# LingBot-World 2.0 stepwise deploy: single diffusion stage for WS /v1/realtime/video.
#
# One AsyncOmni request is one rollout: prepare_encode runs once, then every
# four DMD steps stream one AR block. Step execution keeps AR-Diffusion paged KV
# on ARDiffusionModelRunner, so this stage must stay at max_num_seqs=1.
#
# The cache geometry is fixed at load time, so ar_diffusion_height/width must
# match the height/width a client asks for.

pipeline: lingbot_world
async_chunk: false
distributed_executor_backend: mp
dtype: bfloat16

stages:
  - stage_id: 0
    devices: "0"
    max_num_seqs: 1
    enforce_eager: true
    model_class_name: LingBotWorldCausalDMDPipeline
    # LingBot runs on the AR-Diffusion engine (engine-level paged KV cache).
    engine_backend: vllm_omni.experimental.ar_diffusion.engine.ARDiffusionEngine
    # One request, many streamed chunks. streaming_output implies step_execution,
    # but both are stated here so the topology is readable on its own.
    step_execution: true
    streaming_output: true
    # num_inference_steps is for the step scheduler, not the pipeline: it sizes a
    # request's step budget before the pipeline is reached, so a served request
    # does not have to carry it. The pipeline itself pins the causal-fast
    # checkpoint's four-step DMD schedule and rejects any other count.
    default_sampling_params:
      num_inference_steps: 4
      output_type: pil
      extra_args:
        flow_shift: 5.0
    model_config:
      # Multi-rank width sharding changes floating-point numerics at shard borders.
      # Set false for full-frame VAE decode while keeping DiT sequence parallelism.
      lingbot_vae_spatial_sharding: true
      ar_diffusion_height: 480
      ar_diffusion_width: 832
      ar_diffusion_kv_config:
        gpu_memory_fraction: 0.6
        warmup_cudagraph: true