Skip to content

README.md

Latest commit

 

History

History
58 lines (45 loc) · 3.28 KB

README.md

File metadata and controls

58 lines (45 loc) · 3.28 KB
Raw

TRELLIS on Modal

This is a deployment wrapper around litereality_agent.models.trellis.inference; model inference remains in src/ and the Modal runtime only supplies a GPU, parallel execution, and a persistent weight cache.

Modal associates a deployment with the workspace belonging to the active profile; the workspace cannot be selected by app source code. Confirm which one you are about to bill, then deploy into its main environment:

modal profile list
modal profile activate <name>
modal profile current

modal deploy --env main deploy/modal/trellis/app.py

No Hugging Face account is needed

TRELLIS.2's image encoder is the gated facebook/dinov3-vitl16-pretrain-lvd1689m, which is why this app previously required a huggingface Modal secret. It no longer does. The DINOv3 licence grants the right to "distribute, copy" the trained weights to third parties provided a copy of the licence travels with them, so the image bakes in an ungated mirror instead:

  • repo camenduru/dinov3-vitl16-pretrain-lvd1689m
  • revision 3c276edd87d6f6e569ff0c4400e086807d0f3881
  • model.safetensors sha256 dcb2e45127cccbf1601e5f42fef165eea275c8e5213197e8dcf3f48822718179

The build verifies that digest and fails if it does not match, so a mirror force-push cannot silently swap the weights. Four independent uploaders publish byte-identical copies under that same digest, which is how the weights were confirmed unmodified without holding gate access. Always pin by revision. resolve_dinov3() finds the baked copy through LITEREALITY_DINOV3 and repoints pipeline.json at it, so loading never reaches the gated repo.

microsoft/TRELLIS.2-4B is MIT and ungated, so no token is needed for it either.

The base CUDA image, OS packages, PyTorch, and Python dependencies build on CPU. Modal attaches an H100 only to the separately cached native-extension layer required by the upstream TRELLIS installer; the deployed inference function also uses an H100. Source-only application changes do not rebuild the extension layer. The function is intentionally limited to one container, performs zero application retries, and scales that container down after ten idle seconds to bound spend.

The weights are hosted on Hugging Face, but the model is not currently served by a Hugging Face Inference Provider. The first invocation downloads microsoft/TRELLIS.2-4B into the litereality-trellis-models Volume; later containers reuse that cache. Note that this download runs inside the gpu="H100" function, so several GB transfer at H100 prices — see the known inefficiencies in ../README.md.

Configure the application by setting MODAL_PROFILE to your profile name; the app name, function name, and environment already default. Install client support with uv sync --extra modal.

Measured on a fresh workspace, 2026-08-06: the image built in 11.3 minutes for about $0.45, split almost evenly in time between the CPU dependency layers (318.0s) and the H100 extension compile (318.4s), with the GPU layer carrying roughly 78% of the cost. Redeploy after a source-only change took 2.8 seconds. The first request returned a valid 512-texture GLB from ChairCluster0 — 33,898 vertices, 47,793 faces — in 295.7s including the weights download; later requests skip it. Neither building nor inference is part of the offline test suite.