fix: launch Gradio before load_models so Space leaves BUILDING

The ~77GB checkpoint download ran before demo.launch(), so HF never got a
healthcheck response and stayed on Building. Load models in a background
thread after the module-level demo is constructed; show a status banner.

app.py

@@ -136,9 +136,37 @@ from lora.sources.civarchive import (  # noqa: F401
 )
 
 
-# ── Side-effect: kick off the model load on import unless tests say otherwise ──
+# ── Startup ───────────────────────────────────────────────────────────────────
+# IMPORTANT: do NOT call load_models() at import time. Hugging Face marks the
+# Space as BUILDING until the Gradio HTTP server accepts connections. The H3
+# checkpoint is ~77 GB; loading it before launch() left the Space stuck on
+# "Building" for the entire download. Load models after the UI is up instead.
+
+def _startup_load():
+    """Load transformer + VAEs once the Gradio server is listening."""
+    if os.environ.get("H3_TEST_MODE"):
+        return status()
+    err = load_models()
+    return status() if not err else f"**Load error:** {err}"
+
+
+# Module-level `demo` so the Spaces Gradio runtime can find it (not only under
+# ``if __name__ == "__main__"``).
+demo = build_ui()
+
+# Wire a load callback if the UI exposes a status banner; otherwise load in a
+# background thread so launch() is never blocked.
+import threading as _threading
+
+def _bg_load():
+    try:
+        _startup_load()
+    except Exception as _e:
+        from state import _log
+        _log(f"background load_models failed: {_e}", "ERROR")
+
 if not os.environ.get("H3_TEST_MODE"):
-    load_models()
+    _threading.Thread(target=_bg_load, name="h3-load-models", daemon=True).start()
 
 
 if __name__ == "__main__":
@@ -150,5 +178,4 @@ if __name__ == "__main__":
     #   physical merge→denoise→unmerge critical section so shared weights stay clean.
     # - There is no application-level FIFO job queue; task IDs are isolation/diagnostic
     #   handles, not queue tickets.
-    demo = build_ui()
     demo.launch(show_error=True, theme=gr.themes.Base(), css=CSS)

ui.py

@@ -10,6 +10,7 @@ import os
 
 import gradio as gr
 
+from model import status as _model_status
 from config import (
     CANVASES,
     DEFAULT_CANVAS,
@@ -254,6 +255,9 @@ def build_ui():
     """Construct and return the Gradio `Blocks` instance."""
     with gr.Blocks(title="minimax h3 turbo + lora's") as demo:
         gr.Markdown(INTRO)
+        status_banner = gr.Markdown(
+            value="**Loading model weights** (transformer + VAEs, ~77 GB). The UI is up; generation waits until load finishes — watch Space logs.",
+        )
 
         with gr.Row():
             with gr.Column():
@@ -438,16 +442,20 @@ def build_ui():
             [],
         )
 
-        # ── Refresh community dropdown after startup ────────────────────────────────
-        # _community_lora_choices() runs at build time when the UI is constructed.
-        # The bucket IS mounted by then (load_models() runs at line 2303), but the
-        # choices may still be empty if the bucket listing is slow.  This load()
-        # callback re-populates the dropdown when the page connects.  We return a
-        # gr.Dropdown (not gr.update) — the recommended Gradio 6 pattern.
+        # ── Refresh community dropdown / status after page connect ──────────────
         def _refresh_community_dd():
-            choices = _community_lora_choices() or None
-            return gr.Dropdown(choices=choices, interactive=True)
-        demo.load(_refresh_community_dd, None, community_lora_dropdown, api_name=False)
+            choices = _community_lora_choices() or []
+            return gr.update(choices=choices)
+        def _on_page_load():
+            choices = _community_lora_choices() or []
+            try:
+                from model import status as _st
+                banner = _st()
+            except Exception as _e:
+                banner = f"Status unavailable: {_e}"
+            return gr.update(choices=choices), banner
+
+        demo.load(_on_page_load, None, [community_lora_dropdown, status_banner], api_name=False)
         # Also refresh when the user opens (focuses) the dropdown — covers cases
         # where demo.load() fires before the bucket is fully listed.
         community_lora_dropdown.focus(