A/B confirmed: the guard's in-process forward pass kills subsequent GPU worker forks — restore the subprocess guard

app.py

@@ -73,22 +73,13 @@ PIPE = None
 MANAGER = None
 LOAD_ERROR: str | None = None
 LOADED_IN: float | None = None
-GUARD = None
-GUARD_REPO = "hfmlsoc/ncii-light-guard-v01"
-
-
-def load_guard() -> None:
-    """A 270M CPU text classifier scoring the NCII risk of the prompt on requests carrying a keyframe — the
-    edit-on-a-real-photo case. It runs before the conditioner call and the denoise booking, so a refused
-    prompt costs no GPU time on either half."""
-    global GUARD
-    from transformers import pipeline
-
-    GUARD = pipeline("text-classification", model=GUARD_REPO, device="cpu")
-
-
 def check_prompt(prompt: str) -> None:
-    flag = GUARD(prompt, truncation=True)[0]
+    """The NCII guard, on requests carrying a keyframe — the edit-on-a-real-photo case. It runs before the
+    conditioner call and the denoise booking, so a refused prompt costs no GPU time on either half. The
+    classifier lives in `ncii_guard`'s spawned subprocess — in the main process it kills every GPU worker."""
+    import ncii_guard
+
+    flag = ncii_guard.classify(prompt)
     if flag["label"] == "ncii":
         print(f"[guard] prompt refused (ncii {flag['score']:.2f}): {prompt!r}", flush=True)
         raise gr.Error("This prompt was flagged by a content filter and wasn't run.")
@@ -354,7 +345,9 @@ def _fit_keyframe(image_path, current_canvas):
     return gr.update(value=image_path), gr.update(value=label)
 
 
-load_guard()
+import ncii_guard
+
+ncii_guard.start()
 load_models()
 
 INTRO = """# MiniMax-H3