nvidia

/

Cosmos3-Edge

like 93

Follow

NVIDIA 63.7k

Aligns the Edge card Generation examples to the Nano/Super ### vLLM-Omni convention, replacing the native-PyTorch examples with vLLM-Omni serving examples.

Changes (Usage → Generation only):

• Native cosmos_framework.scripts.inference examples (Image-to-video, Action forward-dynamics, Action inverse-dynamics) → ### vLLM-Omni (Container + General Invocation + Examples), mirroring Nano/Super.
• Removes the cosmos-framework ### Setup section (siblings have none).
• Output assets regenerated on the current vllm/vllm-omni:cosmos3 container. I2V uses the negative prompt; FD/ID follow the sibling convention.

Assets: README.md + 6 regenerated outputs (edge_i2v_output.mp4 [neg-ON], edge_action_fd_umi_2chunk_output.mp4, edge_action_id_av_{0,1}_output.{json,png}).

⚠️ Cross-PR dependency — MUST co-merge: this PR removes ### Setup, which the native ### Reasoning currently depends on. Merge together with the reasoning owner's PR converting ### Reasoning → ### vLLM (its own install). Do not merge this PR alone.

Verification: byte-confined to the Generation region (Reasoning + all other sections byte-identical); example params match the client that produced the embedded outputs; all referenced assets exist in-repo; 3 independent reviews + full mechanical checks passed.