Production Deployment Issues¶

Issues deploying models to production environments.

C++ Deployment¶

Undefined Symbol Errors¶

Solution: Match LibTorch ABI version:

# Check compiler version
g++ --version

# Use cxx11-abi for modern compilers
wget https://download.pytorch.org/libtorch/cpu/libtorch-cxx11-abi-shared-with-deps-latest.zip

CUDA Error: No Kernel Image Available¶

Solution: Match LibTorch CUDA version:

# Check CUDA version
nvidia-smi

# Download matching LibTorch (e.g., CUDA 11.8)
wget https://download.pytorch.org/libtorch/cu118/libtorch-cxx11-abi-shared-with-deps-latest.zip

iOS Deployment¶

Module File Doesn't Exist¶

Solution: Ensure model is in app bundle: - Check "Copy items if needed" when adding to Xcode - Verify target membership in Xcode

Android Deployment¶

Native Method Not Found¶

Solution: Update PyTorch Android version:

implementation 'org.pytorch:pytorch_android:1.13.1'
implementation 'org.pytorch:pytorch_android_torchvision:1.13.1'

ONNX Deployment¶

Export Fails¶

Solutions:

# 1. Use lower opset version
torch.onnx.export(..., opset_version=11)

# 2. Simplify model (remove custom ops)

# 3. Use TorchScript instead
traced_model = torch.jit.trace(model, example_input)
traced_model.save("model.pt")

Model Size Optimization¶

Exported Model Too Large¶

Solutions:

import autotimm as at  # recommended alias
# 1. Use quantization
from autotimm.export import quantize_dynamic
quantize_dynamic("model.onnx", "model_quant.onnx")

# 2. Use smaller backbone
model = ImageClassifier(backbone="resnet34")  # Instead of resnet50

# 3. Use FP16 during export

Related Issues¶

• Export & Inference - Export problems
• Device Errors - Hardware issues