torch.compile End-to-End Tutorial">

intermediate/torch_compile_full_example

Run in Google Colab

Colab

Download Notebook

Notebook

View on GitHub

GitHub

torch.compile End-to-End Tutorial

Author: William Wen

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

This tutorial covers an end-to-end example of training and evaluating a real model with torch.compile. For a gentle introduction to torch.compile, please check out the introduction to torch.compile tutorial.

Required pip Dependencies

• torch >= 2.0

• torchvision

What you will learn

• How to apply torch.compile to a real model

• torch.compile speedups on a real model

• torch.compile’s first few iterations are expected to be slower due to compilation overhead

Prerequisites

• Introduction to torch.compile

# 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."
    )

/__w/tutorials/tutorials/pytorch/tutorials/intermediate_source/torch_compile_full_example.py:51: UserWarning: GPU is not NVIDIA V100, A100, or H100. Speedup numbers may be lower than expected.
  warnings.warn(

Let’s demonstrate how using torch.compile can speed up a real model. 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).cuda(),
        torch.randint(1000, (b,)).cuda(),
    )


N_ITERS = 10

from torchvision.models import densenet121


def init_model():
    return densenet121().cuda()

First, let’s compare inference.

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

model = init_model()

# Note that we generally recommend directly compiling a torch.nn.Module by calling
# its .compile() method.
model_opt = init_model()
model_opt.compile(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])

eager: 0.22596307373046876
/usr/local/lib/python3.12/dist-packages/triton/knobs.py:504: UserWarning: TRITON_DEFAULT_FP_FUSION is unset: defaulting to FP fusion ENABLED for backward compatibility in Triton 3.9. The default will change to DISABLED in Triton 3.10; set TRITON_DEFAULT_FP_FUSION=1 to keep fusion enabled or =0 to adopt the new default early and silence this warning.
  warn_once(FP_FUSION_DEFAULT_WARNING)
compile: 54.31173046875

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)

eager eval time 0: 0.017094720840454103
eager eval time 1: 0.015327072143554688
eager eval time 2: 0.015184896469116211
eager eval time 3: 0.015536128044128418
eager eval time 4: 0.01556275177001953
eager eval time 5: 0.015388671875
eager eval time 6: 0.015490079879760742
eager eval time 7: 0.015164416313171386
eager eval time 8: 0.015232000350952148
eager eval time 9: 0.015424511909484863
~~~~~~~~~~
compile eval time 0: 0.06051942443847656
compile eval time 1: 0.00760316801071167
compile eval time 2: 0.008262592315673828
compile eval time 3: 0.006761472225189209
compile eval time 4: 0.006646783828735352
compile eval time 5: 0.006662112236022949
compile eval time 6: 0.006647808074951172
compile eval time 7: 0.006663167953491211
compile eval time 8: 0.006806528091430664
compile eval time 9: 0.006660096168518067
~~~~~~~~~~
(eval) eager median: 0.01540659189224243, compile median: 0.00671232008934021, speedup: 2.2952707390563116x
~~~~~~~~~~

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.

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())

# Note that because we are compiling a regular Python function, we do not
# call any .compile() method.
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)

eager train time 0: 0.26461798095703126
eager train time 1: 0.04705587387084961
eager train time 2: 0.04315955352783203
eager train time 3: 0.04343500900268555
eager train time 4: 0.0432988166809082
eager train time 5: 0.043041793823242185
eager train time 6: 0.046929920196533206
eager train time 7: 0.04354457473754883
eager train time 8: 0.04309503936767578
eager train time 9: 0.04316159820556641
~~~~~~~~~~
W1106 20:00:03.774000 31018 torch/_logging/_internal.py:1407] [3/0] Profiler record function <class 'torch.autograd.profiler.record_function'> will be ignored
compile train time 0: 147.064078125
compile train time 1: 1.2757032470703125
compile train time 2: 0.021047168731689454
compile train time 3: 0.01965363121032715
compile train time 4: 0.018587583541870116
compile train time 5: 0.018520063400268554
compile train time 6: 0.01856198310852051
compile train time 7: 0.018518207550048828
compile train time 8: 0.018579456329345705
compile train time 9: 0.018945024490356444
~~~~~~~~~~
(train) eager median: 0.043366912841796876, compile median: 0.018766304016113282, speedup: 2.3108925872969346x
~~~~~~~~~~

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.

Conclusion

In this tutorial, we applied torch.compile to training and inference on a real model, demonstrating speedups.

Importantly, we note that the first few iterations of a compiled model are slower than eager mode due to compilation overhead, but subsequent iterations are expected to have speedups.

For a gentle introduction to torch.compile, please check out the introduction to torch.compile tutorial.

To troubleshoot issues and to gain a deeper understanding of how to apply torch.compile to your code, check out the torch.compile programming model.

We hope that you will give torch.compile a try!