Benchmark oxford_spires keble-college-02 with 8GB GPU

[ureeey]

I resolved the discrepancies between lingbot-map.py and predict_long.py (ureeey@6226ed3), then ran the benchmark with the following results.Is this result within the normal range?

lingbot-map.yaml

model: lingbot_map
env: lingbot-map
_checkpoint: /home/jcy/workspace/models/lingbot-map-long.pt
_device: cuda
_mode: windowed
_use_amp: true
_use_sdpa: false  # Force FlashInfer backend (paged KV cache attention)
_image_size: 518
_patch_size: 14
_enable_3d_rope: true
_num_scale_frames: 2
_max_frame_num: 1024
_kv_cache_sliding_window: 48
_kv_cache_scale_frames: 2
_keyframe_interval: auto    # auto = 1 if frames<=320 else ceil(N/320); or a fixed int
_auto_keyframe_threshold: 320    # frame-count threshold driving the 'auto' keyframe_interval (streaming mode)
_area_budget: 255000
_align: 14

Results

{
  "lingbot_map": {
    "ate": 2.0922960340163272,
    "num_scenes": 1,
    "rpe_rot": 0.8845998878433984,
    "rpe_trans": 0.1718545140124919
  }
}

{
  "lingbot_map": {
    "AUC_03": 11.648380355276906,
    "AUC_05": 21.477664576802507,
    "AUC_15": 59.774817136886114,
    "AUC_30": 78.31432863113898,
    "Racc_03": 65.92868338557993,
    "Racc_05": 84.53173981191222,
    "Racc_15": 100.0,
    "Racc_30": 100.0,
    "Tacc_03": 22.71551724137931,
    "Tacc_05": 43.256269592476485,
    "Tacc_15": 92.4960815047022,
    "Tacc_30": 98.52664576802508,
    "num_scenes": 1
  }
}