PyTorch for NVIDIA Jetson (JetPack 6.x)

Precompiled and optimized PyTorch package for NVIDIA Jetson devices running JetPack 6.x (L4T 36.x).
This package includes CUDA, cuDNN, and TensorRT support, built specifically for Jetson Orin platforms.

---

📦 Package Information

• Name: torch-2.5.0a0+gita8d6afb-jetson-jp6.1-a8d6afb.deb

• Architecture: arm64

• Supported Jetson boards:

  • Jetson Orin Nano
  • Jetson Orin NX
  • Jetson AGX Orin

• Supported JetPack versions:

  • JetPack 6.1 (tested)
  • JetPack 6.2 (compatible unless ABI changes)

Note:
This release includes only PyTorch (torch).
It does not include TorchVision or TorchAudio.

---

🚀 Installation

1. Download the .deb package

Download the file from the Releases section:

torch-2.5.0a0+gita8d6afb-jetson-jp6.1-a8d6afb.deb

2. Install it

sudo apt install ./torch-2.5.0a0+gita8d6afb-jetson-jp6.1-a8d6afb.deb

APT will automatically install the package into:

/usr/local/lib/python3.10/dist-packages/

(or the Python version used in your build)

---

🔍 Verify Installation

Run the following test:

python3 - << 'EOF'
import torch
print("PyTorch version:", torch.__version__)
print("CUDA available:", torch.cuda.is_available())
if torch.cuda.is_available():
    print("CUDA device:", torch.cuda.get_device_name(0))
EOF

Expected output:

PyTorch version: 2.5.x+gitXXXX
CUDA available: True
CUDA device: NVIDIA Orin (SM 8.7)

---

🧱 Requirements

• JetPack 6.x (L4T 36.x)
• Python 3.10 (or whatever version your build targets)
• CUDA / cuDNN / TensorRT (already provided by JetPack)

No extra NVIDIA downloads are required.

---

📁 Package Contents

The .deb installs:

/usr/local/lib/python3.10/dist-packages/torch/
/usr/local/lib/python3.10/dist-packages/torch-<VERSION>.dist-info/
/usr/local/lib/python3.10/dist-packages/nvidia/*

Includes:

• CUDA kernels for Jetson Orin
• ATen + cuDNN
• TensorRT-supported layers
• Custom architecture build for Orin (sm_87)

---

🧩 Notes

• FlashAttention is disabled (to prevent JetPack 6.x crashes).
• MPI, Gloo, NCCL, Distributed training → disabled.
• ABI checks disabled to match NVIDIA’s JetPack toolchain.
• Build optimized for stability + compatibility, not multi-GPU features.

---

📝 License

PyTorch is BSD-licensed.
This repository redistributes compiled binaries for developer convenience.

---

✨ Maintainer

[email protected]