# Benchmark deploy config: LingBot-World 2.0 on four GPUs, Ulysses SP-4, compiled.
#
# This is the realtime topology. Two things differ from the single-GPU config,
# and both matter to the number the benchmark reports:
#
#   parallel_config   Pure Ulysses sequence parallelism over four ranks. The
#                     pipeline validates this: LingBot requires
#                     sequence_parallel_size == ulysses_degree, ring_degree and
#                     allgather_degree of 1, ulysses_mode "strict", and
#                     ulysses_a2a_permute disabled. Pipeline, CFG, and VAE
#                     parallel sizes above 1 are rejected. This config uses
#                     TP=1; the validator does not reject larger TP sizes.
#   enforce_eager     false puts LingBotDMDBlockRunner on the compiled path
#                     (regional torch.compile with cudagraph step marking). The
#                     first rollout therefore pays compile and capture cost --
#                     use --warmup-sessions 1 before measured repetitions.
#                     Chunk warmup only excludes the attention-window ramp.
#
# Serve it under a GPU reservation, since it needs four cards:
#
#   gpu run --gpus 4 -- vllm serve robbyant/lingbot-world-v2-14b-causal-fast-diffusers \
#     --omni --deploy-config benchmarks/lingbot_world/configs/usp4_compiled.yaml \
#     --port 8000
#
# `devices` is the rank->card mapping inside the reservation, so it stays
# "0,1,2,3" no matter which physical cards CUDA_VISIBLE_DEVICES selected.

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

stages:
  - stage_id: 0
    devices: "0,1,2,3"
    max_num_seqs: 1
    enforce_eager: false
    model_class_name: LingBotWorldCausalDMDPipeline
    engine_backend: vllm_omni.experimental.ar_diffusion.engine.ARDiffusionEngine
    step_execution: true
    streaming_output: true
    parallel_config:
      tensor_parallel_size: 1
      sequence_parallel_size: 4
      ulysses_degree: 4
      ring_degree: 1
    default_sampling_params:
      num_inference_steps: 4
      output_type: pil
      extra_args:
        flow_shift: 5.0
    model_config:
      ar_diffusion_height: 480
      ar_diffusion_width: 832
      ar_diffusion_kv_config:
        # Each rank holds its own shard of the paged KV cache, so the per-card
        # fraction can stay at the single-GPU value; lower it if the four-rank
        # start OOMs on smaller cards.
        gpu_memory_fraction: 0.6
        warmup_cudagraph: true