# Benchmark deploy config: LingBot-World 2.0 on one GPU, eager.
#
# This is the portable smoke configuration -- it matches the shipped
# vllm_omni/deploy/lingbot_world_v2_stepwise.yaml so a number measured here is
# comparable to the default deployment. It is NOT a realtime configuration: one
# GPU runs all five forwards of every AR block serially and nothing is compiled,
# so expect a chunk interval several times the real-time deadline.
#
# Use configs/usp4_compiled.yaml for the realtime topology.
#
#   vllm serve robbyant/lingbot-world-v2-14b-causal-fast-diffusers \
#     --omni --deploy-config benchmarks/lingbot_world/configs/single_gpu_eager.yaml \
#     --port 8000

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

stages:
  - stage_id: 0
    devices: "0"
    # AR-Diffusion binds one session per runner invocation; the runner rejects
    # anything larger.
    max_num_seqs: 1
    enforce_eager: true
    model_class_name: LingBotWorldCausalDMDPipeline
    engine_backend: vllm_omni.experimental.ar_diffusion.engine.ARDiffusionEngine
    step_execution: true
    streaming_output: true
    default_sampling_params:
      # The causal-fast checkpoint pins a four-step DMD schedule; this sizes the
      # step budget before the pipeline is reached.
      num_inference_steps: 4
      output_type: pil
      extra_args:
        flow_shift: 5.0
    model_config:
      # The AR cache geometry is fixed at load time, so a benchmark request must
      # ask for exactly this resolution.
      ar_diffusion_height: 480
      ar_diffusion_width: 832
      ar_diffusion_kv_config:
        gpu_memory_fraction: 0.6
        warmup_cudagraph: true