The JIT can run and optimize PyTorch programs separate from the Python interpreter. This overview is organized into sections that go over different independent components:
- Core Program Representation - The JIT executes TorchScript, a subset of Python. This section describes how TorchScript programs are represented in the JIT, and serves as the interchange format between components of the JIT.
- Generating Programs - TorchScript programs can be created either through tracing Python code or through directly writing TorchScript. This section describes how Models are created from these frontends.
- Executing Programs - Once created, TorchScript models are optimized and run. Since this is a just-in-time compiler, programs are optimized as they are executed, so this section describes both how programs are optimized and how they get run.
- Saving Programs - TorchScript is often created in Python and then used from C++. This section describes how the save and load process works.
- Python Bindings - TorchScript code is normally created and used from Python, so this section describes how the Python components interact with the code in this directory.
For concepts that are actual classes in the JIT, we use capitalized words in code font, e.g. Graph
or Value
.
Sections start with a reference to the source file where the code related to the section resides.
- Core Program Representation
- Generating Programs
- Executing Programs
- Profiling Programs
- Saving Programs
- Testing Programs
- Python Printer
- Python Bindings
At the top level, all TorchScript programs are represented as a Module. Modules contain:
- named Parameters -
Tensors
used in training such asweight
orbias
- named Buffers -
Tensors
that are part of the training state of a module but do not appear in module.parameters() and do not participate in gradient descent. - named sub-Modules - used for code organization.
- named Attributes - all other attributes that are not in the above three categories. Typically used for configuration and are not saved/restored in the modules
state_dict
. - named Methods - functions that can be run on the module such as
forward
This mirrors the nn.Module
objects used in Python. All TorchScript code is a member of some module. This includes pure functions such as those created by annotating a Python function with @torch.jit.script
, which are represented internally as a Module that has a single method forward
that contains the implementation of the function.
Modules contain Parameter objects, which simply hold a "slot" where a Tensor
can be placed. These Tensors
are accessible by the Methods of the Module or the parent Module.
A Method is a piece of TorchScript code that takes a number of arguments and produces an output value. Methods have several subcomponents. A FunctionSchema describes the types and names of the input arguments and return value. A list of member_inputs
describes which Parameters are accessed by the method (this is blank for pure functions). A Graph
object describes the actual code inside the method. The Method also maintains a GraphExecutor which is used to actually execute the Graph
that defines the method.
The Graph
inside the Method is a pure function. The Parameters used by the Method are added as additional inputs to this graph before it is run. This allows the GraphExecutor to treat method inputs and method parameters the same for the purposes of optimization and execution, simplifying the process for executing programs.
Methods also contain helper functions for inserting calls to the Method from other Method objects.
aten/src/ATen/core/function_schema.h
Each Method has a FunctionSchema that describes the Types of the arguments and return values of a function. Operators (builtin primitives that are called by the Interpreter) also have FunctionSchema. FunctionSchema are analogous to a function declaration in C++. They describe how to call the function but do not provide an implementation.
Graphs are the root of the intermediate representation (IR) used to define the implementation of TorchScript functions. If you are familiar with LLVM, they are analogous to an llvm::Function
object. A Graph
is composed of Nodes
, Blocks
, and Values
. Nodes
are instructions (e.g. do a matrix multiply). Nodes
are organized into Blocks
of sequentially executed Nodes
. Each Node
produces a list of output Values
, and also consumes a list of input Values
. As an example, a user may write the following TorchScript code:
@torch.jit.script
def f(a, b):
c = a + b
d = c * c
e = torch.tanh(d * c)
return d + (e + e)
The frontend, described later in this document will turn into a Graph
:
graph(%0 : Double(2),
%1 : Double(2)):
%2 : int = prim::Constant[value=1]()
%3 : Double(2) = aten::add(%0, %1, %2)
%4 : Double(2) = aten::mul(%3, %3)
%5 : Double(2) = aten::mul(%4, %3)
%6 : Double(2) = aten::tanh(%5)
%7 : Double(2) = aten::add(%6, %6, %2)
%8 : Double(2) = aten::add(%5, %7, %2)
return (%8)
This is the canonical textual representation of the IR. You should be able to easily find (almost all) of the elements we discussed above.
graph
is theGraph
%x
areValue
s%x : Double(2)
is a type annotation ofValue
%x
(see below for a list of supported types).%x : T1, %y : T2 = namespace::name(%z, %w)
is aNode
which represents thenamespace::name
operator (this name is usually referred to as theNode
s kind). It takes%z
and%w
Value
s as inputs, and returns two outputs (%x
,%y
) of typesT1
andT2
respectively.
Finally, nodes can have extra pieces of information assigned to them, which are called attributes. You can see that it's used in the prim::Constant
node, which returns the value
attribute when it's called. There's a fixed list of types you can attach:
int64_t
double
Tensor
Graph
(useful for e.g. slicing subgraphs that are meant to be fused)std::string
- and lists of them (not nested)
Graphs in the JIT are in single-static assignment (SSA) form, meaning that each Value
has precisely one defining Node
that can be looked up directly from the Value
(Node* n = v.node()
).
Ownership Model Blocks
, Nodes
, and Values
are owned by the Graph
they appear in and may only appear in a single Graph
. This is enforced by assertions. Creation and deletion of Nodes
is done via Graph
objects (e.g. Graph::create
). Creation and Deletion of Blocks
and Values
is done via Node
objects (e.g. Node::addOutput
, Node::addBlock
). The code also enforces certain consistency properties. For instance, Node::destroy
removes a Node
, but it is only valid to call this function if the Values
produced by this Node
are no longer used, which can be accomplished using other functions such as Value::replaceAllUsesWith
.
Because Graph
owns all its Nodes
, Values
, and Blocks
, these values are always passed around by raw pointer. Generally developers should not write code that holds Value
, Node
, or Block
objects indefinitely without also holding a shared_ptr
to their owning Graph
.
A Node
represents a single built-in operator such as a matrix multiply or a convolution. NodeKind Node::kind()
identifies the operator the Node
represents. Different operators (e.g. conv vs matrix-multiply) are represented by different kinds rather than via subclassing of Node
, as one would find in LLVM. A NodeKind
is a Symbol
object, which is just an interned string inside some namespace. Symbols can be created from strings, e.g. through Symbol::fromQualString("aten::add")
, so there is not a closed set of NodeKind
values that a Node
might have. This design was chosen to allow the open registration of new operators and user-defined operators.
Code in the JIT should always assume the universe of valid
Node
kinds is open and subject to be expanded.
This reflects the reality of the PyTorch operator library where there are already several hundred valid operators.
Nodes produce output Values
and take input Values
as arguments. For instance, a matrix-multiply will take two input Tensors
and produce one output Tensor
. Nodes
can produce multiple outputs. For instance prim::TupleUnpack
splits a tuple into its components, so it has a number of outputs equal to the number of members of the tuple. Though Nodes
may have multiple outputs, the number of outputs is statically known for each Node
. Operations which may produce a dynamic amount of results, e.g. splitting a Tensor
into chunks of size 2, will be represented as an operator that results a list object.
Because Nodes
are not subclassed per-operator, it is very easy to construct invalid Nodes
, e.g. by forgetting an input or an output, or by passing Values
of the wrong Type. To help avoid this, Graph
provides the method Graph::insert
for constructing Nodes
that guarantees Nodes
have the correct setup. This method uses the database of registered Operators and their FunctionSchema to construct Nodes
using that schema.
PyTorch IR supports function overloading, which means that a single NodeKind may correspond to multiple operators. For example, the kind aten::add
has the following overloads (Scalar
means float
or int
in this case):
aten::add(Tensor self, Tensor other) -> Tensor
aten::add(Tensor self, Scalar other) -> Tensor
aten::add(int self, int other) -> int
aten::add(float self, float other) -> float
For Nodes
representing built-in Operators, the method Node::schema
can also look up the FunctionSchema registered for that Operator.
Each overload corresponds to a different FunctionSchema
object. A Node
can be queried for its schema using the schema()
method (it will check the argument types, and will try to match one of the options for its kind()
).
Note that the chosen overload is not shown in any way in the textual output. If you're unsure which function a node resolves to, you might need to check the type annotations of its input values.
Each node also has a set of attributes which are named integers, strings, floats, Tensors
, subgraphs, or lists of these types. These are used by special primitive operators to encode additional data in the Node
. For instance prim::Constant
defines a compile-time constant value. For Tensor
constants, it will have a single Tensor
attribute with the name attr::value
which contains the value of the constant.
Attributes are rarely used. Operators like convolution or matrix-multiply have no attributes and take their arguments through the input list. This includes things that might be typically thought of as constants, like the stride of the convolution. In PyTorch, any of this information is potentially a dynamic property of the program so Nodes
are always encoded in a way that allows these values to be dynamically determined. However, we recognize that many inputs are almost always constants, so we make it easy to quickly check if an input is constant and get its value with c10::optional<IValue> Node::get(Symbol name)
, which returns an IValue
(a concrete value for the input) in the case the node is constant and nullopt
otherwise.
Nodes are organized into sequentially executed lists inside a Block
. A Node
is a member of precisely one Block
. The Graph
itself has a top-level Graph::block()
, and control-flow nodes (prim::If
and prim::Loop
) also have sub-blocks. While it is possible to design a Graph
representation that does not have a sequential order for nodes (i.e. a sea-of-nodes representation), we find it is much easier to debug and understand Blocks
when there is a specific canonical order for all of the nodes. This does not preclude optimization passes from changing the order when it would improve performance, and the interpreter is allowed to execute Nodes
out-of-order if the re-ordering preserves the semantics (much like an out-of-order processor does). Having the ordering ensures that graphs can always be easily printed, and that we can easily step through the execution of a graph.
Values are Block-scoped. A Value
is in scope for the remainder of the Block
it is defined in, including in the sub-blocks of any Node
defined after it. Values
go out of scope at the end of the block in which they are defined.
When Nodes
are inserted into a Graph
, they are inserted at a special "insertion point" that is part of the state of the Graph
. On construction, this will go to the end of the Graph
.
Each Block
has two dummy nodes that are not included in the list of nodes in the Block
. The prim::Param
node represents the inputs to the Block
and does not have a prev()
or next()
Node
. The prim::Return
Node
represents the outputs of a Block
.
The list of Nodes
in a Block
is implemented as a circular linked list with the prim::Return
Node
serving as the beginning/end sentinel. Inserting and deleting at arbitrary places is efficient. Developers may also encounter implementations inside of IR objects that use this fact (e.g. appending to a Block
is equivalent to putting the node before the prim::Return
Node).
Iterators for the Block::nodes()
list are invalided when the current Node
they point to is moved or deleted. Otherwise iterators remain valid.
Blocks also contain a list of input and output values. The meaning of these values depends on where the Block
is used. For the Graph's top-level Block
, these are inputs and outputs to the Graph
, and line up with the FunctionSchema associated with the Method that owns the Graph
.
Control-flow is represented using sub-Blocks rather than a control-flow graph representation. A prim::If
has one Block
for the true branch and one Block
for the else. A prim:Loop
has a block for the loop body (there is no condition block, instead the end of the loop body computes whether to re-enter the loop body). This representation ensures we have structured control-flow. This limitation makes a lot of optimizations easier and is true for the vast majority of networks. A Node
can look up what Block
it is in, and a Block
and can look up its parent (either the Node
that has it as a sub-Block, or nullptr
for the main Block).
If-statement (prim::If
) Blocks
have no inputs, and the outputs are the new values of variables in the outer block whose values were altered in the if-statement.
Example IR for an if-statement looks like:
%y_1, ..., %y_r = prim::If(%condition)
block0(): # TRUE BRANCH, never takes arguments, has to return r outputs
%t_1, ..., %t_k = some::node(%a_value_from_outer_block)
-> (%t_1, ..., %t_r)
block1(): # FALSE BRANCH, never takes arguments, has to return r outputs
%f_1, ..., %f_m = some::node(%a_value_from_outer_block)
-> (%f_1, ..., %f_r)
Values corresponding to %y_1, ..., %y_r
will become either %t_1, ..., %t_r
, or %f_1, ..., %f_r
depending on the value of %condition
at runtime.
Here's an example translation of a Python program and its corresponding IR:
def f(a, b, c):
d = a + b
if c:
e = d + d
else:
e = b + d
return e
graph(%a : Dynamic,
%b : Dynamic,
%c : Dynamic):
%2 : int = prim::Constant[value=1]()
%3 : Dynamic = aten::add(%a, %b, %2)
%5 : Dynamic = prim::If(%c)
block0():
%6 : int = prim::Constant[value=1]()
%7 : Dynamic = aten::add(%3, %3, %6)
-> (%7)
}
block1():
%8 : int = prim::Constant[value=1]()
%9 : Dynamic = aten::add(%b, %3, %8)
-> (%9)
return (%5)
The outputs of the if-statement serve a role similar to that of a Φ (Phi) function node in traditional SSA control-flow graphs.
Loops are implemented with prim::Loop
which covers both while
and for
loops. A valid instantiation of this node always looks like this:
%y_1, ..., %y_r = prim::Loop(%max_trip_count, %initial_condition, %x_1, ..., %x_r)
block0(%i, %a_1, ..., %a_r):
%b_1, ..., %b_m = some::node(%a_value_from_outer_block, %a_1)
%iter_condition = some::other_node(%a_2)
-> (%iter_condition, %b_1, ..., %b_r)
The simplest way to explain the semantics is to consider this Python-like pseudo-code:
y_1, ..., y_r = x_1, ..., x_r
condition = initial_condition
i = 0
while condition and i < max_trip_count:
a_1, ..., a_r = y_1, ..., y_r
############################################################
# Actual body of the loop
b_1, ..., b_m = some::node(a_value_from_outside_of_the_loop, a_1)
iter_condition = some::node(a_2)
############################################################
y_1, ..., y_r = b_1, ..., b_r
condition = iter_condition
i += 1
Note that translations of
for
loops simply pass in a constanttrue
for both%initial_condition
and%iter_condition
. Translations ofwhile
loops set%max_trip_count
to the largest value ofint64_t
and do not use%i
. Those patterns are recognized by our interpreter and optimized accordingly (e.g.while
loops don't maintain the loop counter).
For example, this program:
def f(x):
z = x
for i in range(x.size(0)):
z = z * z
return z
can be translated as:
graph(%z.1 : Dynamic):
%3 : bool = prim::Constant[value=1]()
%1 : int = prim::Constant[value=0]()
%2 : int = aten::size(%z.1, %1)
%z : Dynamic = prim::Loop(%2, %3, %z.1)
block0(%i : int, %5 : Dynamic):
%z.2 : Dynamic = aten::mul(%5, %5)
-> (%3, %z.2)
return (%z)
With-statements are represented in two different ways. For most of the compilation and optimization process, they are represented as a pair of prim::Enter
and prim::Exit
nodes that wrap the nodes corresponding to the body of the with-statement. However, with-statements are temporarily represented for the duration of the exit_transform
pass using a block-based representation in which a prim::With
node is inserted after the prim::Exit
node, all of the nodes between the prim::Exit
and prim::Enter
are moved into the first block of the prim::With
, and the prim::Exit
is moved into the second block of the prim::With
. For example, this program:
with c as increment:
y = x + increment
can be translated as:
%2 : int = prim::Constant[value=1]()
%increment.1 : int = prim::Enter(%c.1)
%y.1 : Tensor = aten::add(%x.1, %increment.1, %2)
%11 : Tensor = prim::Exit(%c.1)
and will temporarily be transformed to:
%increment.1 : int = prim::Enter(%c.1)
= prim::With()
block0():
%y.1 : Tensor = aten::add(%x.1, %increment.1, %4)
-> ()
block1():
%11 : Tensor = prim::Exit(%c.1)
-> ()
for the duration of the exit_transform
pass.
A Value
represents data flowing through the operations in the program, e.g. the output of a matrix-multiply op. Value
objects are always defined by a single Node
(v.node()
) due to single-static assignment form. For inputs to a Block/Graph, this node is a special prim::Param
node that does not appear anywhere in the block's list of nodes. Value
objects also have a Type (e.g. is it a Tensor? a list? a tuple?) that provides a static guarantee that its value will be of that Type.
A Value
object has methods that return its definition (v.node()
) and to all of its uses (v.uses()
). Each Use has a pointer to the Node
whose input list includes the value. Be careful when iterating over v.uses()
while changing how v
is used because each change to v
will invalidate the v.uses()
iterator.
Values are abstract representations of data in the program. When executing, the actual Tensors
, list, tuples, etc. are stored in IValues
(interpreter values), which are tagged unions of all possible value types in TorchScript. In retrospect the name Value
is a bit confusing because it seems like it should be the tagged union, but it originally came from analogy to llvm::Value
, which serves the same purpose as jit::Value
.
TorchScript, unlike Python, is statically typed, so every Value
has a Type associated with it, and every FunctionSchema has a list of argument types and a return type for a function. Type is the base class of a hierarchy of C++ objects that represent the built-in types of TorchScript. Types provide methods such as Type::isSubtypeOf
that describe the typing relationships. Common type are:
- TensorType - a
Tensor
with optionally refined information. It may know its device, type, requires_grad state, and number of dimensions. If it does know the number of dimensions it may know the size of a particular dimension. - Tuples - e.g.
Tuple[Tensor, Int]
. Each member of the tuple is statically typed and the length of the tuple is statically known. - List[T] - e.g.
List[Tensor]
. Mutable lists of a particular type. - Optional[T] - e.g.
Optional[Tensor]
. Either a value of type T or None. - Dict[K, V] - e.g.
Dict[String, Tensor]
. Dictionaries.
If type S is a subtype of P, then we can substitute an IValue
that has type S anywhere something of type P is expected. This means that all subtyping relationships also require the representation of the IValue
for subtypes to be compatible with the representation for the base type.
JIT programs are created using either the tracing frontend (torch.jit.trace
) or the scripting frontend (torch.jit.script
). In both cases, the result of these frontends is a complete Module that contains all the code in Methods, and all the model weights in the Parameters of the Module. However, each frontend goes through a different pathway for generating those Modules.
The tracer produces graphs by recording what actual operations are done on Tensors
.
The entry point from Python into C++ for tracing using torch.jit.trace
is _create_method_from_trace
.
A thread local instance of the TracingState object maintains a mapping between actual data being computed during the trace (e.g. Tensors) stored in IValues
, and the abstract Value
in the Graph
that would compute each value. The functions void setValueTrace(const IValue&, Value*)
and Value* getValueTrace(const IValue&)
are used by the tracer to maintain this mapping.
An initial IValue
to Value
mapping is set up between the inputs to the function being traced and symbolic Value
inputs to the Graph
being constructed. If we are tracing a torch.nn.Module
, the tracer also adds Parameters and sub-Modules to the Module being constructed that correspond to the Python torch.nn.Module
being traced. Mappings for these values are also added so that uses of the Parameters in the trace will create uses of the Parameters in the Graph
.
As the trace runs, individual operators create Nodes
in the Graph
being traced to record what happens. This code is currently generated per operator in tools/autograd/gen_variable_type.py. It results in code that looks like the following:
torch::jit::Node* node = nullptr;
std::shared_ptr<jit::tracer::TracingState> tracer_state;
if (jit::tracer::isTracing()) {
tracer_state = jit::tracer::getTracingState();
at::Symbol op_name;
op_name = jit::Symbol::fromQualString("aten::__ilshift__");
node = tracer_state->graph->create(op_name, /*num_outputs=*/0);
jit::tracer::recordSourceLocation(node);
jit::tracer::addInputs(node, "self", self);
jit::tracer::addInputs(node, "other", other);
tracer_state->graph->insertNode(node);
jit::tracer::setTracingState(nullptr);
}
TypeDefault::__ilshift__(self, other);
if (tracer_state) {
jit::tracer::setTracingState(std::move(tracer_state));
jit::tracer::addOutput(node, self);
}
The functions addInputs
and addOutput
are overloaded to handle the different data types that operators use.
set/getValueTrace only works on Tensors
and Futures
. Other types are not natively traced. Instead aggregates like tuples or lists are often flattened into Tensors
at the end of a trace and explicitly constructed from individual Tensors
at the beginning of this trace.
The tracer has special behavior when tracing calls to other TorchScript functions. This behavior is implemented in the GraphExecutor right before a Graph
is about to be run. If tracing is enabled while running the graph, the GraphExecutor will disable tracing, run the graph as normal, and then inline the Graph
into the trace. It then hooks up the IValues
computed by running the Graph
to out Values
in the inlined graph.
When a trace calls a TorchScript function, that function is preserved as is, meaning that control-flow is preserved. This makes it possible to work around tracing issues by generating the subset of the program that cannot be traced using the script frontend and having the trace invoke it.
The resulting Graph
created by tracing is installed as the 'forward' method of the Module being created. A Module is produced regardless of whether the thing being traced was a function or a torch.nn.Module
. In the function case, the Module produced will simply have a single forward
function, no Parameters, and no sub-Modules.
The script frontend directly converts Python syntax into Modules. Like many compilers this happens in two phases. First, we generate an abstract syntax tree (AST), which is constructed out of Tree objects. The IR emitter then does semantic analysis on the Tree and lowers it into a Module. We can generate Trees in two ways: (1) using frontend.py, which takes the Python AST and transliterates it into Tree objects, or (2) via the Lexer and Parser which parse Python syntax directly. The Lexer+Parser path may seem redundant but it is crucially important. We need to define builtin functions (frontend/builtin_functions.cpp) when Python is not linked because we allow users to generate TorchScript programs directly from strings containing Python source code (api/include/torch/jit.h) without linking a full Python implementation (e.g. CPython). We also use this Python syntax as the serialization format for TorchScript, since it allows us to make changes to our IR without breaking backward compatibility. Furthermore, the Lexer is reused to implement the FunctionSchema parser, which turns FunctionSchema declarations from strings into FunctionSchema objects.
The following sections look into each the stages in the script frontend in detail.
Our frontends produce ASTs in the form of Tree objects. Trees are similar to s-expressions. Leafs (i.e. Atoms) are always strings. Compound trees have a kind
(e.g TK_CONST
or TK_IDENT
defined in lexer.h) and a list of sub-trees. For instance, the Tree for z.sigmoid() - (x + y)
is:
(-
(+
(variable (ident x))
(variable (ident y)))
(apply
(.
(variable (ident z))
(ident sigmoid))
(list)
(list))))
This is printed in s-expression style with (kind ...)
representing compound trees and string_value
representing strings.
We provide utilities to construct, traverse, and print ASTs without a lot of complicated visitor infrastructure and inheritance.
Each Tree also has a mandatory SourceRange object that describes the range of text that it came from. These will be used for error reporting in the rest of the code.
Trees are easy to construct, visualize and traverse, but extracting information from a large compound tree like that of a function definition is unwieldy since it requires numeric indexing. Tree Views are a small layer on top of a tree that make it possible to create and de-structure trees of particular kinds. For example, here is the tree view for the apply node which provides named accessors for its subtrees: the function being called, the inputs, and the attributes (i.e. kwargs):
struct Apply : public Expr {
Expr callee() const {
return Expr(subtree(0));
}
List<Expr> inputs() const {
return List<Expr>(subtree(1));
}
List<Attribute> attributes() const {
return List<Attribute>(subtree(2));
...
};
The typical way to traverse a tree is to switch
on the kind and then construct the appropriate Tree view:
switch (tree.kind()) {
case TK_VAR:
auto var = Var(tree); // construct tree view
return environment_stack->getSugaredVar(var.name());
case '.': {
auto select = Select(tree); // construct tree view
auto sv = emitSugaredExpr(select.value(), 1);
return sv->attr(select.range(), method, select.selector().name());
}
case TK_APPLY: {
auto apply = Apply(tree); // construct tree view
return emitApplyExpr(apply, n_binders);
} break;
One way we construct Tree objects is directly from Python ASTs. This logic is contained inside frontend.py and is intentionally very minimal.
We endeavor to keep most of the JIT code written in C++, because most of the JIT functionality still needs to work without Python installed.
This code simply constructs the Tree, filtering out the AST nodes of Python that we do not support.
When loading TorchScript code directly from a string, we using a standard Lexer+Parser combo. The Lexer takes an initial string and then exposes a stateful interface for walking the Tokens of the string, providing a standard set of functions:
next()
advances the lexer, returning the current tokencur()
provides the current tokenlookahead()
provides the token coming after the current tokennextIf(int token_kind)
advances the token if it matches token kind.
Similar to Python, the Lexer handles the white-space sensitive nature of Python blocks. The Tokens TK_INDENT
, TK_DEDENT
, and TK_NEWLINE
are injected into the token stream when code first becomes indented, when it dedents, and at the end of a statement. For instance for this stream:
if
.
.
We would get a token stream TK_IF TK_NEWLINE TK_INDENT . TK_NEWLINE . TK_NEWLINE TK_DEDENT
. Unmatched opening brackets disable the injection of these tokens. The result is that the Parser can simply treat TK_INDENT
, TK_DEDENT
and TK_NEWLINE
like C's {
, }
, and ;
.
Tokens are either keywords (def
), operators (+
), literals (3.4
), or identifiers (foo
). A token_kind
integer identifies what it is and is the exact same type as the kind
of a Tree. For single-character Tokens (e.g. +
), the kind is the same as the character, enable statements like:
if (lexer.nextIf('+')) {
// handle + ...
}
Multi-character token kinds are defined in a list, TC_FORALL_TOKEN_KINDS
. Tokens also have a text()
field that records the actual string producing the token and is used by identifiers and literals to construct the actual values (e.g. the numeric value of a floating point literal).
The Parser uses the Lexer to build the AST for function definitions. parseFunction
is the entrypoint for parsing a single def ...
and will return a Def
tree view.
The Parser is written as a top-down precedence parser, or "Pratt" parser. They are simpler and easier to understand than typical parser generators, while still being flexible enough to parse programming languages. For the most part parsing is done by recursive decent. To resolve operator precedence issues, the function to parse an expression is augmented with a precedent p such that calling the function means parse an expression whose operators all have precedence higher than p.
The file ir_emitter.cpp translates Trees into Modules. The main entrypoint is defineMethodsInModule
which takes a list of Def Tree Views representing function definitions and adds them as Methods to the module. During the lowering processing semantic checking occurs. The IR emitter checks that all used variables are defined (sometimes called scope checking), and that all values have compatible types (type-checking). During this process it also emits the graph nodes corresponding to each statement in the Tree and generates a FunctionSchema for the whole definition.
A few helper objects exist in the lowering process. SugaredValues are special values that represent objects that can appear during compilation but that are not first class values. For instance, in TorchScript methods self
refers to the module, and self.weight
refers to a Parameter of the module. Neither are first-class Types and have no corresponding Value
in a graph. Resolver objects are std::functions
that resolve externally-defined variables to SugaredValues. For instance, the identifier torch
which contains most of our built-in ops is looked up through Resolver objects which interact with the Python state of the program.
The Environment tracks the mapping between variable names and the SugaredValues they refer to.
SugaredValues are how the IR emitter represents non-first class values during Graph
creation. These values are things like the Module or a Python function call that do not have corresponding Value
objects in the Graph
. The IR emitter desugars the SugaredValue objects to instructions in the graph based on how they are used. The SugaredValue class has a number of abstract methods on it such as attr
or call
. Consider the expression self.foo
. For methods, self
will resolve to a special SugaredValue subclass, ModuleValue. When the emitter sees self.foo
, it will then call the ModuleValue function sv.attr("foo")
, asking the ModuleValue how it should desugar itself when the attribute "foo"
accessed. If foo
is a parameter, it would then ensure that the parameter was added to the Method being compiled, and return a SimpleValue
sugared value that contains the Value
object representing the parameter as an input. If foo
were a sub-Module then it would return another SugaredModule. The method call
is invoked when the emitter sees the value used as a function call.
SugaredValues are also how we interact with Python runtime during the compilation process. For instance, math.pi
is resolved to 3.1415... by first resolving math
to a SugaredValue representing accesses to Python modules (PythonModuleValue) whose attr
function turns Python numbers into prim::Constant
Nodes
in the graph.
Finally, normal Values
are also represented by the SimpleValue SugaredValue in places where it is valid that either a SugaredValue or a normal Value
will appear.
Any undefined variable during compilation is resolved with a call to an externally-provided Resolver. When called from Python (e.g torch.jit.script
) this resolver interacts with the Python runtime via pybind11 to resolve symbols like torch
and math
to their Python equivalents.
The combination of SugaredValue and Resolver decouples the implementation of the IR emitter from the pybind11 Python bindings that enable its interaction with the Python state.
This makes it possible to use most of the IR emitter functionality when Python is not present.
The Environment object tracks the assignment of variable names during compilation. It is local to the IR emitter file. A stack of environments exist, with a new environment being created for sub-blocks introduced by control flow. The Environment keeps two tables, one for values which are not first class in the type system (SugaredValues) and a type table for values which are. When first class values are set, we emit a prim::Store, and when they are referenced we emit a prim::Load. SugaredValues are not re-assignable.
As explained in the Block section, the IR is represented in structured control flow composed of ifs & loops. This makes it easier to optimize and lower to other compilers which do not support unstructured control flow. We lower Python control flow (break, continue, return) to this simplified form. We do closing over any variables in the environment, so we are able to convert all writes and reads from the environment directly to SSA form.
Conversion to SSA works in multiple parts.
- First, we add loads and stores to control flow operators (ifs & loops).
- Then we erase break & continue statements from the graph and replace them with
prim::LoopContinuation(%loop_continue_condition, %loop_carried_vars)
. Break statements have the continue condition set to false, and continue statements inline the loop condition. %loop_carried_vars are the loop carried variables of the inner most loop that contains the break or continue statement, are added by inserting prim::Load calls at the location of the statement. - Then we inline the loop condition into the graph loops.
- Next we erase loads and stores, removing all stores and replacing all loads with whatever the in-scope value of the variable name is.
- Finally, we remove
prim::LoopContinuation
s andprim::ReturnStmt
s in the exit_transform pass.
This pass takes in a graph where LoopContinuation & ReturnStmts exist in the graph and erases them, correctly setting block outputs. prim::LoopContinuation(*vals)
means that the values are targeting the most recent loop block. prim::ReturnStmt(*vals)
means that the values are targeting the most recent Closure
or Graph
Block
.
If a block has an exit node, no further instructions will be executed until the exit target has been reached. If we encounter a node that contains nested blocks that may have hit an exit node, such as an if statement that exits in one block and does not exit in the other, we use a boolean value to indicate if the exit has been hit or not. Then, we conditionalize further execution.
Python example:
while i < 5:
if i == 3:
i += 1
continue
i += 2
-> transforms to
continue_loop = i < 5
while continue_loop:
if i == 3:
i = i + 1
continue_loop = i < 5
did_exit = True
if did_exit:
pass
else:
i = i + 2
continue_loop = i < 5
The pass also keeps track of nodes or blocks that will always throw Exceptions so that we do not unnecessarily conditionalize execution. In the following example, we can treat the if statement as always Returning and remove the print
statement.
if i < 0:
raise Exception("Negative input")
else:
return math.sqrt(i)
print(i) # unreachable code
In the above example, the if statement will have one output: math.sqrt(i)
on the false branch, and prim::Uninitialized
in the true branch. prim::Uninitialized
is inserted by the compiler when it can prove the value will never be used. It can be introduced by exceptions, breaks, continues, and returns.
We initially considered doing the Transform pass before Loads and Stores were removed from the graph. However, this breaks when a loop carried variable
is captured in a break or continue and then is refined in the rest of the loop body. In the below example, at the point of the continue
, x
has type Optional[int]
but is refined to int
after the continue statement.
...
if cond:
if i < 3:
x = torch.jit.annotate(Optional[int], None)
continue
x = 1
else:
x = 2
print(x)
If we were to rearrange the graph before loads & stores were removed:
if cond:
if i < 3:
x = torch.jit.annotate(Optional[int], None)
did_continue = True
continue
else:
did_continue = False
if not did_continue:
x = 1
else:
x = 2
if not did_continue:
print(x)
The type of x
at the print statement would be Optional[int]
, which breaks its original type.
A set of special SugaredValues are used to translate between objects in the Python environment and Values
in the Graph
during the compilation process. The entry-point for this behavior is toSugaredValue(py::object obj, ...)
which takes a pybind11 Python value and figures out how to turn it into an appropriate SugaredValue. Values
exist to represent Python functions, Python modules, and ScriptModule
objects.
TorchScript is executed using an interpreter attached to a JIT-optimizer and compiler. The entry-point for execution is the GraphExecutor object that is created on demand inside a Method when the method is first called. This section first goes over the semantics of graphs, i.e. what does it mean to execute a graph? And then details how the implementation works.
TorchScript programs implement a very small subset of Python that is necessary to run models.
TorchScript includes immutable value types:
int
float
Tuple[T0, T1, ...]
As well as mutable reference types:
Tensor
List[T]
Dict[K, V]
A value of a reference type points to an underlying memory location where the data for the reference type is stored, and variable assignment for a reference type can cause multiple values to point to the same underlying data. This is similar to Python's class model.
It is important to remember that TorchScript uses these semantics for Tensors
so not all computation on Tensor
is pure. Individual Tensors
may be views of the same underlying data. Views are established by special view creating operations, such as indexing into a Tensor:
t = torch.rand(3, 4)
t2 = t[0] # view of one slice of t
Some builtin operators also mutably write to the underlying Tensor
. In the standard library these operators are always named with a trailing underscore, or take a named out
Tensor
where the result is written:
t2.relu_() # inplace relu operator, note t is modified as well!
torch.add(t, t, out=t) # update t, without using temporary memory if possible
The combination of reference semantics and mutable operators can be more difficult to optimize, but it gives program writers powerful control of the memory usage of their programs. For instance, DenseNets use a concat operation instead of the addition found in a ResNet. Rather than compute a concat of existing Tensors
, many implementations use Tensor
indexing and out
keywords to avoid allocating addition memory for the activations. Ideally a compiler would always be able to do these optimizations, but in practice new ideas are tried all the time that exist outside what compiler writers expect.
In addition to being mutable, Tensors
also have a set of dynamically determined properties (i.e. properties that can vary from run to run) this includes:
- dtype - their data type int, float, double, etc.
- device - where the
Tensor
lives, e.g. the CPU, or CUDA GPU 0 - rank - the number of dimensions that the
Tensor
has - size - the precise size of the Tensor
- requires_grad - whether the
Tensor
is recording its gradient with autograd
Changes in these properties change how operators on Tensors
will evaluate and would make certain optimization invalid. For instance, if we have a fuser capable of generating new CUDA kernels but not CPU kernels, it is only valid to fuse operations where the inputs are known to run only on CUDA devices. The GraphExecutor's job is to still enable optimization even when certain combinations of properties prevent optimizations from occurring.
Nodes in a graph are executed serially in the order they appear in a block. Nodes
may be reordered either during optimization or by the interpreter itself if it can be proven that the new order
is not distinguishable from the original execution order. These semantics are necessary since the combination of mutable Tensors
and potential aliases between Tensors
makes it unsafe to perform arbitrary reordering otherwise. However, the AliasInfo object can accurately track how alias propagate through builtin operators so optimization passes can query when certain reorders or optimizations are safe.
We also provide user-accessible parallel execution through the fork
and wait
primitives. The fork
primitive begins execution of fn
in parallel with the current thread of execution, immediately returning a Future object that will hold the result of the forked function. The wait
method of the future then causes the invoking thread to wait for the value being computed by fn
.
def fn(arg0, arg1, ...):
...
return v
fut = torch.jit.fork(fn, arg0, arg1, ...)
...
v = torch.jit.wait(fut)
Currently, the user is responsible for avoiding races between threads. We encourage users to not write to Tensors
visible from other threads, and may enforce this more strictly in the future.
Optimization passes that wish to exploit multi-threaded execution may automatically convert serial Blocks
into parallel execution by inserting extra fork and wait events. This design enables our users to manually specify parallelism while also allowing optimization passes to exploit it when safe and profitable.
All evaluation involves computation using IValues
, 16-byte tagged unions that can hold the concrete representation of any type in TorchScript. TorchScript is statically typed, so it would be possible to operate on unboxed primitive types, but the interface between interpreter, built-in ops and user functions would be significantly more complicated. A single tagged union keeps these interfaces simple and since most objects are Tensors
anyway, the overhead of storing a tag is small compared to the data stored in the Tensors
.
IValue contains methods to check the type (e.g. isTensor()
) and to convert to particular to type (e.g. toTensor()
). We do not publicly expose the type tag and force clients to use the isX
methods. This enables us to change the underlying implementation of IValue
later, e.g. to use an 8-byte value with NaN-boxing. Most operators work on a specific static type, so dynamic dispatch on the tag is not frequently required.
All builtin operators are represented using a stack machine concept. An operator pops its arguments off the top of the stack and pushes its result to the stack:
using Stack = std::vector<IValue>;
using Operation = std::function<void(Stack*)>;
// schema: example_add(Tensor a, Tensor b) -> Tensor
void example_add(Stack* stack) {
Tensor a, b;
// stack before: ? ? ? a b <- back
pop(stack, a, b); // Templated helper function
// that pops a, b and converts them to Tensor
push(stack, a + b);
// stack after:
// ? ? ? c <- back
}
Most operations, apart from some vararg primitive operators like prim::Unpack, have an associated FunctionSchema that describes how many inputs will be popped and how many will be pushed.
The stack concept makes it easy to define operators with variable numbers of inputs and outputs without the need to allocate vectors of inputs and outputs for each individual operator.
In practice, the interpreter will allocate one Stack, and it will eventually reach a sufficient size such that no more stack-related memory allocations will occur.
The Operator object represents a single registered operator in the system. It combines a FunctionSchema that describes how an Operation executes with a method to look up the corresponding Operation given the Node representing the operator in a Graph. Most Operators are defined by providing a FunctionSchema and an Operation function. However, primitives like prim::Unpack require knowledge of their Node to know how to operate (e.g. how many elements to unpack). These Operators have a function that takes a Node*
and returns an operation.
The interpreter is responsible for the straightforward execution of Graphs
without any optimization. It is composed of two objects: Code and InterpreterState. Code is a linearized representation of the Graph
into simple stack-machine Instructions. Code is shared among all the executions of the Graph
and will include caches for certain operations like the generated CUDA code of FusionGroups.
The InterpreterState is unique to each execution of the Graph
. It holds a list registers with the intermediate IValues
used in the execution, the Stack being used by each Operation, and the program counter tracking the position in the instructions. The information represents the complete state of the interpreter. wait
instructions can cause the interpreter to suspend, and the InterpreterState is used to resume execution where the wait
occurred, potentially on a different thread.
Instructions in the interpreter have three parts: a list of registers from which to gather IValues
onto the stack before the instruction, the Operation to run, and a list of registers in which to store the results of the Operation. Alternatively, we could have used individual instructions to load/store values from the stack to registers, but this design was easier to implement, requires fewer instructions since each instruction does more things, and has not yet been a performance bottleneck. Each Operation returns a potential relative jump to compute the next program counter.
Unlike typical interpreters, we do not attempt careful register allocation. Since Tensors
are reference types, saving registers would only save a few hundred bytes of space in typical applications by cutting down on the number of places a reference could be saved. The data in single a Tensor
is likely significantly bigger than that, so we forgo register allocation to make debugging easier.
However, we do need to ensure that values are destructed immediately after their last use. Because Torch reference counts Tensors
, they will be deallocated immediately when their last reference is gone. To ensure we use a minimum amount of memory we want to ensure that the interpreter releases the reference as soon as it is no longer used. To do this, each Instruction also has a set of flags which indicate the inputs to the operation which will no longer be used after the operation. For these inputs, the IValue
is moved rather than copied from the register file, ensuring the reference will go dead as soon as the Operation no longer needs it. extra instructions may be inserted into the program to explicitly drop references for values whose last use depends on the control flow of the program.
The following is an example program in Graph
form and its equivalent in interpreter Instructions:
graph(%x : Tensor,
%hx : Tensor,
%cx : Tensor,
%w_ih : Tensor,
%w_hh : Tensor,
%b_ih : Tensor,
%b_hh : Tensor):
%7 : int = prim::Constant[value=4]()
%8 : int = prim::Constant[value=1]()
%9 : Tensor = aten::t(%w_ih)
%10 : Tensor = aten::mm(%x, %9)
%11 : Tensor = aten::t(%w_hh)
%12 : Tensor = aten::mm(%hx, %11)
%13 : Tensor = aten::add(%10, %12, %8)
%14 : Tensor = aten::add(%13, %b_ih, %8)
%gates : Tensor = aten::add(%14, %b_hh, %8)
%16 : Tensor[] = aten::chunk(%gates, %7, %8)
%ingate.1 : Tensor, %forgetgate.1 : Tensor, %cellgate.1 : Tensor, %outgate.1 : Tensor = prim::ListUnpack(%16)
%ingate : Tensor = aten::sigmoid(%ingate.1)
%forgetgate : Tensor = aten::sigmoid(%forgetgate.1)
%cellgate : Tensor = aten::tanh(%cellgate.1)
%outgate : Tensor = aten::sigmoid(%outgate.1)
%25 : Tensor = aten::mul(%forgetgate, %cx)
%26 : Tensor = aten::mul(%ingate, %cellgate)
%cy : Tensor = aten::add(%25, %26, %8)
%28 : Tensor = aten::tanh(%cy)
%hy : Tensor = aten::mul(%outgate, %28)
%30 : (Tensor, Tensor) = prim::TupleConstruct(%hy, %cy)
return (%30)
0, 1, 2, 3, 4, 5, 6 = Load
7 = Constant
8 = t move(3)
9 = mm move(0), move(8)
10 = t move(4)
11 = mm move(1), move(10)
12 = add move(9), move(11), 7
13 = add move(12), move(5), 7
14 = add move(13), move(6), 7
15, 16, 17, 18 = ConstantChunk move(14)
19 = sigmoid move(15)
20 = sigmoid move(16)
21 = tanh move(17)
22 = sigmoid move(18)
23 = mul move(20), move(2)
24 = mul move(19), move(21)
25 = add move(23), move(24), move(7)
26 = tanh 25
27 = mul move(22), move(26)
28 = TupleConstruct move(27), move(25)
= Store move(28)
All program execution starts with a graph executor. Its responsible for running optimizations (potentially involving the JIT-compilation of fused kernel code), and then handing the Graph
or subcomponents of it off to an interpreter to actually run.
In this section, we use a running example program that computes one step of an LSTM to show how the graph is transformed:
This section will use an example this LSTM program:
@torch.jit.script
def LSTMCellS(x, hx, cx, w_ih, w_hh, b_ih, b_hh):
gates = x.mm(w_ih.t()) + hx.mm(w_hh.t()) + b_ih + b_hh
ingate, forgetgate, cellgate, outgate = gates.chunk(4, 1)
ingate = torch.sigmoid(ingate)
forgetgate = torch.sigmoid(forgetgate)
cellgate = torch.tanh(cellgate)
outgate = torch.sigmoid(outgate)
cy = (forgetgate * cx) + (ingate * cellgate)
hy = outgate * torch.tanh(cy)
return hy, cy
After going through the the frontend, we start with this unoptimized graph:
graph(%x : Tensor,
%hx : Tensor,
%cx : Tensor,
%w_ih : Tensor,
%w_hh : Tensor,
%b_ih : Tensor,
%b_hh : Tensor):
%7 : int = prim::Constant[value=4]()
%8 : int = prim::Constant[value=1]()
%9 : Tensor = aten::t(%w_ih)
%10 : Tensor = aten::mm(%x, %9)
%11 : Tensor = aten::t(%w_hh)
%12 : Tensor = aten::mm(%hx, %11)
%13 : Tensor = aten::add(%10, %12, %8)
%14 : Tensor = aten::add(%13, %b_ih, %8)
%gates : Tensor = aten::add(%14, %b_hh, %8)
%16 : Tensor[] = aten::chunk(%gates, %7, %8)
%ingate.1 : Tensor, %forgetgate.1 : Tensor, %cellgate.1 : Tensor, %outgate.1 : Tensor = prim::ListUnpack(%16)
%ingate : Tensor = aten::sigmoid(%ingate.1)
%forgetgate : Tensor = aten::sigmoid(%forgetgate.1)
%cellgate : Tensor = aten::tanh(%cellgate.1)
%outgate : Tensor = aten::sigmoid(%outgate.1)
%25 : Tensor = aten::mul(%forgetgate, %cx)
%26 : Tensor = aten::mul(%ingate, %cellgate)
%cy : Tensor = aten::add(%25, %26, %8)
%28 : Tensor = aten::tanh(%cy)
%hy : Tensor = aten::mul(%outgate, %28)
%30 : (Tensor, Tensor) = prim::TupleConstruct(%hy, %cy)
return (%30)
Execution starts in GraphExecutor::run
, which takes a Stack of inputs.
The executor specializes the Graph
for the particular set of inputs. Specialization is handled by the ArgumentSpec
object which extracts a "signature" composed of all the properties being specialized. We only specialize to the properties of Tensors
. The ArgumentSpec
only records properties for Tensors
that either appear directly in the inputs to the graph or inside Tuples that are inputs to the Graph
. The properties recorded are currently:
- dtype
- rank, but not size
- requires_grad
- device type (CPU, CUDA)
- defined - whether the
Tensor
exists or is a placeholder
The ArgumentSpec object is used as a key into a cache that holds pre-optimized Code objects (held in an ExecutionPlan object). On a cache hit, an InterpreterState is created and the Code in the cache is run.
On a code cache miss, we generate a new optimized Graph
on the fly (compileSpec
). It starts by creating a copy of the initial Graph
and setting the input types to the specialized Tensor
types observed in this specialization. TensorType inputs to the Graph
will get refined with types that know the device, number of dimensions, and requires grad state.
# post specialization, inputs are now specialized types
graph(%x : Float(*, *),
%hx : Float(*, *),
%cx : Float(*, *),
%w_ih : Float(*, *),
%w_hh : Float(*, *),
%b_ih : Float(*),
%b_hh : Float(*)):
%7 : int = prim::Constant[value=4]()
%8 : int = prim::Constant[value=1]()
%9 : Tensor = aten::t(%w_ih)
%10 : Tensor = aten::mm(%x, %9)
%11 : Tensor = aten::t(%w_hh)
%12 : Tensor = aten::mm(%hx, %11)
%13 : Tensor = aten::add(%10, %12, %8)
%14 : Tensor = aten::add(%13, %b_ih, %8)
%gates : Tensor = aten::add(%14, %b_hh, %8)
%16 : Tensor[] = aten::chunk(%gates, %7, %8)
%ingate.1 : Tensor, %forgetgate.1 : Tensor, %cellgate.1 : Tensor, %outgate.1 : Tensor = prim::ListUnpack(%16)
%ingate : Tensor = aten::sigmoid(%ingate.1)
%forgetgate : Tensor = aten::sigmoid(%forgetgate.1)
%cellgate : Tensor = aten::tanh(%cellgate.1)
%outgate : Tensor = aten::sigmoid(%outgate.1)
%25 : Tensor = aten::mul(%forgetgate, %cx)
%26 : Tensor = aten::mul(%ingate, %cellgate)
%cy : Tensor = aten::add(%25, %26, %8)
%28 : Tensor = aten::tanh(%cy)
%hy : Tensor = aten::mul(%outgate, %28)
%30 : (Tensor, Tensor) = prim::TupleConstruct(%hy, %cy)
return (%30)
It then runs "required passes", which are graph transformations necessary to generate legal graphs for the interpreter. (Some passes such as differentiation will introduce Nodes
that are not defined by operators and require passes to clean up. The combination of specializeUndef
and LowerGradOf
clean up these operations.) These passes also remove broadcasting "expand" nodes that get implicitly inserted by the tracer but are not valid for all sizes.
It then runs inference passes to calculate properties of the graph given this particular specialization:
- It propagates constants, pre-computing as much as possible
- It propagates the input ranks, dtypes, devices, and requires_grad information to the rest of the graph where possible.
graph(%x : Float(*, *),
%hx : Float(*, *),
%cx : Float(*, *),
%w_ih : Float(*, *),
%w_hh : Float(*, *),
%b_ih : Float(*),
%b_hh : Float(*)):
%8 : int = prim::Constant[value=1]()
%9 : Float(*, *) = aten::t(%w_ih)
%10 : Float(*, *) = aten::mm(%x, %9)
%11 : Float(*, *) = aten::t(%w_hh)
%12 : Float(*, *) = aten::mm(%hx, %11)
%13 : Float(*, *) = aten::add(%10, %12, %8)
%14 : Float(*, *) = aten::add(%13, %b_ih, %8)
%gates : Float(*, *) = aten::add(%14, %b_hh, %8)
%31 : Float(*, *), %32 : Float(*, *), %33 : Float(*, *), %34 : Float(*, *) = prim::ConstantChunk[chunks=4, dim=1](%gates)
%ingate : Float(*, *) = aten::sigmoid(%31)
%forgetgate : Float(*, *) = aten::sigmoid(%32)
%cellgate : Float(*, *) = aten::tanh(%33)
%outgate : Float(*, *) = aten::sigmoid(%34)
%25 : Float(*, *) = aten::mul(%forgetgate, %cx)
%26 : Float(*, *) = aten::mul(%ingate, %cellgate)
%cy : Float(*, *) = aten::add(%25, %26, %8)
%28 : Float(*, *) = aten::tanh(%cy)
%hy : Float(*, *) = aten::mul(%outgate, %28)
%30 : (Float(*, *), Float(*, *)) = prim::TupleConstruct(%hy, %cy)
return (%30)
It then runs a number of derivative preserving optimization passes. If a computation involves Tensors
that have requires_grad
and it is valid to compute its derivative, then these passes are only allowed to replace that computation with another computation that is also differentiable. In other words, these passes cannot break autograd. Algebraic rewrites and peephole optimizations are generally derivative preserving but something that generates code, like pointwise fusion, is not.
Current derivative preserving passes:
- Eliminating dead code
- Eliminating common subexpressions
- Pooling redundant constants into single values
- Peephole optimizations, including some algebraic rewrites into simpler operations
- Unrolling small loops
- Batching matrix multiplications that result from unrolling loops
graph(%x : Float(*, *),
%hx : Float(*, *),
%cx : Float(*, *),
%w_ih : Float(*, *),
%w_hh : Float(*, *),
%b_ih : Float(*),
%b_hh : Float(*)):
%8 : int = prim::Constant[value=1]()
%9 : Float(*, *) = aten::t(%w_ih)
%10 : Float(*, *) = aten::mm(%x, %9)
%11 : Float(*, *) = aten::t(%w_hh)
%12 : Float(*, *) = aten::mm(%hx, %11)
%13 : Float(*, *) = aten::add(%10, %12, %8)
%14 : Float(*, *) = aten::add(%13, %b_ih, %8)
%gates : Float(*, *) = aten::add(%14, %b_hh, %8)
%31 : Float(*, *), %32 : Float(*, *), %33 : Float(*, *), %34 : Float(*, *) = prim::ConstantChunk[chunks=4, dim=1](%gates)
%ingate : Float(*, *) = aten::sigmoid(%31)
%forgetgate : Float(*, *) = aten::sigmoid(%32)
%cellgate : Float(*, *) = aten::tanh(%33)
%outgate : Float(*, *) = aten::sigmoid(%34)
%25 : Float(*, *) = aten::mul(%forgetgate, %cx)
%26 : Float(*, *) = aten::mul(%ingate, %cellgate)
%cy : Float(*, *) = aten::add(%25, %26, %8)
%28 : Float(*, *) = aten::tanh(%cy)
%hy : Float(*, *) = aten::mul(%outgate, %28)
%30 : (Float(*, *), Float(*, *)) = prim::TupleConstruct(%hy, %cy)
return (%30)
The next optimization depends on whether any part of the graph actual requires a gradient to be calculated, which is determined by needsGradient
. In the case where no gradients are required (i.e. for inference graphs), then we can directly apply optimizations that generate graphs that may not have valid gradients defined. For now this is the FuseGraph
pass, which looks for adjacent point-wise operations along with reviewing operations such as split
and concat
, and creates prim::FusionGroup
Nodes
in the graph to replace these operations. The Operator registered to execute prim::FusionGroup
nodes will generate a new CUDA kernel for each unique Node
, which replaces the original separate execution.
Note the two phases for compilation of fusion groups: First, the FuseGraph
pass splits the Graph
into fusible sub-Graphs and returns the resulting Graph
to the graph executor. Second, when the Graph
is turned into Code, the Operation for the FusionGroup node will be looked up and a new CUDA kernel generated for the body. Other compilers should work in a similar way by first introducing a new operator into the Graph
where the compiled code should run, and then registering an Operator that implements that Node
which performs the actual compilation.
In the case where no gradients are required, the optimization process is finished, a Code object is constructed from the Graph
, it is added to the code cache, and then an InterpreterState is constructed and run.
graph(%x : Float(*, *),
%hx : Float(*, *),
%cx : Float(*, *),
%w_ih : Float(*, *),
%w_hh : Float(*, *),
%b_ih : Float(*),
%b_hh : Float(*)):
%9 : Float(*, *) = aten::t(%w_ih)
%10 : Float(*, *) = aten::mm(%x, %9)
%11 : Float(*, *) = aten::t(%w_hh)
%12 : Float(*, *) = aten::mm(%hx, %11)
%77 : Tensor[] = prim::ListConstruct(%b_hh, %b_ih, %10, %12)
%78 : Tensor[] = aten::broadcast_tensors(%77)
%79 : Tensor, %80 : Tensor, %81 : Tensor, %82 : Tensor = prim::ListUnpack(%78)
%hy : Float(*, *), %cy : Float(*, *) = prim::FusionGroup_0(%cx, %82, %81, %80, %79)
%30 : (Float(*, *), Float(*, *)) = prim::TupleConstruct(%hy, %cy)
return (%30);
with prim::FusionGroup_0 = graph(%13 : Float(*, *),
%71 : Tensor,
%76 : Tensor,
%81 : Tensor,
%86 : Tensor):
%87 : Float(*, *), %88 : Float(*, *), %89 : Float(*, *), %90 : Float(*, *) = prim::ConstantChunk[chunks=4, dim=1](%86)
%82 : Float(*, *), %83 : Float(*, *), %84 : Float(*, *), %85 : Float(*, *) = prim::ConstantChunk[chunks=4, dim=1](%81)
%77 : Float(*, *), %78 : Float(*, *), %79 : Float(*, *), %80 : Float(*, *) = prim::ConstantChunk[chunks=4, dim=1](%76)
%72 : Float(*, *), %73 : Float(*, *), %74 : Float(*, *), %75 : Float(*, *) = prim::ConstantChunk[chunks=4, dim=1](%71)
%69 : int = prim::Constant[value=1]()
%70 : Float(*, *) = aten::add(%77, %72, %69)
%66 : Float(*, *) = aten::add(%78, %73, %69)
%62 : Float(*, *) = aten::add(%79, %74, %69)
%58 : Float(*, *) = aten::add(%80, %75, %69)
%54 : Float(*, *) = aten::add(%70, %82, %69)
%50 : Float(*, *) = aten::add(%66, %83, %69)
%46 : Float(*, *) = aten::add(%62, %84, %69)
%42 : Float(*, *) = aten::add(%58, %85, %69)
%38 : Float(*, *) = aten::add(%54, %87, %69)
%34 : Float(*, *) = aten::add(%50, %88, %69)
%30 : Float(*, *) = aten::add(%46, %89, %69)
%26 : Float(*, *) = aten::add(%42, %90, %69)
%ingate : Float(*, *) = aten::sigmoid(%38)
%forgetgate : Float(*, *) = aten::sigmoid(%34)
%cellgate : Float(*, *) = aten::tanh(%30)
%outgate : Float(*, *) = aten::sigmoid(%26)
%14 : Float(*, *) = aten::mul(%forgetgate, %13)
%11 : Float(*, *) = aten::mul(%ingate, %cellgate)
%cy : Float(*, *) = aten::add(%14, %11, %69)
%4 : Float(*, *) = aten::tanh(%cy)
%hy : Float(*, *) = aten::mul(%outgate, %4)
return (%hy, %cy)
Many Graphs
will require gradients (i.e. one of the inputs will have a requires_grad
property set). In this case, it is unsafe to run post-derivative optimizations directly on the Graph
. Instead, our approach is to first split the Graph
into sub-Graphs where symbolic gradient formulas are known and produce an explicit Graph
for the forward pass along with a complementary Graph
that implements the backwards pass using some of the values computed in the forward pass. We can then apply post-derivative optimization to the forward graph. The "gradOutputs" for the backwards graph are only known when the backward pass runs, so we cannot fully optimize it at this time. For instance, we do not know if some of those gradOutputs will also require_grad
meaning that a gradient-of-gradient situation exists. Instead the backward pass will use a new GraphExecutor object to run and optimize its execution. In this way, we can handle an indefinite number of recursive gradient calculations.
The creating of derivative subgraphs is done using a similar approach to finding fusion groups: adjacent operations with known gradient formulas are grouped together into prim::DifferentiableGraph
nodes. We only generate these nodes if we can find a large enough subgraph where optimization is likely to be profitable since there is some overhead involved in entering and exiting a differentiable subgraph.
graph(%x : Float(*, *),
%hx : Float(*, *),
%cx : Float(*, *),
%w_ih : Float(*, *),
%w_hh : Float(*, *),
%b_ih : Float(*),
%b_hh : Float(*)):
%8 : int = prim::Constant[value=1]()
%hy : Float(*, *), %cy : Float(*, *) = prim::DifferentiableGraph_0(%cx, %b_hh, %b_ih, %hx, %w_hh, %x, %w_ih)
%30 : (Float(*, *), Float(*, *)) = prim::TupleConstruct(%hy, %cy)
return (%30)
with prim::DifferentiableGraph_0 = graph(%13 : Float(*, *),
%29 : Float(*),
%33 : Float(*),
%40 : Float(*, *),
%43 : Float(*, *),
%45 : Float(*, *),
%48 : Float(*, *)):
%49 : Float(*, *) = aten::t(%48)
%47 : Float(*, *) = aten::mm(%45, %49)
%44 : Float(*, *) = aten::t(%43)
%42 : Float(*, *) = aten::mm(%40, %44)
%38 : int = prim::Constant[value=1]()
%39 : Float(*, *) = aten::add(%47, %42, %38)
%35 : Float(*, *) = aten::add(%39, %33, %38)
%gates : Float(*, *) = aten::add(%35, %29, %38)
%24 : Float(*, *), %25 : Float(*, *), %26 : Float(*, *), %27 : Float(*, *) = prim::ConstantChunk[chunks=4, dim=1](%gates)
%ingate : Float(*, *) = aten::sigmoid(%24)
%forgetgate : Float(*, *) = aten::sigmoid(%25)
%cellgate : Float(*, *) = aten::tanh(%26)
%outgate : Float(*, *) = aten::sigmoid(%27)
%14 : Float(*, *) = aten::mul(%forgetgate, %13)
%11 : Float(*, *) = aten::mul(%ingate, %cellgate)
%cy : Float(*, *) = aten::add(%14, %11, %38)
%4 : Float(*, *) = aten::tanh(%cy)
%hy : Float(*, *) = aten::mul(%outgate, %4)
return (%hy, %cy)
Logging is a very useful debugging technique, especially in the context of compilers. Compilers perform a series of passes and analyses and logging can help to trace issues such as wrong results or segmentation faults all the way back to the original erroneous transformation.
TorchScript
offers a simple logging facility that can enabled by setting an environment variable PYTORCH_JIT_LOG_LEVEL
.
Logging is enabled on a per file basis. To enable logging in dead_code_elimination.cpp
, PYTORCH_JIT_LOG_LEVEL
should be
set to dead_code_elimination.cpp
or, simply, to dead_code_elimination
(i.e. PYTORCH_JIT_LOG_LEVEL=dead_code_elimination
).
Multiple files can be logged by separating each file name with a colon :
as in the following example, PYTORCH_JIT_LOG_LEVEL=dead_code_elimination:guard_elimination
There are 3 logging levels available for your use ordered by the detail level from lowest to highest.
GRAPH_DUMP
should be used for printing entire graphs after optimization passesGRAPH_UPDATE
should be used for reporting graph transformations (i.e. node deletion, constant folding, etc)GRAPH_DEBUG
should be used for providing information useful for debugging the internals of a particular optimization pass or analysis
The current logging level is GRAPH_UPDATE
meaning that both GRAPH_DUMP
and GRAPH_UPDATE
will be enabled when
one specifies a file(s) in PYTORCH_JIT_LOG_LEVEL
.
GRAPH_DEBUG
can be enabled by prefixing a file name with an >
as in >alias_analysis
.
>>
and >>>
are also valid and currently are equivalent to GRAPH_DEBUG
as there is no logging level that is
higher than GRAPH_DEBUG
.
By default, types in the graph are printed with maximum verbosity. The verbosity level can be controlled via the environment variable PYTORCH_JIT_TYPE_VERBOSITY
. The available settings are:
0
: No type information1
: Types and shapes only2
: Also print strides3
: Also print device type and whether gradient is required
Often times, we need to limit the number of optimizations for any lowering passes for debugging purposes.
TorchScript
offers a simple optimization limit checker that can be configured through environment variable PYTORCH_JIT_OPT_LIMIT
. The purpose is to limit how many optimization you can make per pass. This is useful for debugging any passes.
Opt limit checker is enabled on a per file basis (hence per pass). For example, in constant_propagation.cpp
, PYTORCH_JIT_OPT_LIMIT
should be set to constant_propagation=<opt_limit>
where <opt_limit>
is the number of optimizations you want to make for the pass. (i.e.
PYTORCH_JIT_OPT_LIMIT="constant_propagation=<opt_limit>"
).
Multiple files can be configured by separating each file name with a colon
:
as in the following example,
PYTORCH_JIT_OPT_LIMIT="constant_propagation=<opt_limit>:dead_code_elimination=<opt_limit>"
You can call opt limiter by calling a macro JIT_OPT_ALLOWED
. It will return true if
we haven't reached the optimization limit yet. Typical usage:
if (!JIT_OPT_ALLOWED) {
GRAPH_DUMP(...); //supplied from jit_log
return;
}
A DifferentiableGraphOp combines an explicit forward Graph
f
with a paired backward graph df
. When it runs, the input Tensors
to f
are detached from the autograd, the body of f
is run, and then the autograd graph for the outputs of f
are hooked up to the df
function. The df
function's outputs are also hooked up to the autograd graph.
In PyTorch, Tensors
are reference types. Operators can return "views" of the input Tensor
, creating a new Tensor
object that shares the same underlying storage as the original:
a = torch.rand(2, 3)
b = a
# At this point, `a` and `b` share their storage.
c = b[0]
# `c` shares storage with `a` and `b`, but only sees a slice of the allocated memory.
Some operators will mutate one or more of their operands in-place. These are typically denoted with a trailing underscore, or by taking an out
argument as input:
a = torch.zeros(2, 3)
b = torch.ones(2, 3)
a.add_(b) # in-place add, so `a` is modified.
torch.add(a, b, out=a) # another way to express the same thing
The JIT's FunctionSchema
allows operator writers to add annotations specifying the aliasing and mutation behavior of an operator. Optimization passes will use this information to determine whether transformations are semantics-preserving. This section provides a description of the alias annotation language, assuming that the reader already knows what FunctionSchema
looks like.
First, here is a pure function which always returns new memory:
add(Tensor a, Tensor b) -> Tensor
The type Tensor
with no annotations is sugar for "fresh, read-only Tensor
". So since there are no annotations on anything, we know that this operator creates no aliases and mutates no inputs.
Next, a function that returns an alias to one of the inputs.:
view(Tensor(a) self, int[] size) -> Tensor(a)
The shared (a)
annotation on self
and the output signify that the Tensors
will share the same storage. Another way to say is that self
and the output belong to the same "alias set" a
.
Now a function that writes in-place to one of the inputs (note the trailing underscore):
add_(Tensor(a!) self, Tensor other) -> Tensor(a!)
The !
annotation means that this operator writes to the specified alias set (in this case a
).
Sometimes we don't have enough information to provide an exact alias annotation. For example, here is the operator to extract an element from a list:
list_select(Tensor[] list, int idx) -> Tensor(*)
Note the alias set *
. This is the wildcard set. These are values which we conservatively analyze. Containers, such as lists and dictionaries, Graph inputs, and class attributes are conservatively analyzed to all alias. In most cases, people shouldn't be writing operators with wildcard annotations. They are used as temporary workaround for when our alias analysis isn't sophisticated enough to understand something yet but we don't want to block feature development.
Similarly, we have operators which result in Tensors being contained in a list. In this case, to preserve the relationship between output list and input, we annotate that the input enters the wildcard set with the (a -> *)
syntax.
func: chunk(Tensor(a -> *) self, int chunks, int dim=0) -> Tensor(a)[]
This annotation language is consumed by the FunctionSchema
parser, which produces AliasInfo
objects summarizing the aliasing relationships for each schema Argument
.
An alias analysis pass consumes the per-operator aliasing information to construct a database of aliasing and mutation relationships in a graph, called AliasDb
. This section focuses on the alias analysis pass; the public interface to AliasDb
will be described later.
The core data structure in the AliasDb is called MemoryDAG
, which is a DAG where the edges are "may point to" relationships and the vertices are aliasing Element
s. The most common kind of Element
is an IR Value
, but there are other kinds of things that can alias that aren't first-class Value
s in the IR, like wildcards or contained types (such as in a list or tuple).
The alias analysis pass walks through the nodes in a graph, examining schema AliasInfo
objects and adding edges in the MemoryDAG
accordingly. For example, for the node:
%output : Tensor = aten::view(%self, %size)
the analyzer will examine the schema for view()
:
view(Tensor(a) self, int[] size) -> Tensor(a)
and add an edge from %output
to %self
. The alias analysis pass is flow-insensitive, as we are only adding "points-to" edges when processing a node.
As a more involved example, the following TorchScript snippet:
@torch.jit.script
def foo(a : Tensor, b : Tensor):
c = 2 * b
a += 1
if a.max() > 4:
r = a[0]
else:
r = b[0]
return c, r
Will produce a graph like this:
A few things to note:
- "Graph Input Element" is an example of an
Element
that isn't a first-classValue
. Alias analysis happens on a per-function level, so we don't necessarily know the aliasing relationships of the inputs. The only safe assumption is thata
andb
may alias each other, so they point to a specialElement
that describes "the world outside of this function". r
may point to eithera
orb
, depending on the runtime value ofa.max()
. A givenElement
may point to multiple otherElement
s. This can happen if there is branching control flow (like in this example), or with certain ops likecontiguous()
, which either returns an alias to the input or a freshTensor
, depending on the runtime characteristics of the input.c
is a freshTensor
(i.e. it doesn't point to anything) since it was created using the pure operation2 * b
.
The last point demonstrates a key concept: leaf elements uniquely describe memory locations. Since a leaf element doesn't point to anything, the memory that backs it must have been freshly allocated by some op. Thus we can use leaf elements to represent disjoint memory locations.
So to determine whether a
and b
may alias, we traverse the MemoryDAG
DAG and figure out if a
and b
share any leaf nodes. If they do, then we know a
and b
might point to the same memory location, i.e. a
and b
may alias. This kind of query is common enough that MemoryDAG
does path compression to speed up leaf-finding, so that aliasing queries can be serviced in amortized constant time.
AliasDb
provides a high-level interface to help people write mutability-safe optimization passes.
In particular, moveAfterTopologicallyValid()
(and its moveBefore
variant) will reorder nodes in a way that preserves data dependencies and avoids any data hazards. The rules for this are that all mutable writes to a given memory location must occur in the same order (avoid WAW hazards), and that no reads can be reordered before or after any write (WAR, RAW hazards).
However, reordering of reads across writes is allowed if we can prove that the read cannot alias the thing being written. This happens whenever we have Tensors
that come from functions that produce fresh results (common) inside of the function. It also happens whenever the creation of the mutable Tensor
is seen in the function (so it gets assigned a fresh variable), and all of its writes occur in that function.
The intention is that if you only mutate the graph through AliasDb
, you don't have to think about mutability/aliasing at all in your pass. As we write more passes, the interface to AliasDb
will get richer (one example is transforming an in-place operation to its pure equivalent if we can prove it's safe).
AliasDb
also provides lower level APIs that users of LLVM's alias analysis pass would be familiar with, such as querying whether any two Value
s may alias.
TODO: differentiation, symbolic autograd, fusion, operators
prim::profile
nodes are inserted on every use of a value by ProfilingRecord::instrumentBlock
. Every prim::profile
node runs a lambda that uses a captured, initial type value and the type of an incoming Tensor
and merges the two into a refined TensorType
.
prim::profile
nodes are replaced with prim::Guard
nodes by InsertGuards
. prim::Guard
nodes are inserted to guarantee that beyond the guard a guarded Tensor
will always be of the profiled shape. This guarantee will enable optimizations and code generators to generate more efficient code.
We attempt to reduce the number of prim::Guard
nodes as these nodes may interfere with optimizations.
- First,
GuardElimination::moveGuardsToDefs
tries to moveprim::Guards
to their definitions, so the guards guarding the sameTensor
follow the definition directly or another guard on the sameTensor
. - This ordering allows us to coalesce (done in
GuardElimination::coalesceGuards
) multiple guards into a single one. - After guards are coaslesced ,
GuardElimination::eliminateGuards
attempts to eliminate more guards as follows: it inspects each operation and its inputs. It checks if inputs to the operation are guarded and also if the operation produces the consistent shapes given the guarded inputs. For example, if two inputs toadd
are guaranteed to be of shape(2, 3)
, the output shape will also always be(2, 3)
. If this property holds, we are allowed to remove the guard guarding operation's output.
Lastly, we need to be handle cases when the assumptions about Tensor
shapes fail at runtime. To handle guard failures, we need to be able to run the original code i.e. the code that doesn't rely on assumptions about shapes. As guards can be inserted and moved (by Optimizer) at/to arbitrary points in a computational graph, we need to be able to resume execution starting from those arbitrary points onward.
InsertBailoutNodes
builds deoptimized versions of the original computational graph, that contain the rest of computations starting from their corresponding guard failure points and also captures live values needed to execute those deoptimized graphs. In other words, the pass replaces prim::Guard
nodes with prim::BailOut
nodes which have theattr::Subgraph
attributes set to the deoptimized versions of the remaining computations at their corresponding prim::Guard
s.
When differentiating a graph, each node that has a symbolic gradient will be included in a prim::DifferentiableGraph
. We fall back to using autograd for the node if there isn't a gradient formula for it.
Adding/updating symbolic gradient functions must be tested carefully as it's easy to get CI green by comparing autograd result with itself, but potentially cause an autodiff support regression.
If your PR adds/updates a gradient formula for torch
/nn
functions, you MUST enable/update the corresponding tests in
torch
functions:method_tests
in common_method_tests.pynn
functions:nn_functional_tests
in test_jit.py
To turn on autodiff check, you can add an optional check_ad(should_autodiff_node[bool], nonfusible_nodes[str|list[str]], fusible_nodes[str|list[str]])
tuple after the optional test variant name field.
If should_autodiff_node=True
, the differentiated traced/script forward graph must have a prim::DifferentiableGraph
.
All nodes in nonfusible_nodes
should show up at least once in prim::DifferentiableGraph
subgraphs.
When fusion is enabled, all nodes in fusible_nodes
should show up in one of prim::FusionGroup
graphs attached to prim::DifferentiableGraph
,
otherwise they're checked as nonfusible_nodes
as well.
On the other hand, if should_autodiff_node=False
, the graph can still have prim::DifferentiableGraph
with other nodes, but not nonfusible_nodes
and fusible_nodes
.
To make writing tests easier, you only need to write out node names if it's different from the function name. Below are a few examples:
('conv1d', ...), # No symbolic gradient formula
('avg_pool2d', ..., (True,)), # Has symbolic gradient formula, only has one nonfusible node aten::avg_pool2d
('nll_loss', ..., (True, 'aten::nll_loss_forward')), # Is replaced by a different node in its symbolic gradient formula
('dropout', ..., (True, ['prim::is_CUDA', 'aten::bernoulli_'], ['aten::rand_like', ..., 'aten::div'])), # Some ops are fused when fusion is enabled
Note that even for the same function, different tests could trigger different function schemas (e.g aten::add
) while only a few of them have symbolic gradient formulas.
You should only turn on autodiff checks in tests that have symbolic gradients. If you are not sure, uncomment the debugging line in runtime/symbolic_script.cpp
to check which function schema the test triggers.
serialization/python_print.cpp serialization/import_source.cpp
The Python Printer takes a Graph
and produces Python-like code that represents the same graph. Using some special values in serialization/import_source.cpp, this code can be read back in by the compiler to produce the same Graph
. In Python a ScriptModule
's code
property shows the Python Printed graph.
The table below shows the graph and code for this small ScriptModule
:
class M(torch.jit.ScriptModule):
@torch.jit.script_method
def forward(self, x, y, z):
# type: (Tensor, int, float) -> Tensor
if y > 2:
x = x + z
else:
x = x + y
return x
m = M()
m.graph
graph(%x.1 : Tensor,
%y : int,
%z : float):
%5 : int = prim::Constant[value=1]()
%3 : int = prim::Constant[value=2]()
%4 : bool = aten::gt(%y, %3)
%x : Tensor = prim::If(%4)
block0():
%x.2 : Tensor = aten::add(%x.1, %z, %5)
-> (%x.2)
block1():
%x.3 : Tensor = aten::add(%x.1, %y, %5)
-> (%x.3)
return (%x)
m.code
def forward(self,
x: Tensor,
y: int,
z: float) -> Tensor:
if torch.gt(y, 2):
x0 = torch.add(x, z, 1)
else:
x0 = torch.add(x, y, 1)
return x0
TODO: Script Module, torch.jit.trace, constant handling, weak script modules