Apppublic

nvidia/Cosmos3-Action-Viewer-Prerelease

sourceHugging Faceupdated 3mo agoView on Hugging Face

1likes

App README

Cosmos3 Action Viewer Backend

ZeroGPU inference backend for the Cosmos3 Action Viewer.

This Space is not an interactive demo. The public viewer sends authenticated generation requests here and displays the generated videos in Viser.

Gradio API

—health: lightweight service health and runtime metadata.
—generate: ZeroGPU action-video generation endpoint used by the viewer.

generate uses the generic nvidia/Cosmos3-Nano checkpoint and a persistent in-process Cosmos runner. The first request on a cold ZeroGPU worker can take around four minutes while the model and checkpoint are loaded. Later requests on the same worker are usually much faster.

The checkpoint should be mounted or materialized at /data/models/nvidia--Cosmos3-Nano. Set ACTION_VIZ_REQUIRE_PREWARMED_CHECKPOINT=0 only if downloading the checkpoint during a ZeroGPU generation lease is acceptable.

Environment

—ACTION_VIZ_GENERATE_GPU_DURATION: ZeroGPU reservation duration, capped at 200 seconds for the current large-worker deployment.
—ACTION_VIZ_GENERATE_TIMEOUT_SECONDS: backend generation timeout. Defaults to ten seconds less than the GPU reservation.
—ACTION_VIZ_GENERATE_RUNNER: defaults to persistent; set to subprocess only as an operational fallback.
—ACTION_VIZ_SKIP_RANDOM_INIT: defaults to 1; wraps model construction in transformers.no_init_weights() before checkpoint loading.

Optional diagnostics are available but off by default:

—ACTION_VIZ_LOAD_INSTRUMENT=1
—ACTION_VIZ_DCP_INSTRUMENT=1

Deploy

This repository is configured with the Hugging Face Space remote:

bash

git push origin main