Upload folder using huggingface_hub

README.md

@@ -51,5 +51,11 @@ of [tidalove/cfc-track-instruction](https://huggingface.co/datasets/tidalove/cfc
 - Runs on ZeroGPU; the model is loaded in bfloat16 (~16 GB VRAM).
 - Video is sampled at 2 FPS, max 128 frames, per the model's
   `video_preprocessor_config.json`.
+- The released checkpoint ships a config mismatch: `processor_config.json` sets
+  `use_frame_special_tokens: true` (so each frame is wrapped in
+  `<frame_start>`/`<frame_end>`, matching training) while `config.json` sets it
+  `false`, which makes the model count `<im_end>` instead and abort with
+  `AssertionError: Expected 0 videos, but got 1`. `app.py` aligns the two at
+  load time.
 - A truncated answer (no closing `</tracks>`) means you hit the *Max new tokens* cap —
   raise it in **Advanced**.

app.py

@@ -248,7 +248,20 @@ def _infer(video_path: str, turns: list, max_new_tokens: int):
 # --------------------------------------------------------------------------- #
 # Gradio handlers
 # --------------------------------------------------------------------------- #
-@spaces.GPU(duration=140)
+# Measured on ZeroGPU: ~29 s for ~800 generated tokens, ~48 s for ~1600, plus
+# ~10 s to render the overlay. Runtime is dominated by decoding, so scale the
+# GPU reservation with the token budget instead of over-booking a flat number.
+def _track_duration(video_path=None, correction_hint="", max_new_tokens=1600,
+                    progress=None) -> int:
+    return int(25 + 0.028 * int(max_new_tokens or 1600))
+
+
+def _refine_duration(video_path=None, previous_tracks="", correction="",
+                     max_new_tokens=1600, progress=None) -> int:
+    return int(25 + 0.028 * int(max_new_tokens or 1600))
+
+
+@spaces.GPU(duration=_track_duration)
 def track_fish(
     video_path: str,
     correction_hint: str = "",
@@ -271,7 +284,7 @@ def track_fish(
     return _infer(video_path, [(DEFAULT_PROMPT, None)], int(max_new_tokens))
 
 
-@spaces.GPU(duration=140)
+@spaces.GPU(duration=_refine_duration)
 def refine_tracks(
     video_path: str,
     previous_tracks: str,

requirements.txt

@@ -1,6 +1,5 @@
 transformers==4.57.6
 accelerate
-molmo_utils
 torchvision
 av
 decord