[Solution] decord==0.6.0 installation error on ARM64 / DGX Spark (Pre-compiled wheel provided)

[EvilTCell]

The Problem

When running pip install -r requirements.txt on ARM64 architecture (specifically an Nvidia DGX Spark running Ubuntu 24.04), the installation fails on the decord dependency.

Because there are no official pre-compiled ARM64 wheels for modern Python versions (3.12+), pip attempts to build from source or simply fails with the following error:

ERROR: Could not find a version that satisfies the requirement decord==0.6.0 (from versions: none)
ERROR: No matching distribution found for decord==0.6.0

Attempting to build decord from source manually on modern Ubuntu systems also fails because the source code expects older FFmpeg 4.x headers and throws pointer/casting errors when compiled against FFmpeg 6.x.

The Solution

To save others the headache of patching the C++ source code and compiling against modern FFmpeg libraries, I have forked the repo, patched the FFmpeg 6.x API deprecations, and compiled a ready-to-use .whl file for ARM64 systems.

You can download the pre-compiled wheel here:
👉 Decord 0.6.0 for ARM64 (Ubuntu 24.04 / Python 3.12)

How to use it:

1. Download the .whl file from the release page or install it directly via the URL:
   pip install https://github.com/chappa-ai-llc/decord/releases/download/v0.6.0-ubuntu24.04-arm64/decord-0.6.0-cp312-cp312-linux_aarch64.whl

2. Once decord is installed, you can safely run the Void Model requirements installation:
   pip install -r requirements.txt

System Requirements for this specific wheel:

- Architecture: ARM64 / aarch64
- OS: Ubuntu 24.04
- Python: 3.12.x
- FFmpeg: 6.1.x

Hopefully, this unblocks anyone else trying to spin up void-model on a DGX or similar ARM hardware!