Skip to content

README.md

Latest commit

 

History

History
75 lines (53 loc) · 2.93 KB

README.md

File metadata and controls

75 lines (53 loc) · 2.93 KB
Raw

Modal setup

Modal hosts the GPU-backed model functions while the pipeline stays on the calling machine. The apps remain deployed between calls, but their containers scale to zero when idle.

Modal is rented compute, not a model API: a Modal token gives you GPUs, not models, so nothing is callable until you deploy these apps into your own workspace.

One-time setup

Join the intended Modal workspace, or use your own — nothing is pinned to a particular one, and no workspace is hardcoded in source.

Install the client with uv sync --extra modal, then put your token pair from modal.com/settings/tokens into .env:

MODAL_TOKEN_ID=ak-...
MODAL_TOKEN_SECRET=as-...

Both halves are required — Modal sends them as separate request headers, and the pipeline treats half a pair as unconfigured rather than failing at the first model call.

No Hugging Face account or token is required for either app. DINO's weights are public. TRELLIS.2-4B is MIT and ungated, and TRELLIS.2's gated DINOv3 image encoder is baked into the image from a digest-verified ungated mirror — see trellis/README.md.

Deploy

uv run litereality setup

That is the whole step, and it runs unattended. It checks the client is installed, quotes the build time, and runs both deploys below with the .env tokens injected into each subprocess — no browser login, no profile to manage, and nothing to confirm.

Useful flags: --skip-deploy authenticates and records the profile only, --env <name> targets a non-default Modal environment, and --profile <name> forces a specific ~/.modal.toml profile instead of the tokens.

Deploying by hand

The two commands setup wraps, if you would rather run them yourself:

uv run modal deploy --env main deploy/modal/dino/app.py
uv run modal deploy --env main deploy/modal/trellis/app.py

These need credentials in the environment first — either the same MODAL_TOKEN_ID / MODAL_TOKEN_SECRET pair exported, or a profile:

uv run modal setup
uv run modal profile activate <profile>
export MODAL_PROFILE=<profile>

A profile works everywhere the token pair does; set MODAL_PROFILE=<profile> in .env instead of the tokens. The app names, function names, and environment already default, so that one line is enough. models/registry.py selects Modal whenever either form of credential is configured.

What deploying gets you

The first deployment builds the images; the first invocation downloads model weights into persistent Modal Volumes, so both are slow once and fast afterwards. DINO runs on an L4. TRELLIS runs on one H100 with zero application retries and a ten-second idle scale-down window. A deployed app with zero active containers does not consume GPU compute, though its stored image and Volume data remain.

See dino/README.md and trellis/README.md for model-specific details.