PyTorch Mobile Development: Integration, Optimization, Inference

Machine Learning Development with PyTorch Mobile: Integration, Optimization, Inference

Your data science team trained a model in PyTorch, and now you need to run it on a mobile device. Converting to TFLite or Core ML requires rewriting part of the graph — no time, and non-standard operations aren't supported. We help port PyTorch models to iOS and Android via PyTorch Mobile, preserving accuracy and performance. Our experience — 30+ projects with ML on mobile devices, including NLP, CV, and recommendation systems. We guarantee model compatibility with target devices. Switching to on-device inference reduces cloud AI costs by up to 70%.

PyTorch Mobile is less common compared to TFLite and Core ML, but it wins in several scenarios. Primarily where the data science team works in PyTorch and doesn't want to spend time converting to another format. Or when a model uses custom operations that TFLite doesn't support. Additionally, PyTorch Mobile provides a unified API for both platforms, simplifying maintenance. It also supports quantization via qnnpack, yielding a 2-3x performance boost on convolution operations.

Why PyTorch Mobile Instead of TFLite?

If your model uses custom layers (e.g., custom attention), TFLite requires writing operators in C++ — weeks of work. PyTorch Mobile accepts TorchScript directly, and many operations are built-in. Comparison: for BERT-lite, conversion via ONNX to TFLite would take 2 days, while via TorchScript it takes 3 hours. PyTorch Mobile is better for rapid prototyping and models with unique architectures.

In this article, we'll cover how to convert a model to TorchScript, optimize it for mobile using optimize_for_mobile, perform INT8 quantization for speed, and integrate it into iOS and Android apps with Lite Interpreter. We'll also show a real case with BERT-lite, where the model size decreased from 23 MB to 6 MB, and inference time on a Pixel 6 was 45 ms for 128 tokens. This approach allows using powerful models on mobile devices without compromising the user experience.

TorchScript: The Key Requirement for PyTorch Mobile

PyTorch Mobile only works with TorchScript models — neither eager mode nor torch.fx works. Conversion via torch.jit.trace or torch.jit.script. The difference is fundamental: trace records the execution path for specific inputs and cannot handle data-dependent branching. script analyzes the graph statically and correctly handles if/for, but requires type annotations. If the model contains if x.shape[0] > 1:, trace silently records only one branch. In production, this manifests as incorrect results on certain batch sizes rather than a crash — hard to catch.

After conversion, optimize with optimize_for_mobile:

from torch.utils.mobile_optimizer import optimize_for_mobile
scripted = torch.jit.script(model)
optimized = optimize_for_mobile(scripted)
optimized._save_for_lite_interpreter("model.ptl")

.ptl (Lite Interpreter format) is not the same as .pt. On mobile, you use Lite Interpreter — it doesn't support all PyTorch operations, but has a smaller binary size.

How to Correctly Convert a Model to TorchScript?

The choice between trace and script depends on the model's dynamism. For models with conditional operators, script is the only option. We recommend writing tests after conversion: run the model on synthetic data and compare outputs before and after. If the discrepancy exceeds 1%, we locate the problematic node. This ensures the model works identically on mobile as on the server.

Quantization and Performance

Post-training static quantization for mobile:

model.qconfig = torch.quantization.get_default_qconfig('qnnpack')  # for ARM
torch.quantization.prepare(model, inplace=True)
# run calibration dataset
torch.quantization.convert(model, inplace=True)

qnnpack is the backend for ARM processors (what you need for Android and iOS). fbgemm is for x86 and doesn't work on mobile. This is a common mistake: a developer quantizes with fbgemm on a laptop and wonders why the model doesn't speed up on a phone.

Why Quantization with qnnpack and Not fbgemm?

Because mobile ARM processors use NEON instructions. qnnpack is optimized for them, fbgemm is for AVX2 on x86. Choosing the wrong backend results in zero speedup. In practice, the gain from INT8 on ARM: 2-3x on operations like Linear and Conv2d. On iPhones with Neural Engine, PyTorch Mobile does not leverage it directly — unlike Core ML. If you need Neural Engine on iOS, the correct path is conversion via coremltools, not PyTorch Mobile.

Integration on Android and iOS

Android. Dependency org.pytorch:pytorch_android_lite (Lite Interpreter). Inference:

val module = LiteModuleLoader.load(assetFilePath("model.ptl"))
val inputTensor = TensorImageUtils.bitmapToFloat32Tensor(bitmap, mean, std)
val output = module.forward(IValue.from(inputTensor)).toTensor()

Offload image preprocessing (normalization, resize) to Executors.newSingleThreadExecutor() — not on the main thread.

iOS. CocoaPod LibTorch-Lite. Work through TorchModule:

let module = TorchModule(fileAtPath: modelPath)
let result = module.predict(image: &tensorData)

All inference on DispatchQueue.global(qos: .userInitiated).

Example: NLP task, BERT-lite for classifying reviews inside a corporate app. The DS team worked in PyTorch, converting to TFLite was not desired (custom attention block). We used TorchScript + INT8 quantization (qnnpack). Model size decreased from 23 MB to 6 MB, inference on Pixel 6 — 45 ms for 128 tokens. Sufficient for real-time analytics.

How We Work: From Analysis to Deployment

1. Model analysis — evaluate the graph, operations, size. Identify problematic nodes for TorchScript.
2. Conversion and optimization — write conversion script, test on desktop, then optimize with optimize_for_mobile and quantize.
3. App integration — add Lite Interpreter, write wrapper for inference on a background thread to keep UI responsive.
4. Device testing — verify output correctness (compare with original model), measure performance, temperature, memory usage.
5. Deployment and monitoring — upload to App Store / Google Play, use Firebase for metrics. Updating the model only requires replacing model.ptl.

What's Included in the Work

• Full documentation on conversion and integration (including troubleshooting).
• Source code for iOS (Swift) and Android (Kotlin) integration.
• Instructions for updating the model without rebuilding the app.
• Support during store release (handling issues if the model violates guidelines).
• Guaranteed compatibility with iOS 15+ and Android 12+.

Timelines and Cost

Conversion and integration of a ready PyTorch model into Android or iOS — 1–2 weeks, including TorchScript debugging and device testing. Cost is calculated individually. Contact us for a project estimate — we'll provide approximate time and budget.

Common Mistakes in PyTorch Mobile Integration

• Using fbgemm for quantization instead of qnnpack.
• Forgetting to save the model in Lite Interpreter format (.ptl).
• Running inference on the main thread — causes UI freezes.
• Not checking that the model uses unsupported operations (e.g., torch.einsum may not be available).
• Not accounting for model size during loading — large models (>50 MB) may be evicted from memory.

We have been working with mobile development since the release of PyTorch Mobile and have completed 30+ projects with ML integration. Our experience spans over 10 years. We guarantee model compatibility with target devices.

Want to port your model to mobile? Contact us for a free consultation on timelines and complexity.

Machine Learning in Mobile Apps: CoreML, TFLite, and On-Device Models

We distinguish two fundamentally different approaches: an app with on-device AI and an app that simply calls a cloud API. The former works without internet, does not send user data to third-party servers, and responds within 50 milliseconds. The latter depends on network latency and pricing plans. Choosing the architecture is a key step that directly affects cost, privacy, and user experience in machine learning in mobile apps. Our experience shows that in 70% of projects, on-device inference is cheaper in the long run due to eliminating server costs.

How to Choose Between CoreML and TFLite for On-Device Inference?

CoreML — Apple's native framework for running ML models on device. Supports Neural Engine (starting with A11 Bionic), GPU, and CPU as fallback. Models are converted to .mlmodel format via coremltools from PyTorch, ONNX, or TensorFlow. Conversion is not always trivial: custom layers require implementing MLCustomLayer, and INT8 quantization can sometimes noticeably reduce accuracy on specific data. We ensure the final model passes validation on real data before and after conversion.

TensorFlow Lite — cross-platform alternative for Android and Flutter. On Android it uses NNAPI (Neural Networks API) for hardware acceleration — since Android 10 NNAPI is more stable; before that it's better to explicitly use GPU delegate via GpuDelegate. A typical mistake: the model is trained on normalized data in range [0,1], but the app feeds [0,255] — inference runs but produces meaningless results without any error. We include an automatic input data validation module in the SDK.

For image classification, object detection, and segmentation tasks, ready-to-use optimized models are available. YOLOv8 in CoreML format runs detection on a 640×640 frame in 15–20 ms on iPhone 14 Neural Engine. MobileNetV3 on TFLite with GPU delegate runs around 8 ms on Pixel 7 for classification.

What If You Need Text Generation On-Device?

Running small language models on device has become a reality in the last few years. Apple Intelligence uses its own models via Private Cloud Compute, but for third-party developers other paths are available.

llama.cpp with Metal backend on iOS is a working approach for phi-3-mini (3.8B parameters, 4-bit quantization, ~2.3 GB). Inference: 15–25 tokens/second on iPhone 15 Pro. For integration in Swift, use the Swift Package llama.swift or a wrapper via C interface llama.h. The binary is not bundled with the app — the model is downloaded on first launch and stored in Application Support. Our certified developers configure incremental download to avoid blocking the first launch.

On Android, the analog is Google AI Edge (formerly MediaPipe LLM Inference API) supporting Gemma-2B. It works via GPU delegate, on Tensor G3 chip Pixel 8 Pro — about 20 tokens/second.

Limitations are real: models larger than 4B parameters are still slow on mobile devices. For complex reasoning tasks, on-device LLM falls behind GPT-4o in quality. A hybrid approach — on-device for short tasks and private data, cloud for complex queries — is often optimal. We will evaluate your case and propose a balance of performance and privacy — contact us.

How Does On-Device Inference Compare to Cloud in Terms of Cost and Performance?

On-device inference is typically 10x cheaper per request than cloud APIs for image recognition tasks, while also eliminating latency variability and privacy risks. The table below summarizes the trade-offs.

For an app with 100k MAU running 10 image recognitions per user per month, on-device inference can save up to $5,000 monthly compared to cloud API. Get a free consultation on your ML architecture today.

Integrating OpenAI API and Other Cloud Models

For scenarios where cloud inference is acceptable, integrating OpenAI, Anthropic, or Google Gemini is an HTTP client + streaming SSE. In Swift, AsyncThrowingStream is convenient for streaming responses. In Kotlin, use Flow.

Critically: API keys must never be stored in the app bundle. Even an obfuscated key can be extracted from the IPA in 10 minutes using strings or frida. Correct architecture: mobile app → your own backend → OpenAI API. The backend controls rate limiting, logs requests, and protects the key.

What Is Included in the Work (Deliverables)

• Trained and quantized model for the target device (documentation with metrics)
• SDK for integration (Swift/Kotlin/Flutter) with call examples
• Performance tests on 3–5 real devices
• Instructions for OTA model updates
• Support during App Store / Google Play moderation (compliance with Guidelines 4.2, 5.1)
• 2 weeks of technical support after release

Typical Project Pipeline

1. Task analysis — measure latency, privacy, size, supported devices.
2. Model prototyping — in Python, evaluate accuracy on target data.
3. Conversion and quantization — for CoreML/TFLite with validation.
4. Integration into the app — model wrapped in a service layer (easy to swap CoreML ↔ TFLite ↔ cloud).
5. Testing — on real devices, measure FPS, RAM, battery.
6. Deployment — via TestFlight / Firebase App Distribution, monitor metrics.

Timelines: integration of a ready CoreML/TFLite model — 1–2 weeks, development of a custom model with mobile optimization — from 6 weeks, on-device LLM chat with personalization — 4–8 weeks.

Why We Take on Complex Cases?

10+ years of experience in mobile development, 50+ implemented AI/ML solutions, guarantee of compatibility with current iOS and Android versions. All projects undergo code review and load testing. The cost includes preparation of moderation documentation and training of your team.

Contact us — we will help you choose the architecture and implement ML in your app turnkey. Order an audit of your existing solution — we will assess the potential for server cost savings free of charge. In some projects, savings can reach significant amounts per month.