Introduction to torch.compile

Author: William Wen

torch.compile is the latest method to speed up your PyTorch code! torch.compile makes PyTorch code run faster by JIT-compiling PyTorch code into optimized kernels, all while requiring minimal code changes.

In this tutorial, we cover basic torch.compile usage, and demonstrate the advantages of torch.compile over previous PyTorch compiler solutions, such as TorchScript and FX Tracing.

Contents

Required pip Dependencies

• torch >= 2.0

• torchvision

• numpy

• scipy

• tabulate

NOTE: a modern NVIDIA GPU (H100, A100, or V100) is recommended for this tutorial in order to reproduce the speedup numbers shown below and documented elsewhere.

import torch
import warnings

gpu_ok = False
if torch.cuda.is_available():
    device_cap = torch.cuda.get_device_capability()
    if device_cap in ((7, 0), (8, 0), (9, 0)):
        gpu_ok = True

if not gpu_ok:
    warnings.warn(
        "GPU is not NVIDIA V100, A100, or H100. Speedup numbers may be lower "
        "than expected."
    )

/workspace/tutorials-kr/intermediate_source/torch_compile_tutorial.py:48: UserWarning:

GPU is not NVIDIA V100, A100, or H100. Speedup numbers may be lower than expected.

Basic Usage

torch.compile is included in the latest PyTorch. Running TorchInductor on GPU requires Triton, which is included with the PyTorch 2.0 nightly binary. If Triton is still missing, try installing torchtriton via pip (pip install torchtriton --extra-index-url "https://download.pytorch.org/whl/nightly/cu117" for CUDA 11.7).

Arbitrary Python functions can be optimized by passing the callable to torch.compile. We can then call the returned optimized function in place of the original function.

def foo(x, y):
    a = torch.sin(x)
    b = torch.cos(y)
    return a + b
opt_foo1 = torch.compile(foo)
print(opt_foo1(torch.randn(10, 10), torch.randn(10, 10)))

Traceback (most recent call last):
  File "/workspace/tutorials-kr/intermediate_source/torch_compile_tutorial.py", line 72, in <module>
    print(opt_foo1(torch.randn(10, 10), torch.randn(10, 10)))
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/eval_frame.py", line 451, in _fn
    return fn(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py", line 921, in catch_errors
    return callback(frame, cache_entry, hooks, frame_state, skip=1)
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py", line 786, in _convert_frame
    result = inner_convert(
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py", line 400, in _convert_frame_assert
    return _compile(
  File "/usr/lib/python3.10/contextlib.py", line 79, in inner
    return func(*args, **kwds)
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py", line 676, in _compile
    guarded_code = compile_inner(code, one_graph, hooks, transform)
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/utils.py", line 262, in time_wrapper
    r = func(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py", line 535, in compile_inner
    out_code = transform_code_object(code, transform)
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/bytecode_transformation.py", line 1036, in transform_code_object
    transformations(instructions, code_options)
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py", line 165, in _fn
    return fn(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py", line 500, in transform
    tracer.run()
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/symbolic_convert.py", line 2149, in run
    super().run()
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/symbolic_convert.py", line 810, in run
    and self.step()
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/symbolic_convert.py", line 773, in step
    getattr(self, inst.opname)(inst)
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/symbolic_convert.py", line 2268, in RETURN_VALUE
    self.output.compile_subgraph(
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/output_graph.py", line 981, in compile_subgraph
    self.compile_and_call_fx_graph(tx, list(reversed(stack_values)), root)
  File "/usr/lib/python3.10/contextlib.py", line 79, in inner
    return func(*args, **kwds)
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/output_graph.py", line 1178, in compile_and_call_fx_graph
    compiled_fn = self.call_user_compiler(gm)
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/utils.py", line 262, in time_wrapper
    r = func(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/output_graph.py", line 1251, in call_user_compiler
    raise BackendCompilerFailed(self.compiler_fn, e).with_traceback(
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/output_graph.py", line 1232, in call_user_compiler
    compiled_fn = compiler_fn(gm, self.example_inputs())
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/repro/after_dynamo.py", line 117, in debug_wrapper
    compiled_gm = compiler_fn(gm, example_inputs)
  File "/usr/local/lib/python3.10/dist-packages/torch/__init__.py", line 1731, in __call__
    return compile_fx(model_, inputs_, config_patches=self.config)
  File "/usr/lib/python3.10/contextlib.py", line 79, in inner
    return func(*args, **kwds)
  File "/usr/local/lib/python3.10/dist-packages/torch/_inductor/compile_fx.py", line 1330, in compile_fx
    return aot_autograd(
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/backends/common.py", line 58, in compiler_fn
    cg = aot_module_simplified(gm, example_inputs, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/_functorch/aot_autograd.py", line 903, in aot_module_simplified
    compiled_fn = create_aot_dispatcher_function(
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/utils.py", line 262, in time_wrapper
    r = func(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/_functorch/aot_autograd.py", line 628, in create_aot_dispatcher_function
    compiled_fn = compiler_fn(flat_fn, fake_flat_args, aot_config, fw_metadata=fw_metadata)
  File "/usr/local/lib/python3.10/dist-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py", line 443, in aot_wrapper_dedupe
    return compiler_fn(flat_fn, leaf_flat_args, aot_config, fw_metadata=fw_metadata)
  File "/usr/local/lib/python3.10/dist-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py", line 648, in aot_wrapper_synthetic_base
    return compiler_fn(flat_fn, flat_args, aot_config, fw_metadata=fw_metadata)
  File "/usr/local/lib/python3.10/dist-packages/torch/_functorch/_aot_autograd/jit_compile_runtime_wrappers.py", line 119, in aot_dispatch_base
    compiled_fw = compiler(fw_module, updated_flat_args)
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/utils.py", line 262, in time_wrapper
    r = func(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/_inductor/compile_fx.py", line 1257, in fw_compiler_base
    return inner_compile(
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/repro/after_aot.py", line 83, in debug_wrapper
    inner_compiled_fn = compiler_fn(gm, example_inputs)
  File "/usr/local/lib/python3.10/dist-packages/torch/_inductor/debug.py", line 304, in inner
    return fn(*args, **kwargs)
  File "/usr/lib/python3.10/contextlib.py", line 79, in inner
    return func(*args, **kwds)
  File "/usr/lib/python3.10/contextlib.py", line 79, in inner
    return func(*args, **kwds)
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/utils.py", line 262, in time_wrapper
    r = func(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/_inductor/compile_fx.py", line 438, in compile_fx_inner
    compiled_graph = fx_codegen_and_compile(
  File "/usr/local/lib/python3.10/dist-packages/torch/_inductor/compile_fx.py", line 714, in fx_codegen_and_compile
    compiled_fn = graph.compile_to_fn()
  File "/usr/local/lib/python3.10/dist-packages/torch/_inductor/graph.py", line 1307, in compile_to_fn
    return self.compile_to_module().call
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/utils.py", line 262, in time_wrapper
    r = func(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/_inductor/graph.py", line 1254, in compile_to_module
    mod = PyCodeCache.load_by_key_path(
  File "/usr/local/lib/python3.10/dist-packages/torch/_inductor/codecache.py", line 2160, in load_by_key_path
    exec(code, mod.__dict__, mod.__dict__)
  File "/tmp/torchinductor_root/mm/cmmtcbmmfakm37djocu6jtpak3f6yuldla55frjautjluv52xvyz.py", line 59, in <module>
    async_compile.wait(globals())
  File "/usr/local/lib/python3.10/dist-packages/torch/_inductor/codecache.py", line 2715, in wait
    scope[key] = result.result()
  File "/usr/lib/python3.10/concurrent/futures/_base.py", line 458, in result
    return self.__get_result()
  File "/usr/lib/python3.10/concurrent/futures/_base.py", line 403, in __get_result
    raise self._exception
  File "/usr/lib/python3.10/concurrent/futures/thread.py", line 58, in run
    result = self.fn(*self.args, **self.kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/_inductor/codecache.py", line 2074, in load_pybinding
    result = cls.load(source_code + suffix, cuda)
  File "/usr/local/lib/python3.10/dist-packages/torch/_inductor/codecache.py", line 1948, in load
    compile_file(input_path, output_path, cmd)
  File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/utils.py", line 262, in time_wrapper
    r = func(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/_inductor/codecache.py", line 1888, in compile_file
    raise exc.CppCompileError(cmd, output) from e
torch._dynamo.exc.BackendCompilerFailed: backend='inductor' raised:
CppCompileError: C++ compile error

Command:
g++ /tmp/torchinductor_root/pg/cpg44lbh7nfyakvqvjrcqw4qoozmn277q2j4vwqoke5ugrzrdp45.cpp -shared -fPIC -Wall -std=c++17 -Wno-unused-variable -Wno-unknown-pragmas -D_GLIBCXX_USE_CXX11_ABI=0 -I/usr/local/lib/python3.10/dist-packages/torch/include -I/usr/local/lib/python3.10/dist-packages/torch/include/torch/csrc/api/include -I/usr/local/lib/python3.10/dist-packages/torch/include/TH -I/usr/local/lib/python3.10/dist-packages/torch/include/THC -I/usr/include/python3.10 -L/usr/local/lib/python3.10/dist-packages/torch/lib -L/usr/lib/x86_64-linux-gnu -L/usr/local/lib/python3.10/dist-packages/torch/lib -ltorch -ltorch_cpu -lgomp -ltorch_python -lc10 -mavx2 -mfma -DCPU_CAPABILITY_AVX2 -O3 -DNDEBUG -ffast-math -fno-finite-math-only -fno-unsafe-math-optimizations -ffp-contract=off -march=native -fopenmp -D C10_USING_CUSTOM_GENERATED_MACROS -o /tmp/torchinductor_root/pg/cpg44lbh7nfyakvqvjrcqw4qoozmn277q2j4vwqoke5ugrzrdp45.so

Output:
/tmp/torchinductor_root/pg/cpg44lbh7nfyakvqvjrcqw4qoozmn277q2j4vwqoke5ugrzrdp45.cpp:32:10: fatal error: Python.h: No such file or directory
   32 | #include <Python.h>
      |          ^~~~~~~~~~
compilation terminated.


Set TORCH_LOGS="+dynamo" and TORCHDYNAMO_VERBOSE=1 for more information


You can suppress this exception and fall back to eager by setting:
    import torch._dynamo
    torch._dynamo.config.suppress_errors = True

Alternatively, we can decorate the function.

@torch.compile
def opt_foo2(x, y):
    a = torch.sin(x)
    b = torch.cos(y)
    return a + b
print(opt_foo2(torch.randn(10, 10), torch.randn(10, 10)))

We can also optimize torch.nn.Module instances.

class MyModule(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.lin = torch.nn.Linear(100, 10)

    def forward(self, x):
        return torch.nn.functional.relu(self.lin(x))

mod = MyModule()
opt_mod = torch.compile(mod)
print(opt_mod(torch.randn(10, 100)))

Demonstrating Speedups

Let’s now demonstrate that using torch.compile can speed up real models. We will compare standard eager mode and torch.compile by evaluating and training a torchvision model on random data.

Before we start, we need to define some utility functions.

# Returns the result of running `fn()` and the time it took for `fn()` to run,
# in seconds. We use CUDA events and synchronization for the most accurate
# measurements.
def timed(fn):
    start = torch.cuda.Event(enable_timing=True)
    end = torch.cuda.Event(enable_timing=True)
    start.record()
    result = fn()
    end.record()
    torch.cuda.synchronize()
    return result, start.elapsed_time(end) / 1000

# Generates random input and targets data for the model, where `b` is
# batch size.
def generate_data(b):
    return (
        torch.randn(b, 3, 128, 128).to(torch.float32).cuda(),
        torch.randint(1000, (b,)).cuda(),
    )

N_ITERS = 10

from torchvision.models import densenet121
def init_model():
    return densenet121().to(torch.float32).cuda()

First, let’s compare inference.

Note that in the call to torch.compile, we have have the additional mode argument, which we will discuss below.

model = init_model()

# Reset since we are using a different mode.
import torch._dynamo
torch._dynamo.reset()

model_opt = torch.compile(model, mode="reduce-overhead")

inp = generate_data(16)[0]
with torch.no_grad():
    print("eager:", timed(lambda: model(inp))[1])
    print("compile:", timed(lambda: model_opt(inp))[1])

Notice that torch.compile takes a lot longer to complete compared to eager. This is because torch.compile compiles the model into optimized kernels as it executes. In our example, the structure of the model doesn’t change, and so recompilation is not needed. So if we run our optimized model several more times, we should see a significant improvement compared to eager.

eager_times = []
for i in range(N_ITERS):
    inp = generate_data(16)[0]
    with torch.no_grad():
        _, eager_time = timed(lambda: model(inp))
    eager_times.append(eager_time)
    print(f"eager eval time {i}: {eager_time}")

print("~" * 10)

compile_times = []
for i in range(N_ITERS):
    inp = generate_data(16)[0]
    with torch.no_grad():
        _, compile_time = timed(lambda: model_opt(inp))
    compile_times.append(compile_time)
    print(f"compile eval time {i}: {compile_time}")
print("~" * 10)

import numpy as np
eager_med = np.median(eager_times)
compile_med = np.median(compile_times)
speedup = eager_med / compile_med
assert(speedup > 1)
print(f"(eval) eager median: {eager_med}, compile median: {compile_med}, speedup: {speedup}x")
print("~" * 10)

And indeed, we can see that running our model with torch.compile results in a significant speedup. Speedup mainly comes from reducing Python overhead and GPU read/writes, and so the observed speedup may vary on factors such as model architecture and batch size. For example, if a model’s architecture is simple and the amount of data is large, then the bottleneck would be GPU compute and the observed speedup may be less significant.

You may also see different speedup results depending on the chosen mode argument. The "reduce-overhead" mode uses CUDA graphs to further reduce the overhead of Python. For your own models, you may need to experiment with different modes to maximize speedup. You can read more about modes here.

You may might also notice that the second time we run our model with torch.compile is significantly slower than the other runs, although it is much faster than the first run. This is because the "reduce-overhead" mode runs a few warm-up iterations for CUDA graphs.

For general PyTorch benchmarking, you can try using torch.utils.benchmark instead of the timed function we defined above. We wrote our own timing function in this tutorial to show torch.compile’s compilation latency.

Now, let’s consider comparing training.

model = init_model()
opt = torch.optim.Adam(model.parameters())

def train(mod, data):
    opt.zero_grad(True)
    pred = mod(data[0])
    loss = torch.nn.CrossEntropyLoss()(pred, data[1])
    loss.backward()
    opt.step()

eager_times = []
for i in range(N_ITERS):
    inp = generate_data(16)
    _, eager_time = timed(lambda: train(model, inp))
    eager_times.append(eager_time)
    print(f"eager train time {i}: {eager_time}")
print("~" * 10)

model = init_model()
opt = torch.optim.Adam(model.parameters())
train_opt = torch.compile(train, mode="reduce-overhead")

compile_times = []
for i in range(N_ITERS):
    inp = generate_data(16)
    _, compile_time = timed(lambda: train_opt(model, inp))
    compile_times.append(compile_time)
    print(f"compile train time {i}: {compile_time}")
print("~" * 10)

eager_med = np.median(eager_times)
compile_med = np.median(compile_times)
speedup = eager_med / compile_med
assert(speedup > 1)
print(f"(train) eager median: {eager_med}, compile median: {compile_med}, speedup: {speedup}x")
print("~" * 10)

Again, we can see that torch.compile takes longer in the first iteration, as it must compile the model, but in subsequent iterations, we see significant speedups compared to eager.

We remark that the speedup numbers presented in this tutorial are for demonstration purposes only. Official speedup values can be seen at the TorchInductor performance dashboard.

Comparison to TorchScript and FX Tracing

We have seen that torch.compile can speed up PyTorch code. Why else should we use torch.compile over existing PyTorch compiler solutions, such as TorchScript or FX Tracing? Primarily, the advantage of torch.compile lies in its ability to handle arbitrary Python code with minimal changes to existing code.

One case that torch.compile can handle that other compiler solutions struggle with is data-dependent control flow (the if x.sum() < 0: line below).

def f1(x, y):
    if x.sum() < 0:
        return -y
    return y

# Test that `fn1` and `fn2` return the same result, given
# the same arguments `args`. Typically, `fn1` will be an eager function
# while `fn2` will be a compiled function (torch.compile, TorchScript, or FX graph).
def test_fns(fn1, fn2, args):
    out1 = fn1(*args)
    out2 = fn2(*args)
    return torch.allclose(out1, out2)

inp1 = torch.randn(5, 5)
inp2 = torch.randn(5, 5)

TorchScript tracing f1 results in silently incorrect results, since only the actual control flow path is traced.

traced_f1 = torch.jit.trace(f1, (inp1, inp2))
print("traced 1, 1:", test_fns(f1, traced_f1, (inp1, inp2)))
print("traced 1, 2:", test_fns(f1, traced_f1, (-inp1, inp2)))

FX tracing f1 results in an error due to the presence of data-dependent control flow.

import traceback as tb
try:
    torch.fx.symbolic_trace(f1)
except:
    tb.print_exc()

If we provide a value for x as we try to FX trace f1, then we run into the same problem as TorchScript tracing, as the data-dependent control flow is removed in the traced function.

fx_f1 = torch.fx.symbolic_trace(f1, concrete_args={"x": inp1})
print("fx 1, 1:", test_fns(f1, fx_f1, (inp1, inp2)))
print("fx 1, 2:", test_fns(f1, fx_f1, (-inp1, inp2)))

Now we can see that torch.compile correctly handles data-dependent control flow.

# Reset since we are using a different mode.
torch._dynamo.reset()

compile_f1 = torch.compile(f1)
print("compile 1, 1:", test_fns(f1, compile_f1, (inp1, inp2)))
print("compile 1, 2:", test_fns(f1, compile_f1, (-inp1, inp2)))
print("~" * 10)

TorchScript scripting can handle data-dependent control flow, but this solution comes with its own set of problems. Namely, TorchScript scripting can require major code changes and will raise errors when unsupported Python is used.

In the example below, we forget TorchScript type annotations and we receive a TorchScript error because the input type for argument y, an int, does not match with the default argument type, torch.Tensor.

def f2(x, y):
    return x + y

inp1 = torch.randn(5, 5)
inp2 = 3

script_f2 = torch.jit.script(f2)
try:
    script_f2(inp1, inp2)
except:
    tb.print_exc()

However, torch.compile is easily able to handle f2.

compile_f2 = torch.compile(f2)
print("compile 2:", test_fns(f2, compile_f2, (inp1, inp2)))
print("~" * 10)

Another case that torch.compile handles well compared to previous compilers solutions is the usage of non-PyTorch functions.

import scipy
def f3(x):
    x = x * 2
    x = scipy.fft.dct(x.numpy())
    x = torch.from_numpy(x)
    x = x * 2
    return x

TorchScript tracing treats results from non-PyTorch function calls as constants, and so our results can be silently wrong.

inp1 = torch.randn(5, 5)
inp2 = torch.randn(5, 5)
traced_f3 = torch.jit.trace(f3, (inp1,))
print("traced 3:", test_fns(f3, traced_f3, (inp2,)))

TorchScript scripting and FX tracing disallow non-PyTorch function calls.

try:
    torch.jit.script(f3)
except:
    tb.print_exc()

try:
    torch.fx.symbolic_trace(f3)
except:
    tb.print_exc()

In comparison, torch.compile is easily able to handle the non-PyTorch function call.

compile_f3 = torch.compile(f3)
print("compile 3:", test_fns(f3, compile_f3, (inp2,)))

TorchDynamo and FX Graphs

One important component of torch.compile is TorchDynamo. TorchDynamo is responsible for JIT compiling arbitrary Python code into FX graphs, which can then be further optimized. TorchDynamo extracts FX graphs by analyzing Python bytecode during runtime and detecting calls to PyTorch operations.

Normally, TorchInductor, another component of torch.compile, further compiles the FX graphs into optimized kernels, but TorchDynamo allows for different backends to be used. In order to inspect the FX graphs that TorchDynamo outputs, let us create a custom backend that outputs the FX graph and simply returns the graph’s unoptimized forward method.

from typing import List
def custom_backend(gm: torch.fx.GraphModule, example_inputs: List[torch.Tensor]):
    print("custom backend called with FX graph:")
    gm.graph.print_tabular()
    return gm.forward

# Reset since we are using a different backend.
torch._dynamo.reset()

opt_model = torch.compile(init_model(), backend=custom_backend)
opt_model(generate_data(16)[0])

Using our custom backend, we can now see how TorchDynamo is able to handle data-dependent control flow. Consider the function below, where the line if b.sum() < 0 is the source of data-dependent control flow.

def bar(a, b):
    x = a / (torch.abs(a) + 1)
    if b.sum() < 0:
        b = b * -1
    return x * b

opt_bar = torch.compile(bar, backend=custom_backend)
inp1 = torch.randn(10)
inp2 = torch.randn(10)
opt_bar(inp1, inp2)
opt_bar(inp1, -inp2)

The output reveals that TorchDynamo extracted 3 different FX graphs corresponding the following code (order may differ from the output above):

1. x = a / (torch.abs(a) + 1)

2. b = b * -1; return x * b

3. return x * b

When TorchDynamo encounters unsupported Python features, such as data-dependent control flow, it breaks the computation graph, lets the default Python interpreter handle the unsupported code, then resumes capturing the graph.

Let’s investigate by example how TorchDynamo would step through bar. If b.sum() < 0, then TorchDynamo would run graph 1, let Python determine the result of the conditional, then run graph 2. On the other hand, if not b.sum() < 0, then TorchDynamo would run graph 1, let Python determine the result of the conditional, then run graph 3.

This highlights a major difference between TorchDynamo and previous PyTorch compiler solutions. When encountering unsupported Python features, previous solutions either raise an error or silently fail. TorchDynamo, on the other hand, will break the computation graph.

We can see where TorchDynamo breaks the graph by using torch._dynamo.explain:

# Reset since we are using a different backend.
torch._dynamo.reset()
explain_output = torch._dynamo.explain(bar)(torch.randn(10), torch.randn(10))
print(explain_output)

In order to maximize speedup, graph breaks should be limited. We can force TorchDynamo to raise an error upon the first graph break encountered by using fullgraph=True:

opt_bar = torch.compile(bar, fullgraph=True)
try:
    opt_bar(torch.randn(10), torch.randn(10))
except:
    tb.print_exc()

And below, we demonstrate that TorchDynamo does not break the graph on the model we used above for demonstrating speedups.

opt_model = torch.compile(init_model(), fullgraph=True)
print(opt_model(generate_data(16)[0]))

We can use torch.export (from PyTorch 2.1+) to extract a single, exportable FX graph from the input PyTorch program. The exported graph is intended to be run on different (i.e. Python-less) environments. One important restriction is that the torch.export does not support graph breaks. Please check this tutorial for more details on torch.export.

Conclusion

In this tutorial, we introduced torch.compile by covering basic usage, demonstrating speedups over eager mode, comparing to previous PyTorch compiler solutions, and briefly investigating TorchDynamo and its interactions with FX graphs. We hope that you will give torch.compile a try!