Skip to content

COLAB.md

Latest commit

 

History

History
executable file
·
101 lines (72 loc) · 4.31 KB

COLAB.md

File metadata and controls

executable file
·
101 lines (72 loc) · 4.31 KB
Raw

Running DedrisGenAI on Google Colab

DedrisGenAI can run on a free Google Colab GPU, giving you the full PHP web UI backed by a CUDA GPU — no local install required.

Open the notebook: engine/dedrisgenai_colab.ipynb · Open in Colab

One-click run

  1. Click the Open in Colab badge above.
  2. Set a GPU runtime: Runtime → Change runtime type → Hardware accelerator: GPU.
  3. Runtime → Run all.
  4. When the last cells finish, click the printed link to open the UI in a new tab.

That's it. The notebook clones the repo, installs everything, starts both services, and surfaces a public URL.

What the notebook does

  1. Clones the repository into /content/DedrisGenAI.
  2. Installs the PHP CLI (Colab does not ship PHP) and the CUDA build of PyTorch plus the engine's Python requirements.
  3. Starts the engine (engine/server.py, CWD = engine/) on 127.0.0.1:7866 and waits for /api/health.
  4. Starts the PHP web UI on 127.0.0.1:8888 (php -S ... webui/public/router.php).
  5. Exposes the UI through Colab's port proxy and prints a clickable link. A cloudflared tunnel fallback is included.

The cells are idempotent: re-running them will not start a second engine or UI if one is already healthy.

Troubleshooting

Make sure you are on a GPU runtime

Before running anything, set Runtime → Change runtime type → Hardware accelerator: GPU, then Runtime → Run all. The dependency cell prints whether CUDA is visible:

torch 2.x.x | CUDA available: True
GPU: Tesla T4

If it prints CUDA available: False, you are on a CPU runtime — switch to GPU and re-run. DedrisGenAI works on CPU only in theory; generation would be impractically slow.

First run is slow (model download)

The first engine start downloads the default SDXL model (several GB) and primes caches before /api/health reports ready. This commonly takes a few minutes; the start cell polls and prints ...waiting for engine (downloading model on first run) while it works. This only happens once per session. Watch the live engine log with:

!tail -n 50 /content/dedris_engine.log

If the engine process exits early, the cell prints the last log lines and stops — read them to see what failed (usually a download timeout; just re-run the cell).

Enhance / auto-mask features may be unavailable (rembg / cupy)

The Enhance tab and automatic mask generation depend on rembg, which pulls in pymatting → cupy, plus the SAM / GroundingDINO models. Colab ships a preinstalled cupy that is frequently ABI-incompatible with NumPy, so importing that chain can fail.

This is non-blocking for text-to-image: the engine imports those heavy modules lazily (only when a mask is actually generated), so the engine starts and plain generation works regardless. If you try the Enhance/auto-mask feature and it errors on Colab, that is the expected limitation — stick to standard generation, or run DedrisGenAI locally where the runtime is provisioned cleanly.

The Colab proxy link does not open

Some browsers, extensions, or corporate networks block Colab's proxyPort links. Use the cloudflared tunnel fallback cell instead — it downloads cloudflared and opens a temporary public https://<random>.trycloudflare.com URL to the UI. Treat that URL as public for the life of the session.

UI loads but generation fails / "engine unreachable"

The PHP UI proxies to the engine on port 7866. If the UI is up but calls fail, the engine probably isn't healthy yet (still downloading the model) — wait for the engine start cell to finish, then reload the UI. Check both logs:

!tail -n 50 /content/dedris_engine.log
!tail -n 50 /content/dedris_ui.log

Ports

Defaults are UI 8888 and engine 7866. Override before the start cells by setting DEDRIS_UI_PORT / DEDRIS_ENGINE_PORT in the environment (e.g. os.environ['DEDRIS_UI_PORT']='9000').

Session limits

Colab runtimes are temporary. When the session ends (idle timeout, closed tab, or the free-tier limit), the runtime — and both services — stop, and downloaded models are discarded. Re-running the notebook re-downloads them. Keep the tab open while you work.