INFERENCE_CONTRACT.json

{
  "schema_version": "inference_contract.v1",
  "full_inference_implemented": true,
  "inference_strategy": "diffusers_pipeline",
  "health_endpoint": "/health",
  "primary_api_name": "/generate",
  "expected_output_type": "video",
  "validation_level": "smoke_test_hardware_limited",
  "requires_gpu": true,
  "recommended_target_space_hardware": "manual_required",
  "estimated_vram": "50_plus_gb",
  "blockers_count": 0,
  "model_family": "diffusers_full_pipeline",
  "artifact_role": "full_model",
  "loader_strategy": "DiffusionPipeline_or_family_specific_from_pretrained",
  "app_runtime_contract_version": "app_runtime_contract.v198_26_30",
  "health_required_fields": [
    "status",
    "model_ready",
    "pipeline_ready",
    "model_family",
    "loader_strategy",
    "last_error",
    "expected_output_type"
  ],
  "generate_not_ready_behavior": "raises RuntimeError containing the concrete pipeline load failure or a clear 'Pipeline not loaded' message with last_error included",
  "recipe_conformance_notes": [
    "Preserves model_family 'diffusers_full_pipeline' and loader_strategy from MODEL_RECIPE.",
    "Implements ModularPipeline.from_pretrained + pipe.load_components exactly as the model card instructs, with a documented ZeroGPU-safe fallback for device_map.",
    "Health endpoint is cheap (reads global state only) and exposes all required contract fields.",
    "Generate attempts a real forward pass through 14B transformer + UMT5 text encoder + Wan VAE if pipeline loaded successfully."
  ]
}