Upload app.py with huggingface_hub

app.py

@@ -372,6 +372,28 @@ def decode_frames(video_path, target_fps=TARGET_FPS, max_frames=MAX_FRAMES):
 # ---------------------------------------------------------------------------
 # GPU inference.
 # ---------------------------------------------------------------------------
+def _as_tensor(out):
+    """Coerce a get_*_features() return value to a plain feature tensor.
+
+    Depending on the transformers version, get_image_features / get_text_features
+    may return a tensor directly or a ModelOutput wrapping it.
+    """
+    if isinstance(out, torch.Tensor):
+        return out
+    for attr in ("image_embeds", "text_embeds", "pooler_output", "last_hidden_state"):
+        val = getattr(out, attr, None)
+        if isinstance(val, torch.Tensor):
+            return val
+    # Fall back to first tensor element (ModelOutput is tuple-like).
+    try:
+        first = out[0]
+        if isinstance(first, torch.Tensor):
+            return first
+    except Exception:
+        pass
+    raise TypeError(f"Cannot extract feature tensor from {type(out)}")
+
+
 def _encode_siglip_images(frames, batch_size=64):
     feats = []
     for i in range(0, len(frames), batch_size):
@@ -379,7 +401,7 @@ def _encode_siglip_images(frames, batch_size=64):
         inputs = siglip_processor(images=batch, return_tensors="pt")
         pixel_values = inputs["pixel_values"].to("cuda", dtype=torch.bfloat16)
         with torch.no_grad():
-            f = siglip_model.get_image_features(pixel_values=pixel_values)
+            f = _as_tensor(siglip_model.get_image_features(pixel_values=pixel_values))
         f = f.float()
         f = f / f.norm(dim=-1, keepdim=True)
         feats.append(f.cpu())
@@ -396,7 +418,7 @@ def _encode_siglip_text(query):
     )
     input_ids = inputs["input_ids"].to("cuda")
     with torch.no_grad():
-        f = siglip_model.get_text_features(input_ids=input_ids)
+        f = _as_tensor(siglip_model.get_text_features(input_ids=input_ids))
     f = f.float()
     f = f / f.norm(dim=-1, keepdim=True)
     return f.cpu().squeeze(0)  # (D,)