RuntimeError
# RuntimeError: Could not get name of python class object

Type Annotation

Since TorchScript is statically typed, programmers need to annotate types at strategic points of TorchScript code so that every local variable or instance data attribute has a static type, and every function and method has a statically typed signature.

When to Annotate Types

In general, type annotations are only needed in places where static types cannot be automatically inferred (e.g., parameters or sometimes return types to methods or functions). Types of local variables and data attributes are often automatically inferred from their assignment statements. Sometimes an inferred type may be too restrictive, e.g., x being inferred as NoneType through assignment x = None, whereas x is actually used as an Optional. In such cases, type annotations may be needed to overwrite auto inference, e.g., x: Optional[int] = None. Note that it is always safe to type annotate a local variable or data attribute even if its type can be automatically inferred. The annotated type must be congruent with TorchScript’s type-checking.

When a parameter, local variable, or data attribute is not type annotated and its type cannot be automatically inferred, TorchScript assumes it to be a default type of TensorType, List[TensorType], or Dict[str, TensorType].

Annotate Function Signature

Since a parameter may not be automatically inferred from the body of the function (including both functions and methods), they need to be type annotated. Otherwise, they assume the default type TensorType.

TorchScript supports two styles for method and function signature type annotation:

  • Python3-style annotates types directly on the signature. As such, it allows individual parameters to be left unannotated (whose type will be the default type of TensorType), or allows the return type to be left unannotated (whose type will be automatically inferred).

Python3Annotation ::= "def" Identifier [ "(" ParamAnnot* ")" ] [ReturnAnnot] ":"
                            FuncOrMethodBody
ParamAnnot        ::= Identifier [ ":" TSType ] ","
ReturnAnnot       ::= "->" TSType

Note that when using Python3 style, the type self is automatically inferred and should not be annotated.

  • Mypy style annotates types as a comment right below the function/method declaration. In the Mypy style, since parameter names do not appear in the annotation, all parameters have to be annotated.

MyPyAnnotation ::= "# type:" "(" ParamAnnot* ")" [ ReturnAnnot ]
ParamAnnot     ::= TSType ","
ReturnAnnot    ::= "->" TSType

Example 1

In this example:

  • a is not annotated and assumes the default type of TensorType.

  • b is annotated as type int.

  • The return type is not annotated and is automatically inferred as type TensorType (based on the type of the value being returned).

import torch

def f(a, b: int):
    return a+b

m = torch.jit.script(f)
print("TorchScript:", m(torch.ones([6]), 100))

Example 2

The following example uses Mypy style annotation. Note that parameters or return values must be annotated even if some of them assume the default type.

import torch

def f(a, b):
    # type: (torch.Tensor, int) → torch.Tensor
    return a+b

m = torch.jit.script(f)
print("TorchScript:", m(torch.ones([6]), 100))

Annotate Variables and Data Attributes

In general, types of data attributes (including class and instance data attributes) and local variables can be automatically inferred from assignment statements. Sometimes, however, if a variable or attribute is associated with values of different types (e.g., as None or TensorType), then they may need to be explicitly type annotated as a wider type such as Optional[int] or Any.

Local Variables

Local variables can be annotated according to Python3 typing module annotation rules, i.e.,

LocalVarAnnotation ::= Identifier [":" TSType] "=" Expr

In general, types of local variables can be automatically inferred. In some cases, however, you may need to annotate a multi-type for local variables that may be associated with different concrete types. Typical multi-types include Optional[T] and Any.

Example

import torch

def f(a, setVal: bool):
    value: Optional[torch.Tensor] = None
    if setVal:
        value = a
    return value

ones = torch.ones([6])
m = torch.jit.script(f)
print("TorchScript:", m(ones, True), m(ones, False))

Instance Data Attributes

For ModuleType classes, instance data attributes can be annotated according to Python3 typing module annotation rules. Instance data attributes can be annotated (optionally) as final via Final.

"class" ClassIdentifier "(torch.nn.Module):"
InstanceAttrIdentifier ":" ["Final("] TSType [")"]
...

Where:

  • InstanceAttrIdentifier is the name of an instance attribute.

  • Final indicates that the attribute cannot be re-assigned outside of __init__ or overridden in subclasses.

Example

import torch

class MyModule(torch.nn.Module):
    offset_: int

def __init__(self, offset):
    self.offset_ = offset

...

Type Annotation APIs

torch.jit.annotate(T, expr)

This API annotates type T to an expression expr. This is often used when the default type of an expression is not the type intended by the programmer. For instance, an empty list (dictionary) has the default type of List[TensorType] (Dict[TensorType, TensorType]), but sometimes it may be used to initialize a list of some other types. Another common use case is for annotating the return type of tensor.tolist(). Note, however, that it cannot be used to annotate the type of a module attribute in __init__; torch.jit.Attribute should be used for this instead.

Example

In this example, [] is declared as a list of integers via torch.jit.annotate (instead of assuming [] to be the default type of List[TensorType]):

import torch
from typing import List

def g(l: List[int], val: int):
    l.append(val)
    return l

def f(val: int):
    l = g(torch.jit.annotate(List[int], []), val)
    return l

m = torch.jit.script(f)
print("Eager:", f(3))
print("TorchScript:", m(3))

See torch.jit.annotate() for more information.

Type Annotation Appendix

TorchScript Type System Definition

TSAllType       ::= TSType | TSModuleType
TSType          ::= TSMetaType | TSPrimitiveType | TSStructuralType | TSNominalType

TSMetaType      ::= "Any"
TSPrimitiveType ::= "int" | "float" | "double" | "complex" | "bool" | "str" | "None"

TSStructualType ::=  TSTuple | TSNamedTuple | TSList | TSDict |
                     TSOptional | TSFuture | TSRRef
TSTuple         ::= "Tuple" "[" (TSType ",")* TSType "]"
TSNamedTuple    ::= "namedtuple" "(" (TSType ",")* TSType ")"
TSList          ::= "List" "[" TSType "]"
TSOptional      ::= "Optional" "[" TSType "]"
TSFuture        ::= "Future" "[" TSType "]"
TSRRef          ::= "RRef" "[" TSType "]"
TSDict          ::= "Dict" "[" KeyType "," TSType "]"
KeyType         ::= "str" | "int" | "float" | "bool" | TensorType | "Any"

TSNominalType   ::= TSBuiltinClasses | TSCustomClass | TSEnum
TSBuiltinClass  ::= TSTensor | "torch.device" | "torch.stream"|
                    "torch.dtype" | "torch.nn.ModuleList" |
                    "torch.nn.ModuleDict" | ...
TSTensor        ::= "torch.tensor" and subclasses

Unsupported Typing Constructs

TorchScript does not support all features and types of the Python3 typing module. Any functionality from the typing module that is not explicitly specified in this documentation is unsupported. The following table summarizes typing constructs that are either unsupported or supported with restrictions in TorchScript.

Item

Description

typing.Any

In development

typing.NoReturn

Not supported

typing.Union

In development

typing.Callable

Not supported

typing.Literal

Not supported

typing.ClassVar

Not supported

typing.Final

Supported for module attributes, class attribute, and annotations, but not for functions.

typing.AnyStr

Not supported

typing.overload

In development

Type aliases

Not supported

Nominal typing

In development

Structural typing

Not supported

NewType

Not supported

Generics

Not supported

Expressions

The following section describes the grammar of expressions that are supported in TorchScript. It is modeled after the expressions chapter of the Python language reference.

Arithmetic Conversions

There are a number of implicit type conversions that are performed in TorchScript:

  • A Tensor with a float or int data type can be implicitly converted to an instance of FloatType or IntType provided that it has a size of 0, does not have require_grad set to True, and will not require narrowing.

  • Instances of StringType can be implicitly converted to DeviceType.

  • The implicit conversion rules from the two bullet points above can be applied to instances of TupleType to produce instances of ListType with the appropriate contained type.

Explicit conversions can be invoked using the float, int, bool, and str built-in functions that accept primitive data types as arguments and can accept user-defined types if they implement __bool__, __str__, etc.

Atoms

Atoms are the most basic elements of expressions.

atom      ::=  identifier | literal | enclosure
enclosure ::=  parenth_form | list_display | dict_display

Identifiers

The rules that dictate what is a legal identifer in TorchScript are the same as their Python counterparts.

Literals

literal ::=  stringliteral | integer | floatnumber

Evaluation of a literal yields an object of the appropriate type with the specific value (with approximations applied as necessary for floats). Literals are immutable, and multiple evaluations of identical literals may obtain the same object or distinct objects with the same value. stringliteral, integer, and floatnumber are defined in the same way as their Python counterparts.

Parenthesized Forms

parenth_form ::=  '(' [expression_list] ')'

A parenthesized expression list yields whatever the expression list yields. If the list contains at least one comma, it yields a Tuple; otherwise, it yields the single expression inside the expression list. An empty pair of parentheses yields an empty Tuple object (Tuple[]).

List and Dictionary Displays

list_comprehension ::=  expression comp_for
comp_for           ::=  'for' target_list 'in' or_expr
list_display       ::=  '[' [expression_list | list_comprehension] ']'
dict_display       ::=  '{' [key_datum_list | dict_comprehension] '}'
key_datum_list     ::=  key_datum (',' key_datum)*
key_datum          ::=  expression ':' expression
dict_comprehension ::=  key_datum comp_for

Lists and dicts can be constructed by either listing the container contents explicitly or by providing instructions on how to compute them via a set of looping instructions (i.e. a comprehension). A comprehension is semantically equivalent to using a for loop and appending to an ongoing list. Comprehensions implicitly create their own scope to make sure that the items of the target list do not leak into the enclosing scope. In the case that container items are explicitly listed, the expressions in the expression list are evaluated left-to-right. If a key is repeated in a dict_display that has a key_datum_list, the resultant dictionary uses the value from the rightmost datum in the list that uses the repeated key.

Primaries

primary ::=  atom | attributeref | subscription | slicing | call

Attribute References

attributeref ::=  primary '.' identifier

The primary must evaluate to an object of a type that supports attribute references that have an attribute named identifier.

Subscriptions

subscription ::=  primary '[' expression_list ']'

The primary must evaluate to an object that supports subscription.

  • If the primary is a List, Tuple, or str, the expression list must evaluate to an integer or slice.

  • If the primary is a Dict, the expression list must evaluate to an object of the same type as the key type of the Dict.

  • If the primary is a ModuleList, the expression list must be an integer literal.

  • If the primary is a ModuleDict, the expression must be a stringliteral.

Slicings

A slicing selects a range of items in a str, Tuple, List, or Tensor. Slicings may be used as expressions or targets in assignment or del statements.

slicing      ::=  primary '[' slice_list ']'
slice_list   ::=  slice_item (',' slice_item)* [',']
slice_item   ::=  expression | proper_slice
proper_slice ::=  [expression] ':' [expression] [':' [expression] ]

Slicings with more than one slice item in their slice lists can only be used with primaries that evaluate to an object of type Tensor.

Calls

call          ::=  primary '(' argument_list ')'
argument_list ::=  args [',' kwargs] | kwargs
args          ::=  [arg (',' arg)*]
kwargs        ::=  [kwarg (',' kwarg)*]
kwarg         ::=  arg '=' expression
arg           ::=  identifier

The primary must desugar or evaluate to a callable object. All argument expressions are evaluated before the call is attempted.

Power Operator

power ::=  primary ['**' u_expr]

The power operator has the same semantics as the built-in pow function (not supported); it computes its left argument raised to the power of its right argument. It binds more tightly than unary operators on the left, but less tightly than unary operators on the right; i.e. -2 ** -3 == -(2 ** (-3)). The left and right operands can be int, float or Tensor. Scalars are broadcast in the case of scalar-tensor/tensor-scalar exponentiation operations, and tensor-tensor exponentiation is done elementwise without any broadcasting.

Unary and Arithmetic Bitwise Operations

u_expr ::=  power | '-' power | '~' power

The unary - operator yields the negation of its argument. The unary ~ operator yields the bitwise inversion of its argument. - can be used with int, float, and Tensor of int and float. ~ can only be used with int and Tensor of int.

Binary Arithmetic Operations

m_expr ::=  u_expr | m_expr '*' u_expr | m_expr '@' m_expr | m_expr '//' u_expr | m_expr '/' u_expr | m_expr '%' u_expr
a_expr ::=  m_expr | a_expr '+' m_expr | a_expr '-' m_expr

The binary arithmetic operators can operate on Tensor, int, and float. For tensor-tensor ops, both arguments must have the same shape. For scalar-tensor or tensor-scalar ops, the scalar is usually broadcast to the size of the tensor. Division ops can only accept scalars as their right-hand side argument, and do not support broadcasting. The @ operator is for matrix multiplication and only operates on Tensor arguments. The multiplication operator (*) can be used with a list and integer in order to get a result that is the original list repeated a certain number of times.

Shifting Operations

shift_expr ::=  a_expr | shift_expr ( '<<' | '>>' ) a_expr

These operators accept two int arguments, two Tensor arguments, or a Tensor argument and an int or float argument. In all cases, a right shift by n is defined as floor division by pow(2, n), and a left shift by n is defined as multiplication by pow(2, n). When both arguments are Tensors, they must have the same shape. When one is a scalar and the other is a Tensor, the scalar is logically broadcast to match the size of the Tensor.

Binary Bitwise Operations

and_expr ::=  shift_expr | and_expr '&' shift_expr
xor_expr ::=  and_expr | xor_expr '^' and_expr
or_expr  ::=  xor_expr | or_expr '|' xor_expr

The & operator computes the bitwise AND of its arguments, the ^ the bitwise XOR, and the | the bitwise OR. Both operands must be int or Tensor, or the left operand must be Tensor and the right operand must be int. When both operands are Tensor, they must have the same shape. When the right operand is int, and the left operand is Tensor, the right operand is logically broadcast to match the shape of the Tensor.

Comparisons

comparison    ::=  or_expr (comp_operator or_expr)*
comp_operator ::=  '<' | '>' | '==' | '>=' | '<=' | '!=' | 'is' ['not'] | ['not'] 'in'

A comparison yields a boolean value (True or False), or if one of the operands is a Tensor, a boolean Tensor. Comparisons can be chained arbitrarily as long as they do not yield boolean Tensors that have more than one element. a op1 b op2 c ... is equivalent to a op1 b and b op2 c and ....

Value Comparisons

The operators <, >, ==, >=, <=, and != compare the values of two objects. The two objects generally need to be of the same type, unless there is an implicit type conversion available between the objects. User-defined types can be compared if rich comparison methods (e.g., __lt__) are defined on them. Built-in type comparison works like Python:

  • Numbers are compared mathematically.

  • Strings are compared lexicographically.

  • lists, tuples, and dicts can be compared only to other lists, tuples, and dicts of the same type and are compared using the comparison operator of corresponding elements.

Membership Test Operations

The operators in and not in test for membership. x in s evaluates to True if x is a member of s and False otherwise. x not in s is equivalent to not x in s. This operator is supported for lists, dicts, and tuples, and can be used with user-defined types if they implement the __contains__ method.

Identity Comparisons

For all types except int, double, bool, and torch.device, operators is and is not test for the object’s identity; x is y is True if and and only if x and y are the same object. For all other types, is is equivalent to comparing them using ==. x is not y yields the inverse of x is y.

Boolean Operations

or_test  ::=  and_test | or_test 'or' and_test
and_test ::=  not_test | and_test 'and' not_test
not_test ::=  'bool' '(' or_expr ')' | comparison | 'not' not_test

User-defined objects can customize their conversion to bool by implementing a __bool__ method. The operator not yields True if its operand is false, False otherwise. The expression x and y first evaluates x; if it is False, its value (False) is returned; otherwise, y is evaluated and its value is returned (False or True). The expression x or y first evaluates x; if it is True, its value (True) is returned; otherwise, y is evaluated and its value is returned (False or True).

Conditional Expressions

conditional_expression ::=  or_expr ['if' or_test 'else' conditional_expression]
 expression            ::=  conditional_expression

The expression x if c else y first evaluates the condition c rather than x. If c is True, x is evaluated and its value is returned; otherwise, y is evaluated and its value is returned. As with if-statements, x and y must evaluate to a value of the same type.

Expression Lists

expression_list ::=  expression (',' expression)* [',']
starred_item    ::=  '*' primary

A starred item can only appear on the left-hand side of an assignment statement, e.g., a, *b, c = ....

Simple Statements

The following section describes the syntax of simple statements that are supported in TorchScript. It is modeled after the simple statements chapter of the Python language reference.

Expression Statements

expression_stmt    ::=  starred_expression
starred_expression ::=  expression | (starred_item ",")* [starred_item]
starred_item       ::=  assignment_expression | "*" or_expr

Assignment Statements

assignment_stmt ::=  (target_list "=")+ (starred_expression)
target_list     ::=  target ("," target)* [","]
target          ::=  identifier
                     | "(" [target_list] ")"
                     | "[" [target_list] "]"
                     | attributeref
                     | subscription
                     | slicing
                     | "*" target

Augmented Assignment Statements

augmented_assignment_stmt ::= augtarget augop (expression_list)
augtarget                 ::= identifier | attributeref | subscription
augop                     ::= "+=" | "-=" | "*=" | "/=" | "//=" | "%=" |
                              "**="| ">>=" | "<<=" | "&=" | "^=" | "|="

Annotated Assignment Statements

annotated_assignment_stmt ::= augtarget ":" expression
                              ["=" (starred_expression)]

The raise Statement

raise_stmt ::=  "raise" [expression ["from" expression]]

Raise statements in TorchScript do not support try\except\finally.

The assert Statement

assert_stmt ::=  "assert" expression ["," expression]

Assert statements in TorchScript do not support try\except\finally.

The return Statement

return_stmt ::=  "return" [expression_list]

Return statements in TorchScript do not support try\except\finally.

The del Statement

del_stmt ::=  "del" target_list

The pass Statement

pass_stmt ::= "pass"

The print Statement

print_stmt ::= "print" "(" expression  [, expression] [.format{expression_list}] ")"

The break Statement

break_stmt ::= "break"

The continue Statement:

continue_stmt ::= "continue"

Compound Statements

The following section describes the syntax of compound statements that are supported in TorchScript. The section also highlights how Torchscript differs from regular Python statements. It is modeled after the compound statements chapter of the Python language reference.

The if Statement

Torchscript supports both basic if/else and ternary if/else.

Basic if/else Statement

if_stmt ::= "if" assignment_expression ":" suite
            ("elif" assignment_expression ":" suite)
            ["else" ":" suite]

elif statements can repeat for an arbitrary number of times, but it needs to be before else statement.

Ternary if/else Statement

if_stmt ::= return [expression_list] "if" assignment_expression "else" [expression_list]

Example 1

A tensor with 1 dimension is promoted to bool:

import torch

@torch.jit.script
def fn(x: torch.Tensor):
    if x: # The tensor gets promoted to bool
        return True
    return False
print(fn(torch.rand(1)))

The example above produces the following output:

True

Example 2

A tensor with multi dimensions are not promoted to bool:

import torch

# Multi dimensional Tensors error out.

@torch.jit.script
def fn():
    if torch.rand(2):
        print("Tensor is available")

    if torch.rand(4,5,6):
        print("Tensor is available")

print(fn())

Running the above code yields the following RuntimeError.

RuntimeError: The following operation failed in the TorchScript interpreter.
Traceback of TorchScript (most recent call last):
@torch.jit.script
def fn():
    if torch.rand(2):
       ~~~~~~~~~~~~ <--- HERE
        print("Tensor is available")
RuntimeError: Boolean value of Tensor with more than one value is ambiguous

If a conditional variable is annotated as final, either the true or false branch is evaluated depending on the evaluation of the conditional variable.

Example 3

In this example, only the True branch is evaluated, since a is annotated as final and set to True:

import torch

a : torch.jit.final[Bool] = True

if a:
    return torch.empty(2,3)
else:
    return []

The while Statement

while_stmt ::=  "while" assignment_expression ":" suite

while…else statements are not supported in Torchscript. It results in a RuntimeError.

The for-in Statement

for_stmt ::=  "for" target_list "in" expression_list ":" suite
              ["else" ":" suite]

for...else statements are not supported in Torchscript. It results in a RuntimeError.

Example 1

For loops on tuples: these unroll the loop, generating a body for each member of the tuple. The body must type-check correctly for each member.

import torch
from typing import Tuple

@torch.jit.script
def fn():
    tup = (3, torch.ones(4))
    for x in tup:
        print(x)

fn()

The example above produces the following output:

3
 1
 1
 1
 1
[ CPUFloatType{4} ]

Example 2

For loops on lists: for loops over a nn.ModuleList will unroll the body of the loop at compile time, with each member of the module list.

class SubModule(torch.nn.Module):
    def __init__(self):
        super(SubModule, self).__init__()
        self.weight = nn.Parameter(torch.randn(2))

    def forward(self, input):
        return self.weight + input

class MyModule(torch.nn.Module):

    def __init__(self):
        super(MyModule, self).init()
        self.mods = torch.nn.ModuleList([SubModule() for i in range(10)])

    def forward(self, v):
        for module in self.mods:
            v = module(v)
        return v

model = torch.jit.script(MyModule())

The with Statement

The with statement is used to wrap the execution of a block with methods defined by a context manager.

with_stmt ::=  "with" with_item ("," with_item) ":" suite
with_item ::=  expression ["as" target]
  • If a target was included in the with statement, the return value from the context manager’s __enter__() is assigned to it. Unlike python, if an exception caused the suite to be exited, its type, value, and traceback are not passed as arguments to __exit__(). Three None arguments are supplied.

  • try, except, and finally statements are not supported inside with blocks.

  • Exceptions raised within with block cannot be suppressed.

The tuple Statement

tuple_stmt ::= tuple([iterables])
  • Iterable types in TorchScript include Tensors, lists, tuples, dictionaries, strings, torch.nn.ModuleList, and torch.nn.ModuleDict.

  • You cannot convert a List to Tuple by using this built-in function.

Unpacking all outputs into a tuple is covered by:

abc = func() # Function that returns a tuple
a,b = func()

The getattr Statement

getattr_stmt ::= getattr(object, name[, default])
  • Attribute name must be a literal string.

  • Module type object is not supported (e.g., torch._C).

  • Custom class object is not supported (e.g., torch.classes.*).

The hasattr Statement

hasattr_stmt ::= hasattr(object, name)
  • Attribute name must be a literal string.

  • Module type object is not supported (e.g., torch._C).

  • Custom class object is not supported (e.g., torch.classes.*).

The zip Statement

zip_stmt ::= zip(iterable1, iterable2)
  • Arguments must be iterables.

  • Two iterables of same outer container type but different length are supported.

Example 1

Both the iterables must be of the same container type:

a = [1, 2] # List
b = [2, 3, 4] # List
zip(a, b) # works

Example 2

This example fails because the iterables are of different container types:

a = (1, 2) # Tuple
b = [2, 3, 4] # List
zip(a, b) # Runtime error

Running the above code yields the following RuntimeError.

RuntimeError: Can not iterate over a module list or
    tuple with a value that does not have a statically determinable length.

Example 3

Two iterables of the same container Type but different data type is supported:

a = [1.3, 2.4]
b = [2, 3, 4]
zip(a, b) # Works

Iterable types in TorchScript include Tensors, lists, tuples, dictionaries, strings, torch.nn.ModuleList, and torch.nn.ModuleDict.

The enumerate Statement

enumerate_stmt ::= enumerate([iterable])
  • Arguments must be iterables.

  • Iterable types in TorchScript include Tensors, lists, tuples, dictionaries, strings, torch.nn.ModuleList and torch.nn.ModuleDict.

Python Values

Resolution Rules

When given a Python value, TorchScript attempts to resolve it in the following five different ways:

  • Compilable Python Implementation:
    • When a Python value is backed by a Python implementation that can be compiled by TorchScript, TorchScript compiles and uses the underlying Python implementation.

    • Example: torch.jit.Attribute

  • Op Python Wrapper:
    • When a Python value is a wrapper of a native PyTorch op, TorchScript emits the corresponding operator.

    • Example: torch.jit._logging.add_stat_value

  • Python Object Identity Match:
    • For a limited set of torch.* API calls (in the form of Python values) that TorchScript supports, TorchScript attempts to match a Python value against each item in the set.

    • When matched, TorchScript generates a corresponding SugaredValue instance that contains lowering logic for these values.

    • Example: torch.jit.isinstance()

  • Name Match:
    • For Python built-in functions and constants, TorchScript identifies them by name, and creates a corresponding SugaredValue instance that implements their functionality.

    • Example: all()

  • Value Snapshot:
    • For Python values from unrecognized modules, TorchScript attempts to take a snapshot of the value and converts it to a constant in the graph of the function(s) or method(s) that are being compiled.

    • Example: math.pi

Python Built-in Functions Support

TorchScript Support for Python Built-in Functions

Built-in Function

Support Level

Notes

abs()

Partial

Only supports Tensor/Int/Float type inputs. | Doesn’t honor __abs__ override.

all()

Full

any()

Full

ascii()

None

bin()

Partial

Only supports Int type input.

bool()

Partial

Only supports Tensor/Int/Float type inputs.

breakpoint()

None

bytearray()

None

bytes()

None

callable()

None

chr()

Partial

Only ASCII character set is supported.

classmethod()

Full

compile()

None

complex()

None

delattr()

None

dict()

Full

dir()

None

divmod()

Full

enumerate()

Full

eval()

None

exec()

None

filter()

None

float()

Partial

Doesn’t honor __index__ override.

format()

Partial

Manual index specification not supported. | Format type modifier not supported.

frozenset()

None

getattr()

Partial

Attribute name must be string literal.

globals()

None

hasattr()

Partial

Attribute name must be string literal.

hash()

Full

Tensor’s hash is based on identity not numeric value.

hex()

Partial

Only supports Int type input.

id()

Full

Only supports Int type input.

input()

None

int()

Partial

base argument not supported. | Doesn’t honor __index__ override.

isinstance()

Full

torch.jit.isintance provides better support when checking against container types like Dict[str, int].

issubclass()

None

iter()

None

len()

Full

list()

Full

ord()

Partial

Only ASCII character set is supported.

pow()

Full

print()

Partial

separate, end and file arguments are not supported.

property()

None

range()

Full

repr()

None

reversed()

None

round()

Partial

ndigits argument is not supported.

set()

None

setattr()

None

slice()

Full

sorted()

Partial

key argument is not supported.

staticmethod()

Full

str()

Partial

encoding and errors arguments are not supported.

sum()

Full

super()

Partial

It can only be used in nn.Module’s __init__ method.

type()

None

vars()

None

zip()

Full

__import__()

None

Python Built-in Values Support

TorchScript Support for Python Built-in Values

Built-in Value

Support Level

Notes

False

Full

True

Full

None

Full

NotImplemented

None

Ellipsis

Full

torch.* APIs

Remote Procedure Calls

TorchScript supports a subset of RPC APIs that supports running a function on a specified remote worker instead of locally.

Specifically, following APIs are fully supported:

  • torch.distributed.rpc.rpc_sync()
    • rpc_sync() makes a blocking RPC call to run a function on a remote worker. RPC messages are sent and received in parallel to execution of Python code.

    • More details about its usage and examples can be found in rpc_sync().

  • torch.distributed.rpc.rpc_async()
    • rpc_async() makes a non-blocking RPC call to run a function on a remote worker. RPC messages are sent and received in parallel to execution of Python code.

    • More deatils about its usage and examples can be found in rpc_async().

  • torch.distributed.rpc.remote()
    • remote.() executes a remote call on a worker and gets a Remote Reference RRef as the return value.

    • More details about its usage and examples can be found in remote().

Asynchronous Execution

TorchScript enables you to create asynchronous computation tasks to make better use of computation resources. This is done via supporting a list of APIs that are only usable within TorchScript:

  • torch.jit.fork()
    • Creates an asynchronous task executing func and a reference to the value of the result of this execution. Fork will return immediately.

    • Synonymous to torch.jit._fork(), which is only kept for backward compatibility reasons.

    • More deatils about its usage and examples can be found in fork().

  • torch.jit.wait()
    • Forces completion of a torch.jit.Future[T] asynchronous task, returning the result of the task.

    • Synonymous to torch.jit._wait(), which is only kept for backward compatibility reasons.

    • More deatils about its usage and examples can be found in wait().

Type Annotations

TorchScript is statically-typed. It provides and supports a set of utilities to help annotate variables and attributes:

  • torch.jit.annotate()
    • Provides a type hint to TorchScript where Python 3 style type hints do not work well.

    • One common example is to annotate type for expressions like []. [] is treated as List[torch.Tensor] by default. When a different type is needed, you can use this code to hint TorchScript: torch.jit.annotate(List[int], []).

    • More details can be found in annotate()

  • torch.jit.Attribute
    • Common use cases include providing type hint for torch.nn.Module attributes. Because their __init__ methods are not parsed by TorchScript, torch.jit.Attribute should be used instead of torch.jit.annotate in the module’s __init__ methods.

    • More details can be found in Attribute()

  • torch.jit.Final
    • An alias for Python’s typing.Final. torch.jit.Final is kept only for backward compatibility reasons.

Meta Programming

TorchScript provides a set of utilities to facilitate meta programming:

  • torch.jit.is_scripting()
    • Returns a boolean value indicating whether the current program is compiled by torch.jit.script or not.

    • When used in an assert or an if statement, the scope or branch where torch.jit.is_scripting() evaluates to False is not compiled.

    • Its value can be evaluated statically at compile time, thus commonly used in if statements to stop TorchScript from compiling one of the branches.

    • More details and examples can be found in is_scripting()

  • @torch.jit.ignore
    • This decorator indicates to the compiler that a function or method should be ignored and left as a Python function.

    • This allows you to leave code in your model that is not yet TorchScript compatible.

    • If a function decorated by @torch.jit.ignore is called from TorchScript, ignored functions will dispatch the call to the Python interpreter.

    • Models with ignored functions cannot be exported.

    • More details and examples can be found in ignore()

  • @torch.jit.unused
    • This decorator indicates to the compiler that a function or method should be ignored and replaced with the raising of an exception.

    • This allows you to leave code in your model that is not yet TorchScript compatible and still export your model.

    • If a function decorated by @torch.jit.unused is called from TorchScript, a runtime error will be raised.

    • More details and examples can be found in unused()

Type Refinement

  • torch.jit.isinstance()
    • Returns a boolean indicating whether a variable is of the specified type.

    • More deatils about its usage and examples can be found in isinstance().

Docs

Access comprehensive developer documentation for PyTorch

View Docs

Tutorials

Get in-depth tutorials for beginners and advanced developers

View Tutorials

Resources

Find development resources and get your questions answered

View Resources