hao-ai-lab-FastVideo/docs/cookbook/options.yaml

# Authored control metadata. These standard JSON Schema fragments describe the
# available fields; deployment.options selects which ones appear in a recipe.
# Config values come from the native baseline, not these informational defaults.
# CUDA declarations follow fastvideo/api/schema.py; MLX declarations follow
# fastvideo/entrypoints/openai/mlx_server.py.
# GPU count, frames, dimensions, and FPS have practical, overridable cookbook
# limits. These UI choices do not describe engine capacity or guarantee GPU fit.
# Integer transport precision is checked by the serializer, not modeled as range.
options:
  server.host:
    title: Host
    type: string
    default: 0.0.0.0
  server.port:
    title: Port
    type: integer
    default: 8000
    minimum: 1
    maximum: 65535
  server.output_dir:
    title: Output directory
    type: string
    default: outputs/
  server.served_model_name:
    # CUDA permits null; MLX requires a nonempty string. Each runtime completes it.
    title: Served model name

runtimes:
  fastvideo-cuda-rest:
    metadata:
      id: fastvideo-cuda-rest
      label: CUDA REST
      backend: cuda
      interface: rest
      launch_argv: &cuda_launch [fastvideo, serve, --config, config.yaml]
      install_url: https://haoailab.com/FastVideo/getting_started/installation/
      topology_defaults: &cuda_topology
        num_gpus: 1
        tp_size: -1
        sp_size: -1
        hsdp_replicate_dim: 1
        hsdp_shard_dim: -1
      server_defaults: &cuda_server
        host: 0.0.0.0
        port: 8000
        served_model_name: null
    workloads: [t2v, i2v]
    options: &cuda_options
      server.served_model_name:
        anyOf:
          - type: string
          - type: "null"
        default: null
      generator.engine.num_gpus:
        title: GPU count
        type: integer
        default: 1
        minimum: 1
        maximum: 8
      generator.engine.parallelism.tp_size:
        title: Tensor parallelism
        description: Splits supported text encoders across GPUs; this does not shard every model's DiT.
        type: integer
        default: -1
        minimum: -1
        not: {const: 0}
      generator.engine.parallelism.sp_size:
        title: Sequence parallelism
        type: integer
        default: -1
        minimum: -1
        not: {const: 0}
      generator.engine.parallelism.hsdp_replicate_dim:
        title: HSDP replication
        type: integer
        default: 1
        minimum: 1
      generator.engine.parallelism.hsdp_shard_dim:
        title: HSDP sharding
        type: integer
        default: -1
        minimum: -1
        not: {const: 0}
      generator.engine.offload.dit_layerwise:
        title: Offload DiT layers to CPU
        type: boolean
        default: true
      generator.engine.offload.text_encoder:
        title: Offload text encoder to CPU
        type: boolean
        default: true
      generator.engine.offload.vae:
        title: Offload VAE to CPU
        type: boolean
        default: true
      generator.engine.compile.enabled:
        title: Compile transformer
        type: boolean
        default: false
      default_request.sampling.num_frames:
        title: Frames
        type: integer
        default: 125
        minimum: 1
        maximum: 512
      default_request.sampling.height:
        title: Height
        type: integer
        default: 720
        minimum: 1
        maximum: 4096
      default_request.sampling.width:
        title: Width
        type: integer
        default: 1280
        minimum: 1
        maximum: 4096
      default_request.sampling.fps:
        title: FPS
        type: integer
        default: 24
        minimum: 1
        maximum: 120
      default_request.sampling.seed:
        title: Seed
        type: integer
        default: 1024

  fastvideo-mlx-rest:
    metadata:
      id: fastvideo-mlx-rest
      label: MLX REST
      backend: mlx
      interface: rest
      hardware_label: Apple Silicon
      launch_argv: [python, -m, fastvideo.entrypoints.openai.mlx_server, --config, config.yaml]
      install_url: https://haoailab.com/FastVideo/getting_started/installation/mlx/
      server_defaults:
        host: 127.0.0.1
        port: 8000
        served_model_name: fasth3
    workloads: [t2v]
    options:
      server.host:
        default: 127.0.0.1
      server.port:
        minimum: 1
        maximum: 65535
      server.output_dir:
        default: outputs/mlx_fasth3
      server.served_model_name:
        type: string
        minLength: 1
        default: fasth3
      generator.model_root:
        title: Model directory
        type: string
      generator.mlx_checkpoint:
        title: MLX checkpoint
        type: string
      generator.prompt_cache_dir:
        title: Prompt cache directory
        type: string
        default: outputs/h3_prompt_cache
      generator.vae_dtype:
        title: VAE precision
        type: string
        enum: [fp32, fp16, bf16]
        default: fp32
      generator.vsa_sparsity:
        title: VSA sparsity
        type: number
        minimum: 0.0
        exclusiveMaximum: 1.0
        default: 0.9
      default_request.sampling.num_frames:
        title: Frames
        type: integer
        minimum: 108
        maximum: 362
        description: Configured counts align upward to the next 17n+5 value; explicit HTTP counts must already be aligned.
      default_request.sampling.height:
        title: Height
        type: integer
        minimum: 32
        maximum: 4096
        multipleOf: 32
        description: H3 serving requires multiples of 32; its canvas-area limit also applies.
      default_request.sampling.width:
        title: Width
        type: integer
        minimum: 32
        maximum: 4096
        multipleOf: 32
        description: H3 serving requires multiples of 32; its canvas-area limit also applies.
      default_request.sampling.seed:
        title: Seed
        type: integer
        minimum: 0
        maximum: 4294967295

  fastvideo-cuda-streaming:
    metadata:
      id: fastvideo-cuda-streaming
      label: CUDA streaming
      backend: cuda
      interface: websocket
      launch_argv: *cuda_launch
      install_url: https://haoailab.com/FastVideo/getting_started/installation/
      topology_defaults: *cuda_topology
      client_guide_url: ../../design/server_contracts/streaming/
      server_defaults: *cuda_server
    workloads: [t2v]
    options:
      <<: *cuda_options
      streaming.session_timeout_seconds:
        title: Client message timeout (seconds)
        type: integer
        minimum: 1
        default: 300
        description: Time allowed for client initialization, an idle client message, or worker acquisition; not an ongoing-generation deadline.
      streaming.generation_segment_cap:
        title: Maximum segments per session
        type: integer
        minimum: 1
        default: 6
        description: Stop the session after this many generated segments.