Add a one-click Colab WebUI with a CUDA T4 release build

[DrewThomasson]

Summary

• add a single-cell Google Colab notebook that launches the native WebUI through a public Cloudflare tunnel
• add a release job for an Ubuntu 22.04, CUDA 12.8, compute capability 7.5 server binary
• discover the Colab asset from the latest official GitHub release and verify its published SHA-256 digest
• fall back to a source build when the official prebuilt is unavailable or the runtime is incompatible

Release build

The Colab binary targets NVIDIA T4 (sm_75) and statically links the GCC C++ runtimes for compatibility with the Colab Ubuntu environment. CI checks that the executable has no missing libraries or dynamic libstdc++/libgcc_s dependency before packaging it. The normal release aggregation job publishes the resulting audio-v*-bin-ubuntu-x64-cuda12.8-colab.tar.gz asset.

Validation

• tested the equivalent packaged binary with an NVIDIA GPU in a clean CUDA 12.8 / Ubuntu 22.04 container
• confirmed the CUDA backend health response and WebUI HTTP 200 response
• validated notebook JSON, Python syntax, T4 accelerator metadata, and release workflow YAML
• verified source compilation remains available as a fallback

[DrewThomasson]

This is confirmed working on my end and in my personal testing of it

Google Colab free NVIDIA T4 works (CPU should work too but is too slow)

Building on the Google Colab natively takes too long (1+ hours), I have solved this by having your workflow supply the Colab build for users

Now it launches the Colab in under a minute

[DrewThomasson]

Omnivoice result in google Colab on Free Nvidia T4 GPU

ab0ad835-44e7-419e-9706-06d6c4efb899.mp4

[0xShug0]

Thanks @DrewThomasson! I will test tomorrow.