Upload app.py with huggingface_hub

app.py

@@ -464,6 +464,32 @@ def analyze(
                 desc=f"Value pass {len(pred_value)}/{len(eval_indices)}",
             )
 
+    # --- DEBUG: single-pass per-frame readout (model card "Programmatic Usage") ---
+    dbg_idx = np.linspace(0, total - 1, num_frames, dtype=int)
+    dbg_sample = processor.process_episode(
+        instruction=instruction,
+        images=[model_images[j] for j in dbg_idx],
+        robot_description=robot_description,
+        camera_description=camera_description,
+    )
+    with torch.inference_mode():
+        dbg_out = model(
+            input_ids=dbg_sample["input_ids"].to("cuda").long(),
+            attention_mask=dbg_sample["attention_mask"].to("cuda").long(),
+            pixel_values=dbg_sample["pixel_values"].flatten(0, 1).to("cuda"),
+            image_grid_thw=dbg_sample["image_grid_thw"].flatten(0, 1).to("cuda").long(),
+        )
+    dbg_pred = dbg_out.value.pred_value.float().reshape(-1).tolist()
+    dbg_rel = dbg_out.relative.pred_value.float().reshape(-1).tolist()
+    debug_lines = [
+        "```",
+        "SINGLE-PASS values: " + ", ".join(f"{v:.2f}" for v in dbg_pred),
+        "SINGLE-PASS deltas: " + ", ".join(f"{v:.2f}" for v in dbg_rel),
+        "PREFIX last-slot  : " + ", ".join(f"{v:.2f}" for v in pred_value),
+        f"fps={fps:.2f} total={total}",
+        "```",
+    ]
+
     progress(0.82, desc="Generating analysis...")
     input_ids = final_sample["input_ids"].to("cuda").long()
     with torch.inference_mode():
@@ -506,6 +532,8 @@ def analyze(
             f"{pred_value[-1]:.2f} s at the last frame "
             f"(video is {(total - 1) / fps:.1f} s long, {total} frames decoded).",
             "",
+            "\n".join(debug_lines),
+            "",
             f"<sub>{len(eval_indices)} prefix evaluations · {num_frames} frames per prefix · "
             f"{gpu_seconds:.1f} s on GPU</sub>",
         ]