live/pipelines: Move prepare models logic to pipeline (#849)

* .github: Rename docker build actions for consistency

* Add scope pipeline scaffold for live-video-to-video

This commit adds the initial scaffold for the Scope pipeline integration,
following the same patterns as StreamDiffusion (not ComfyUI). The scaffold
includes:

Pipeline Implementation:
- Created runner/app/live/pipelines/scope/ with __init__.py, pipeline.py,
  and params.py
- ScopeParams extends BaseParams for future configuration
- Scope pipeline class based on Noop implementation
- Updated loader.py to register scope pipeline and params

Docker & CI:
- Added Dockerfile.live-base-scope that installs scope via uv in the
  comfystream conda environment and builds the frontend
- Created .github/workflows/ai-runner-docker-scope.yaml to build and push
  scope Docker images (live-base-scope and live-app-scope)

Documentation:
- Updated docs/live-ai-local-dev.md examples to use scope instead of
  comfyui

The implementation follows StreamDiffusion patterns and does not include
any ComfyUI-style special cases. Scope is treated as a standard pipeline
without parameter wrapping or resolution restrictions.

* docker: Fix scope package installation

* Fix --active flag to uv

Also adda a check on pipe startup

* Add scope model prepare tooling

* Fix scope prepare script path handling

* Add diffusers deps to live AI requirements

* Pin uv to latest version 0.9.9

* Fix scope pipeline interface import path

* Add /workspace/scope to PYTHONPATH for lib module imports

* Remove sys.path manipulation from prepare_models, rely on PYTHONPATH

* feat: unify pipeline model preparation

Add a shared prepare_models entrypoint that loads pipelines via loader.load_pipeline and lets each pipeline expose prepare_models(). StreamDiffusion now owns a full preparation module that downloads dependencies, compiles Depth/Pose/RAFT engines, and iterates over all model/controlnet/IP-Adapter combinations to build TensorRT engines sequentially. Scope implements the hook with its smoke test, ComfyUI/Noop have no-ops, and the download scripts call the new tool so the old bespoke streamdiffusion/scope helpers are removed.

* Human simplify dl_checkpoints

* Human plane proof read (WIP)

* [ai] Stop passing model_dir around

* Rename cwd_models to just StreamDiffusion--models

* Review rest of prepare.py

* Fix ipa_type

Author: victorges

.github/workflows/ai-runner-docker-live-noop.yaml

@@ -6,6 +6,7 @@ on:
       - "runner/docker/Dockerfile.live-*"
       - "runner/app/**"
       - "runner/images/**"
+      - "runner/VERSION"
   push:
     branches:
       - main
@@ -15,6 +16,7 @@ on:
       - "runner/docker/Dockerfile.live-*"
       - "runner/app/**"
       - "runner/images/**"
+      - "runner/VERSION"
   workflow_dispatch:
 
 concurrency:

.github/workflows/ai-runner-docker-live-streamdiffusion.yaml

@@ -16,6 +16,7 @@ on:
       - "runner/docker/Dockerfile.live-*"
       - "runner/app/**"
       - "runner/images/**"
+      - "runner/VERSION"
   workflow_dispatch:
 
 concurrency:

.github/workflows/ai-runner-docker-scope.yaml

@@ -16,6 +16,7 @@ on:
       - "runner/docker/Dockerfile.live-*"
       - "runner/app/**"
       - "runner/images/**"
+      - "runner/VERSION"
   workflow_dispatch:
 
 concurrency:

runner/app/live/pipelines/comfyui/pipeline.py

@@ -1,8 +1,9 @@
-import os
-import torch
 import asyncio
 import logging
+import os
+from pathlib import Path
 
+import torch
 from comfystream.client import ComfyStreamClient
 
 from ..interface import Pipeline
@@ -103,3 +104,7 @@ async def stop(self):
         finally:
             self.client = None
             logging.info("ComfyUI pipeline stopped")
+
+    @classmethod
+    def prepare_models(cls):
+        logging.info("ComfyUI prepare_models is currently a no-op")

runner/app/live/pipelines/interface.py

@@ -1,6 +1,9 @@
 from asyncio import Task
 from abc import ABC, abstractmethod
+from pathlib import Path
+
 from pydantic import BaseModel, Field
+
 from ..trickle import VideoFrame, VideoOutput, DEFAULT_WIDTH, DEFAULT_HEIGHT
 
 
@@ -104,3 +107,9 @@ async def stop(self):
         Called once when the pipeline is no longer needed.
         """
         pass
+
+    @classmethod
+    @abstractmethod
+    def prepare_models(cls):
+        """Download and/or compile any assets required for this pipeline."""
+        pass

runner/app/live/pipelines/noop.py

@@ -1,5 +1,6 @@
-import logging
 import asyncio
+import logging
+from pathlib import Path
 
 from .interface import Pipeline
 from ..trickle import VideoFrame, VideoOutput
@@ -24,3 +25,7 @@ async def update_params(self, **params):
 
   async def stop(self):
     logging.info("Stopping pipeline")
+
+  @classmethod
+  def prepare_models(cls):
+    logging.info("Noop pipeline does not require model preparation")

runner/app/live/pipelines/scope/pipeline.py

@@ -1,5 +1,11 @@
+<<<<<<< HEAD
+import asyncio
+import logging
+from pathlib import Path
+=======
 import logging
 import asyncio
+>>>>>>> main
 
 from ..interface import Pipeline
 from ...trickle import VideoFrame, VideoOutput
@@ -37,3 +43,16 @@ async def stop(self):
         # clear the frame queue
         self.frame_queue = asyncio.Queue()
 
+    @classmethod
+    def prepare_models(cls):
+        logging.info("Preparing Scope models")
+        try:
+            from lib.schema import HealthResponse  # type: ignore
+
+            assert HealthResponse is not None
+        except ImportError as exc:
+            raise RuntimeError(
+                "Scope Python dependencies are not installed inside this image."
+            ) from exc
+        logging.info("Scope model preparation complete")
+

runner/app/live/pipelines/streamdiffusion/pipeline.py

@@ -3,6 +3,7 @@
 import asyncio
 import base64
 import re
+from pathlib import Path
 from typing import Dict, List, Optional, Any, cast
 
 import torch
@@ -23,6 +24,9 @@
     LCM_LORAS_BY_TYPE
 )
 
+ENGINES_DIR = Path("./engines")
+# this one is used only for realesrgan_trt which has ./models hardcoded
+LOCAL_MODELS_DIR = Path("./models")
 
 class StreamDiffusion(Pipeline):
     def __init__(self):
@@ -258,6 +262,12 @@ async def stop(self):
             self.params = None
             self.frame_queue = asyncio.Queue()
 
+    @classmethod
+    def prepare_models(cls):
+        from .prepare import prepare_streamdiffusion_models
+
+        prepare_streamdiffusion_models()
+
 
 def _prepare_controlnet_configs(params: StreamDiffusionParams) -> Optional[List[Dict[str, Any]]]:
     """Prepare ControlNet configurations for wrapper"""
@@ -386,7 +396,7 @@ def load_streamdiffusion_sync(
     params: StreamDiffusionParams,
     min_batch_size=1,
     max_batch_size=4,
-    engine_dir="engines",
+    engine_dir: str | Path = ENGINES_DIR,
     build_engines=False,
 ) -> StreamDiffusionWrapper:
     pipe = StreamDiffusionWrapper(