feat: run Comfy+weights bootstrap at Space import; ANIMA_SKIP_STARTUP_BOOTSTRAP for tests

.pytest_cache/v/cache/nodeids

@@ -1,4 +1,5 @@
 [
+  "tests/test_app_import.py::test_import_app_with_startup_skipped",
   "tests/test_generation_smoke.py::test_comfy_entrypoint_receives_params",
   "tests/test_generation_smoke.py::test_workflow_includes_ksampler_and_split_loaders",
   "tests/test_validation.py::test_cfg_respects_step",

README.md

@@ -23,10 +23,11 @@ This Hugging Face **Gradio** Space replaces the ComfyUI canvas with a focused UI
 
 ## Runtime behavior
 
-1. On first generation, the app bootstraps **ComfyUI** into `ANIMA_COMFY_ROOT` (default: `./ComfyUI` under the Space working directory) if `main.py` is missing.
-2. Model files are downloaded with the same **retry / min-size / `.part` streaming** strategy as the original Anima Comfy Space (`circlestone-labs/Anima`, optional `Bedovyy/Anima-FP8`, Civitai RDBT URL).
+1. **On Space startup** (while the container is still in “Starting…”), the app imports, then **blocks** until Comfy is cloned, dependencies are installed, model files are downloaded, and the headless Comfy server is responding on the loopback port. The Gradio **HTTP server** starts only after this import finishes, so the **Generate** action does not wait on downloads.
+2. Model files use the same **retry / min-size / `.part` streaming** strategy as the original Anima Comfy Space (`circlestone-labs/Anima`, optional `Bedovyy/Anima-FP8`, Civitai RDBT URL).
 3. ComfyUI listens on **`ANIMA_COMFY_HOST`:`ANIMA_COMFY_PORT`** (default `127.0.0.1:8188`). Gradio serves on **`PORT`** (Hugging Face sets this; default **7860**).
 4. Inference is wrapped in **`@spaces.GPU`** with a **dynamic duration** based on steps × batch (cap `ANIMA_MAX_GPU_DURATION`, default 300s).
+5. For **tests** or local hacking without large downloads, set **`ANIMA_SKIP_STARTUP_BOOTSTRAP=1`** to defer bootstrap to the first generation (or to skip the import-time block entirely).
 
 ### Backend limitation & fallback
 
@@ -54,6 +55,7 @@ RDBT checkpoint on disk: **`rdbt-anima-p3-v024f-16step-dmd2.safetensors`** (plus
 | Variable | Purpose |
 |----------|---------|
 | `PORT` | Gradio listen port (HF sets this; usually **7860**). |
+| `ANIMA_SKIP_STARTUP_BOOTSTRAP` | Set to `1` to **skip** import-time Comfy/weights setup (e.g. unit tests, local dev). First **Generate** then runs the full `ensure_prepared()` path. |
 | `SKIP_CIVITAI` | Set to `1` to skip Civitai download of the RDBT file (only if you provide that file another way). |
 | `CIVITAI_TOKEN` | Optional bearer token for Civitai API downloads. |
 | `ANIMA_COMFY_ROOT` | Path to ComfyUI checkout (default `./ComfyUI`). |

app.py

@@ -208,7 +208,8 @@ def build_ui() -> gr.Blocks:
                     height=600,
                 )
                 status = gr.Markdown(
-                    "Ready. **First run** may take several minutes (ComfyUI clone + model downloads)."
+                    "ComfyUI, Python deps, and model weights are prepared **while the Space is starting** "
+                    "(see build logs if startup is slow). **Generate** runs inference once the UI is up."
                 )
                 go = gr.Button("Generate", variant="primary")
         go.click(
@@ -225,6 +226,13 @@ def build_ui() -> gr.Blocks:
 demo = build_ui()
 demo.queue()
 
+# Prepare Comfy + weights at container import time so the first HTTP request does
+# not pay download cost; the Space "Starting…" state covers this (skip in tests).
+if not config.skip_startup_bootstrap():
+    from src.comfy_backend import run_at_container_startup
+
+    run_at_container_startup()
+
 # Gradio 6: `theme` is a launch() argument. ZeroGPU calls `demo.launch()` without
 # running our `if __name__` block, so wrap to always default the Soft theme.
 _base_launch = demo.launch

src/comfy_backend.py

@@ -58,6 +58,8 @@ def _start_process() -> None:
     with _lock:
         if _proc is not None and _proc.poll() is None:
             return
+        if _proc is not None and _proc.poll() is not None:
+            _proc = None
         root = config.comfy_root()
         main_py = os.path.join(root, "main.py")
         if not os.path.isfile(main_py):
@@ -106,6 +108,25 @@ def _teardown_process() -> None:
         _proc = None
 
 
+def run_at_container_startup() -> None:
+    """
+    Block until ComfyUI is cloned, deps installed, weights downloaded, and the
+    Comfy process responds on HTTP. Intended to run from app import so the
+    Gradio server starts only after the stack is ready.
+    """
+    print(
+        "[startup] Cloning ComfyUI (if needed), installing deps, downloading weights, "
+        "starting Comfy — this can take many minutes on first deploy…",
+        flush=True,
+    )
+    try:
+        ensure_prepared()
+    except Exception as e:
+        print(f"[startup] Failed: {e!s}", flush=True)
+        raise
+    print("[startup] ComfyUI and model artifacts are ready.", flush=True)
+
+
 def ensure_prepared() -> None:
     """Idempotent: clone Comfy, pip, weights, start server. Thread-safe."""
     global _prepared, _bootstrapped

src/config.py

@@ -159,6 +159,11 @@ def skip_civitai() -> bool:
     return os.environ.get("SKIP_CIVITAI", "").strip() == "1"
 
 
+def skip_startup_bootstrap() -> bool:
+    """If true, skip clone/pip/weights/Comfy at import (tests, local dev without downloads)."""
+    return os.environ.get("ANIMA_SKIP_STARTUP_BOOTSTRAP", "").strip() == "1"
+
+
 @dataclass(frozen=True)
 class GenerationParams:
     prompt: str

tests/conftest.py

@@ -0,0 +1,4 @@
+import os
+
+# Do not clone Comfy or download multi-GB weights when tests import `app` (if they do).
+os.environ.setdefault("ANIMA_SKIP_STARTUP_BOOTSTRAP", "1")

tests/test_app_import.py

@@ -0,0 +1,6 @@
+"""Ensure app.py imports in CI without running multi-GB bootstrap."""
+
+def test_import_app_with_startup_skipped() -> None:
+    import app  # noqa: F401
+
+    assert app.demo is not None