Graph Compiler#

⚠️ Warning: The compiler and the transpiler are not publicly available yet, so certain parts of this doc won’t work as expected as of now!

When we call an Ivy function, there is always a small performance hit due to added Python wrapping. This overhead becomes increasingly noticeable when we use large models with multiple function calls. The Graph Compiler improves the performance of Ivy by removing the extra wrapping around each function call.

The Graph Compiler takes in any Ivy function, framework-specific (backend) function, or composition of both, and produces a simplified executable computation graph composed of functions from the backend functional API only, which results in:

  • Simplified code: The Graph Compiler simplifies the code by removing all the wrapping and functions that don’t contribute to the output: print statements, loggers, etc.

  • Improved performance: The compiled graph has no performance overhead due to Ivy’s function wrapping, likewise, redundant operations from the original function are also removed, increasing its overall performance.

Compiler API#

ivy.compile(*objs, stateful=None, arg_stateful_idxs=None, kwarg_stateful_idxs=None, to=None, include_generators=True, array_caching=True, return_backend_compiled_fn=False, static_argnums=None, static_argnames=None, args=None, kwargs=None)#

Compiles a Callable or set of them into an Ivy graph. If args or kwargs are specified, compilation is performed eagerly, otherwise, compilation will happen lazily.

  • objs (Callable) – Callable(s) to compile and create a graph of.

  • stateful (Optional[List]) – List of instances to be considered stateful during the graph compilation.

  • arg_stateful_idxs (Optional[List]) – Positional arguments to be considered stateful during the graph compilation.

  • kwarg_stateful_idxs (Optional[List]) – Keyword arguments to be considered stateful during the graph compilation.

  • to (Optional[str]) – Backend that the graph will be compiled to. If not specified, the current backend will be used.

  • include_generators (bool) – Include array creation/generation functions as part of the graph.

  • array_caching (bool) – Cache the constant arrays that appear as arguments to the functions in the graph.

  • return_backend_compiled_fn (bool) – Whether to apply the native compilers, i.e. tf.function, after ivy’s compilation.

  • static_argnums (Optional[Union[int, Iterable[int]]]) – For jax’s jit compilation.

  • static_argnames (Optional[Union[str, Iterable[str]]]) – For jax’s jit compilation.

  • args (Optional[Tuple]) – Positional arguments for obj.

  • kwargs (Optional[dict]) – Keyword arguments for obj.

Return type:

Union[Graph, LazyGraph, ivy.Module, ModuleType]


A compiled Graph or a non-initialized LazyGraph. If the object is an ivy.Module, the forward pass will be compiled and the same module will be returned. If the object is a ModuleType, the function will return a copy of the module with every method lazily compiled.

Using the compiler#

To use the ivy.compile() function, you need to pass a callable object and the corresponding inputs to the function.

Let’s start with a simple function:

import ivy


def fn(x, y):
    z = x**y
    k = x * y
    j = ivy.concat([x, z, y])
    sum_j = ivy.sum(j)
    return z

x = ivy.array([1, 2, 3])
y = ivy.array([2, 3, 4])

# Compile the function
compiled_fn = ivy.compile(fn, args=(x, y))

In this case, the compiled graph would be:

From the graph, we can observe that:

  1. As x and y are the only variables used when calculating the returned value z, the non-contributing variable(s), k was not included in the graph. Function calls that don’t contribute to the output like the print function were also excluded.

  2. As we set the backend to torch during the compilation process, the compiled functions are torch functions, and the input and output types are torch tensors.

  3. The tensor shape in the graph only indicates the shape of the inputs the graph was traced with. The compiler doesn’t impose additional restrictions on the shape or datatype of the input array(s).

# Original set of inputs
out = compiled_fn(x, y)

# Inputs of different shape
a = ivy.array([[1., 2.]])
b = ivy.array([[2., 3.]])

# New set of inputs
out = compiled_fn(x, y)

Eager vs lazy Compilation#

The graph compiler runs the original function under the hood and tracks its computation to create the compiled graph. The eager compilation method traces the graph in the corresponding function call with the specified inputs before we use the compiled function.

Instead of compiling functions before using them, Ivy also allows you to compile the function dynamically. This can be done by passing only the function to the compile method and not including the function arguments. In this case, the output will be a LazyGraph instead of a Graph instance. When this LazyGraph object is first invoked with function arguments, it compiles the function and returns the output of the compiled function. Once the graph has been initialized, calls to the LazyGraph object will use the compiled function to compute the outputs directly.

# Compile the function eagerly (compilation happens here)
eager_graph = ivy.compile(fn, args=(x, y))

# Compile the function lazily (compilation does not happen here)
lazy_graph = ivy.compile(fn)

# Compile and return the output
out = lazy_graph(x, y)

To sum up, lazy compilation enables you to delay the compilation process until you have the necessary inputs during execution. This is particularly useful in cases like compiling libraries, where it’s not feasible to provide valid arguments for every function call.

Now let’s look at additional functionalities that you can find in the compiler.

Array caching#

The compiler is able to cache constant arrays and their operations through the array_caching flag, reducing computation time after compilation.

import ivy


def fn(x):
    b = ivy.array([2])
    a = ivy.array([2])
    z = x ** (a + b)
    return z

comp_func = ivy.compile(fn, args=(x,))

When calling ivy.compile(), the array_caching argument is set to True by default, which returns the following graph.

This shows that by caching the constant operation in the graph, a simpler graph can be obtained. However, if desired, this argument can be set to False, which results in the graph below. This ultimately results in a trade-off between time and memory, as cached results need to be stored in memory but if they are not cached these operations need to be performed.


By using the include_generators argument, you can choose whether generator functions are included as nodes or “baked” into the graph.

import ivy


def fn(x):
    a = torch.randint(0, 100, size=[1])
    z = x * a
    return z + torch.rand([1])

comp_func = ivy.compile(fn, include_generators=True, args=(x,))


And instead,

import ivy


def fn(x):
    a = torch.randint(0, 100, size=[1])
    z = x * a
    return z + torch.rand([1])

comp_func = ivy.compile(fn, include_generators=False, args=(x,))



Finally, you can also track __setattr__ and __getattr__ methods of arbitrary classes using the stateful parameters.

import ivy


def fn(cont, x):
    cont.new_attribute = x
    return x + 1

x = torch.tensor([0])
cont = ivy.Container(x=x)

args = (cont.cont_deep_copy(), x)
comp_func = ivy.compile(fn, arg_stateful_idxs=[[0]], args=args)

Sharp bits#

As some parts of the graph compiler are still under development, there are some sharp bits to take into account when using it. All of these points are WIP, so they’ll be removed soon!

  1. Dynamic control flow: The compiled graph is built using function tracing at the moment, so dynamic control flow such as conditional branches or conditional loops will not be registered correctly. As an example, if there is a while loop in your code that depends on a changing value, the number of iterations in the final graph will be the same as the number of iterations performed with the input passed to the compile function.

  2. Non-framework-specific code: As the compiler traces the function using the functional API of the underlying framework, any piece of code inside the model that is not from said framework will not be correctly registered, this includes other frameworks code (such as NumPy statements inside a torch model) or python statements such as len().

  3. Incorrectly cached parts of the graph: There are certain cases where compilation can succeed but hide some cached parts of the graph which shouldn’t really be cached. To check this, it’s recommended to compile with a noise array of the same shape and then check if the output of the original function and the compiled graph with another input is the same. If you find out that the graph is not right, feel free to open an issue with a minimal example and we’ll look into it!


Below, we compile a ResNet50 model from Hugging Face and use it to classify the breed of a cat.

import ivy
from transformers import AutoImageProcessor, ResNetForImageClassification
from datasets import load_dataset

# Set backend to torch

# Download the input image
dataset = load_dataset("huggingface/cats-image")
image = dataset["test"]["image"][0]

# Setting the model
image_processor = AutoImageProcessor.from_pretrained("microsoft/resnet-50")
model = ResNetForImageClassification.from_pretrained("microsoft/resnet-50")

# Preprocessing the input image
inputs = image_processor(image, return_tensors="pt")

Normally, we would then feed these inputs to the model itself without compiling it

# Normal flow using pytorch
with torch.no_grad():
logits = model(**inputs).logits

With ivy, you can compile your model to a computation graph for increased performance.

# Compiling the model
compiled_graph = ivy.compile(model, args=(**inputs,))

# Using the compiled function
logits = compiled_graph(**inputs).logits

Time for the final output of our computation graph.

predicted_label = logits.argmax(-1).item()