fix

app.py

@@ -9,6 +9,15 @@ from video_utils import (
     merge_seconds_union, redact_with_ffmpeg
 )
 
+import os
+try:
+    from huggingface_hub import login
+    tok = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_HUB_TOKEN")
+    if tok: login(tok)
+except Exception:
+    pass
+
+
 APP_TITLE = "Content Moderation Demo (Image & Video)"
 APP_DESC = """
 Minimal prototype: image/video analysis, model & category selection, and threshold control.

requirements.txt

@@ -9,3 +9,5 @@ gradio>=4.44.0
 opencv-python-headless
 ffmpeg-python
 timm
+sentencepiece
+

src/models/clip_lp.py

@@ -10,6 +10,7 @@ class CLIPLinearProbe:
     def __init__(self, head_path):
         self.device = "cuda" if torch.cuda.is_available() else "cpu"
         self.torch_dtype = torch.float16 if self.device == "cuda" else torch.float32
+
         self.model, _, self.preprocess = open_clip.create_model_and_transforms(
             "ViT-L-14", pretrained="openai", device=self.device
         )
@@ -18,6 +19,7 @@ class CLIPLinearProbe:
         self.w = torch.from_numpy(npz["w"]).to(self.device).float()
         self.b = torch.from_numpy(npz["b"]).to(self.device).float()
 
+        self.use_amp = False
         if self.device == "cuda":
             torch.backends.cuda.matmul.allow_tf32 = True
             torch.backends.cudnn.benchmark = True

src/models/clip_multilabel.py

@@ -16,7 +16,7 @@ class CLIPMultiLabel:
         self.model.eval().requires_grad_(False)
 
         ckpt = torch.load(head_path, map_location=self.device, weights_only=True)
-        state = ckpt.get("model_state", ckpt)
+        state = ckpt.get("state_dict", ckpt)
 
         w = state["head.weight"].to(self.device).float()
         b = state["head.bias"].to(self.device).float()
@@ -24,6 +24,7 @@ class CLIPMultiLabel:
 
         self.w, self.b = w, b
 
+        self.use_amp = False
         if self.device == "cuda":
             torch.backends.cuda.matmul.allow_tf32 = True
             torch.backends.cudnn.benchmark = True

src/models/eva_headpreserving.py

@@ -21,7 +21,6 @@ class EVAHeadPreserving:
                  tag_csv: str = "selected_tags.csv"):
         self.device = "cuda" if torch.cuda.is_available() else "cpu"
         self.torch_dtype = torch.float16 if self.device == "cuda" else torch.float32
-        self.use_amp = (self.device == "cuda")
 
         self.categories = list(categories)
         self.tag_csv = tag_csv
@@ -54,9 +53,11 @@ class EVAHeadPreserving:
         with torch.no_grad():
             self.custom_head.weight.copy_(w)
             self.custom_head.bias.copy_(b)
+            self.use_amp = True
 
         self.tag_names = load_tag_names(T, self.tag_csv)
 
+        self.use_amp = False
         if self.device == "cuda":
             torch.backends.cuda.matmul.allow_tf32 = True
             torch.backends.cudnn.allow_tf32 = True

src/models/siglip_lp.py

@@ -19,6 +19,7 @@ class SigLIPLinearProbe:
         self.w = torch.from_numpy(npz["w"]).to(self.device).float()
         self.b = torch.from_numpy(npz["b"]).to(self.device).float()
 
+        self.use_amp = False
         if self.device == "cuda":
             torch.backends.cuda.matmul.allow_tf32 = True
             torch.backends.cudnn.benchmark = True