2026-02-19 15:41:48,629 - INFO - === Loading & Creating Swap Pairs === 2026-02-19 15:41:51,137 - INFO - Swap pair creation stats: 2026-02-19 15:41:51,137 - INFO - left: 616/616 2026-02-19 15:41:51,137 - INFO - right: 620/620 2026-02-19 15:41:51,137 - INFO - above: 596/596 2026-02-19 15:41:51,137 - INFO - under: 602/602 2026-02-19 15:41:51,137 - INFO - far: 594/594 2026-02-19 15:41:51,137 - INFO - close: 612/612 2026-02-19 15:41:51,137 - INFO - Total pairs: 3640 2026-02-19 15:41:51,498 - INFO - PyTorch version 2.3.0 available. 2026-02-19 15:41:51,917 - INFO - Loading HF dataset: FlagEval/EmbSpatial-Bench 2026-02-19 15:42:10,591 - INFO - Built bbox cache: 3640 entries (sample keys: ['mp3d_0', 'mp3d_1', 'mp3d_2', 'mp3d_3', 'mp3d_4']) 2026-02-19 15:42:10,649 - INFO - Matched 1206/1206 question_ids between TSV and HF dataset 2026-02-19 15:42:10,659 - INFO - Cross-group quads: 1039/1206 (ambiguous=112, no_bbox=55) 2026-02-19 15:42:10,659 - INFO - ============================================================ 2026-02-19 15:42:10,659 - INFO - Processing nvila - vanilla 2026-02-19 15:42:10,659 - INFO - Model path: /data/shared/Qwen/mydisk/NVILA-Lite-2B 2026-02-19 15:42:10,659 - INFO - ============================================================ [2026-02-19 15:42:12,889] [INFO] [real_accelerator.py:110:get_accelerator] Setting ds_accelerator to cuda (auto detect) 2026-02-19 15:42:19,682 - INFO - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk). You are attempting to use Flash Attention 2.0 with a model not initialized on GPU. Make sure to move the model to GPU after initializing it on CPU with `model.to('cuda')`. The new embeddings will be initialized from a multivariate normal distribution that has old embeddings' mean and covariance. As described in this article: https://nlp.stanford.edu/~johnhew/vocab-expansion.html. To disable this, use `mean_resizing=False` 2026-02-19 15:43:49,343 - INFO - Loaded NVILA from /data/shared/Qwen/mydisk/NVILA-Lite-2B 2026-02-19 15:43:49,344 - INFO - Model has 28 layers. Extracting ALL. 2026-02-19 15:43:49,345 - INFO - --- Phase A: Extracting swap pair features --- Swap pairs: 0%| | 0/1200 [00:00