compile.rst

Compiled Graph Neural Networks

torch.compile is the latest method to speed up your PyTorch code in torch >= 2.0.0! torch.compile makes PyTorch code run faster by JIT-compiling it into optimized kernels, all while required minimal code changes.

Under the hood, torch.compile captures PyTorch programs via TorchDynamo, canonicalizes over 2,000 PyTorch operators via PrimTorch, and finally generates fast code out of it across multiple accelerators and backends via the deep learning compiler TorchInductor.

Note

See here for a general tutorial on how to leverage torch.compile, and here for a description of its interface.

In this tutorial, we show how to optimize your custom PyG model via torch.compile.

Note

From PyG 2.5 (and onwards), torch.compile is now fully compatible with all PyG GNN layers. If you are on an earlier version of PyG, consider using torch_geometric.compile instead.

Basic Usage

Once you have a PyG model defined, simply wrap it with torch.compile to obtain its optimized version:

import torch
from torch_geometric.nn import GraphSAGE

model = GraphSAGE(in_channels, hidden_channels, num_layers, out_channels)
model = model.to(device)

model = torch.compile(model)

and execute it as usual:

from torch_geometric.datasets import Planetoid

dataset = Planetoid(root, name="Cora")
data = dataset[0].to(device)

out = model(data.x, data.edge_index)

Maximizing Performance

The torch.compile method provides two important arguments to be aware of:

• Most of the mini-batches observed in PyG are dynamic by nature, meaning that their shape varies across different mini-batches. For these scenarios, we can enforce dynamic shape tracing in PyTorch via the dynamic=True argument:

  torch.compile(model, dynamic=True)

  With this, PyTorch will up-front attempt to generate a kernel that is as dynamic as possible to avoid recompilations when sizes change across mini-batches changes. Note that when dynamic is set to False, PyTorch will never generate dynamic kernels, and thus only works when graph sizes are guaranteed to never change (e.g., in full-batch training on small graphs). By default, dynamic is set to None in PyTorch >= 2.1.0, and PyTorch will automatically detect if dynamism has occured. Note that support for dynamic shape tracing requires PyTorch >= 2.1.0 to be installed.

• In order to maximize speedup, graphs breaks in the compiled model should be limited. We can force compilation to raise an error upon the first graph break encountered by using the fullgraph=True argument:

  torch.compile(model, fullgraph=True)

  It is generally a good practice to confirm that your written model does not contain any graph breaks. Importantly, there exists a few operations in PyG that will currently lead to graph breaks (but workaround exists), e.g.:

  1. ~torch_geometric.nn.pool.global_mean_pool (and other pooling operators) perform device synchronization in case the batch size size is not passed, leading to a graph break.
  2. ~torch_geometric.utils.remove_self_loops and ~torch_geometric.utils.add_remaining_self_loops mask the given edge_index, leading to a device synchronization to compute its final output shape. As such, we recommend to augment your graph before inputting it into your GNN, e.g., via the ~torch_geometric.transforms.AddSelfLoops or ~torch_geometric.transforms.GCNNorm transformations, and setting add_self_loops=False/normalize=False when initializing layers such as ~torch_geometric.nn.conv.GCNConv.

Exampe Scripts

We have incorporated multiple examples in examples/compile that further show the practical usage of torch.compile:

1. Node Classification via ~torch_geometric.nn.models.GCN (dynamic=False)
2. Graph Classification via ~torch_geometric.nn.models.GIN (dynamic=True)

If you notice that torch.compile fails for a certain PyG model, do not hesitate to reach out either on null GitHub or null Slack. We are very eager to improve torch.compile support across the whole PyG code base.

Benchmark

torch.compile works fantastically well for many PyG models. Overall, we observe runtime improvements of nearly up to 300%.

Specifically, we benchmark ~torch_geometric.nn.models.GCN, ~torch_geometric.nn.models.GraphSAGE and ~torch_geometric.nn.models.GIN and compare runtimes obtained from traditional eager mode and torch.compile. We use a synthetic graph with 10,000 nodes and 200,000 edges, and a hidden feature dimensionality of 64. We report runtimes over 500 optimization steps:

Model	Mode	Forward	Backward	Total	Speedup
~torch_geometric.nn.models.GCN	Eager	2.6396s	2.1697s	4.8093s
~torch_geometric.nn.models.GCN	Compiled	1.1082s	0.5896s	1.6978s	2.83x
~torch_geometric.nn.models.GraphSAGE	Eager	1.6023s	1.6428s	3.2451s
~torch_geometric.nn.models.GraphSAGE	Compiled	0.7033s	0.7465s	1.4498s	2.24x
~torch_geometric.nn.models.GIN	Eager	1.6701s	1.6990s	3.3690s
~torch_geometric.nn.models.GIN	Compiled	0.7320s	0.7407s	1.4727s	2.29x

To reproduce these results, run

python test/nn/models/test_basic_gnn.py

from the root folder of your checked out PyG repository from GitHub.