Skip to main content
React Native ExecuTorch supports running custom AI models beyond the pre-configured ones. This guide explains how to export your PyTorch models to the .pte format and integrate them into your React Native application.

​
Overview

To run custom models with ExecuTorch, you need to:
  1. Export your PyTorch model to the .pte format
  2. Optionally optimize the model for mobile devices
  3. Load the model using useExecutorchModule or ExecutorchModule

​
Exporting Models to .pte Format

​
Using the Python API

The ExecuTorch Python API provides tools to export PyTorch models. Here’s a basic example: 
import torch
from executorch.exir import to_edge
from torch.export import export
from executorch.exir import EdgeProgramManager

# 1. Load or define your model
model = YourModel()
model.eval()

# 2. Create example inputs
example_inputs = (torch.randn(1, 3, 224, 224),)

# 3. Export to ExecuTorch
aten_dialect = export(model, example_inputs)
edge_program = to_edge(aten_dialect)
executorch_program = edge_program.to_executorch()

# 4. Save to .pte file
with open("model.pte", "wb") as f:
    f.write(executorch_program.buffer)

​
Key Export Steps

Step 1: Prepare Your Model Ensure your model is in evaluation mode and all operations are supported by ExecuTorch: 
model.eval()
# Freeze model parameters
for param in model.parameters():
    param.requires_grad = False
Step 2: Define Representative Inputs Provide example inputs that match your expected input shape: 
# For image models
example_inputs = (torch.randn(1, 3, 224, 224),)

# For text models
example_inputs = (torch.randint(0, 1000, (1, 512)),)
Step 3: Export with Backend Delegation You can specify backends like XNNPACK for CPU optimization: 
from executorch.exir.backend.backend_api import to_backend
from executorch.backends.xnnpack.partition.xnnpack_partitioner import XnnpackPartitioner

# Export to edge
edge_program = to_edge(aten_dialect)

# Apply XNNPACK backend
edge_program = edge_program.to_backend(XnnpackPartitioner())

# Convert to ExecuTorch
executorch_program = edge_program.to_executorch()

​
Using optimum-executorch

optimum-executorch is a library by Hugging Face that simplifies exporting models:

​
Installation

pip install optimum-executorch

​
Export Hugging Face Models

from optimum.executorch import ExecuTorchModelForImageClassification

# Export a pre-trained model
model = ExecuTorchModelForImageClassification.from_pretrained(
    "google/mobilenet_v2_1.0_224",
    export=True
)

# Save the exported model
model.save_pretrained("./exported_model")

​
Export Custom Models

from optimum.executorch.exporter import export_to_executorch
import torch

# Your custom model
model = MyCustomModel()

# Export with optimum
export_to_executorch(
    model=model,
    output="model.pte",
    task="image-classification",
    example_inputs=torch.randn(1, 3, 224, 224)
)

​
Advanced Export Options

​
Quantization During Export

Reduce model size with quantization: 
from executorch.exir import EdgeCompileConfig
from torch.ao.quantization.quantize_pt2e import convert_pt2e, prepare_pt2e
from torch.ao.quantization.quantizer.xnnpack_quantizer import (
    get_symmetric_quantization_config,
    XNNPACKQuantizer,
)

# Prepare model for quantization
quantizer = XNNPACKQuantizer()
quantization_config = get_symmetric_quantization_config(is_per_channel=True)
quantizer.set_global(quantization_config)

# Export with quantization
aten_dialect = export(model, example_inputs)
m = prepare_pt2e(aten_dialect, quantizer)
m = convert_pt2e(m)

# Continue with edge export
edge_program = to_edge(m)
executorch_program = edge_program.to_executorch()

​
Core ML Backend (iOS)

For iOS devices, you can use Core ML for hardware acceleration: 
from executorch.exir.backend.backend_api import to_backend
from executorch.backends.apple.coreml.partition.coreml_partitioner import CoreMLPartitioner

edge_program = to_edge(aten_dialect)
edge_program = edge_program.to_backend(CoreMLPartitioner())
executorch_program = edge_program.to_executorch()

​
Loading Custom Models in React Native

​
Using useExecutorchModule Hook

Once you have your .pte file, load it in React Native: 
import { useExecutorchModule } from 'react-native-executorch';

function MyComponent() {
  const module = useExecutorchModule({
    modelSource: 'https://your-server.com/model.pte',
    // or from local assets
    // modelSource: require('./assets/model.pte'),
  });

  const runInference = async () => {
    if (!module.isReady) return;

    const input = /* prepare your input tensor */;
    const output = await module.forward([input]);
    console.log('Model output:', output);
  };

  return (
    // Your UI
  );
}

​
Using TypeScript API

For more control, use the TypeScript API directly: 
import { ExecutorchModule } from 'react-native-executorch';

const module = new ExecutorchModule();

await module.load({
  modelSource: 'https://your-server.com/model.pte',
});

const output = await module.forward([inputTensor]);

​
Input/Output Handling

​
Preparing Inputs

Inputs must match the model’s expected format: 
import { Tensor } from 'react-native-executorch';

// Create a tensor from a flat array
const input = Tensor.from({
  data: new Float32Array([/* your data */]),
  shape: [1, 3, 224, 224], // batch, channels, height, width
  dtype: 'float32',
});

​
Processing Outputs

const outputs = await module.forward([input]);

// Access output data
const outputData = outputs[0].data;
const outputShape = outputs[0].shape;

// Process results
const result = processModelOutput(outputData, outputShape);

​
Debugging Export Issues

​
Common Problems

Unsupported Operations If your model uses operations not supported by ExecuTorch: 
# Check which ops are used
from executorch.exir import ExecutorchProgramManager

# This will show unsupported ops during export
try:
    executorch_program = edge_program.to_executorch()
except Exception as e:
    print(f"Export failed: {e}")
Shape Mismatches Ensure dynamic shapes are properly handled: 
from torch.export import Dim

# Define dynamic dimensions
batch = Dim("batch", min=1, max=32)
dynamic_shapes = {"input": {0: batch}}

aten_dialect = export(
    model,
    example_inputs,
    dynamic_shapes=dynamic_shapes
)

​
Testing Your Exported Model

​
Validate with Python

Before using in React Native, test with Python: 
from executorch.extension.pybindings.portable_lib import (
    _load_for_executorch
)

# Load the .pte file
module = _load_for_executorch("model.pte")

# Test inference
test_input = torch.randn(1, 3, 224, 224)
output = module.forward((test_input,))
print("Output shape:", output[0].shape)

​
Best Practices

  1. Start Simple: Begin with basic models before attempting complex architectures
  2. Test Incrementally: Validate each export step before proceeding
  3. Check Operator Support: Verify all operations are supported by ExecuTorch
  4. Optimize for Mobile: Use quantization and backend delegation for better performance
  5. Version Control: Keep track of export scripts and model versions

​
Resources

​
Next Steps

Build docs developers (and LLMs) love

Get started for freeTalk to us