Add granular step-by-step logging in _taro_gpu_infer to find exact GPU abort point

app.py

@@ -889,17 +889,23 @@ def _taro_gpu_infer(video_file, seed_val, cfg_scale, num_steps, mode,
     total_dur_s = ctx["total_dur_s"]
     print(f"[_taro_gpu_infer] tmp_dir={tmp_dir!r} silent_video={silent_video!r} segments={segments} total_dur_s={total_dur_s}")
 
+    print(f"[_taro_gpu_infer] calling _load_taro_feature_extractors")
     extract_cavp, onset_model = _load_taro_feature_extractors(device)
+    print(f"[_taro_gpu_infer] extractors loaded, calling extract_cavp")
     cavp_feats  = extract_cavp(silent_video, tmp_path=tmp_dir)
+    print(f"[_taro_gpu_infer] cavp done, calling extract_onset")
     # Onset features depend only on the video — extract once for all samples
     onset_feats = extract_onset(silent_video, onset_model, tmp_path=tmp_dir, device=device)
+    print(f"[_taro_gpu_infer] onset done, freeing extractors")
 
     # Free feature extractors before loading the heavier inference models
     del extract_cavp, onset_model
     if torch.cuda.is_available():
         torch.cuda.empty_cache()
 
+    print(f"[_taro_gpu_infer] calling _load_taro_models")
     model, vae, vocoder, latents_scale = _load_taro_models(device, weight_dtype)
+    print(f"[_taro_gpu_infer] models loaded")
 
     results = []   # list of (wavs, onset_feats) per sample
     for sample_idx in range(num_samples):