# Cosmos3 policy (DROID) deploy: single diffusion stage serving the OpenPI # realtime API (/v1/realtime/robot/openpi). # # Topology (COSMOS3_POLICY_PIPELINE) is declared in # vllm_omni/model_executor/models/cosmos3/pipeline.py. Policy checkpoints share # their HF metadata (model_type=cosmos3_omni, model_index.json _class_name= # Cosmos3OmniDiffusersPipeline) with the T2I/video Cosmos3 checkpoints, which # must keep using the default single-stage diffusion fallback — so this # pipeline cannot be auto-detected and must be selected explicitly: # # vllm serve nvidia/Cosmos3-Nano-Policy-DROID --omni \ # --deploy-config /absolute/path/to/vllm-omni/vllm_omni/deploy/cosmos3_policy_droid.yaml # # Per-checkpoint model_config tweaks (e.g. policy_server_config values) can # still be layered on top with --stage-overrides; dict overrides deep-merge # with this file. pipeline: cosmos3_policy # async_chunk defaults to True, which the deploy validator rejects for a # single-stage pipeline (no next-stage processor). Must be false here. async_chunk: false distributed_executor_backend: mp stages: - stage_id: 0 devices: "0" # OpenPI policy serving is one observation -> one action reply. max_num_seqs: 1 model_class_name: Cosmos3OmniDiffusersPipeline # JSON prompt formatting is a property of this DROID checkpoint's training # recipe, not a generic server default. default_sampling_params: extra_args: format_prompt_as_json: true model_config: # Policy serving emits robot actions, not text/video; the Cosmos3 # guardrail stack (gated nvidia/Cosmos-1.0-Guardrail repo) is not part # of it. Equivalent to launching with --no-guardrails. guardrails: false # OpenPI handshake metadata advertised to the robot client. Sent verbatim # (the pipeline never reads these); they describe the camera inputs and # action space the client must produce for this checkpoint. policy_server_config: image_resolution: [540, 640] n_external_cameras: 2 needs_wrist_camera: true needs_stereo_camera: false needs_session_id: true action_space: joint_position