=== Profiling parquet ===
File: /lustre/fsw/portfolios/dir/projects/dir_cosmos_base_lustre/cosmos3_action_datasets/bridge_raw/data/chunk-000/file-000.parquet

Rows: 1,496,285  Row groups: 2  Columns: 7
Chunk size (15 fps, 76 frames): 76 rows per sample
File size on disk: 84.22 MiB (59 bytes/row)

Columns hand_pose uses (present in parquet):
  task_index
  episode_index
  frame_index
  timestamp

Columns in parquet NOT in hand_pose list:
  observation.state
  action
  index

Arrow table size (in-memory): 146.09 MiB (102 bytes/row)
  observation.state: 48037.5 KiB  (list<element: float>)
  action: 48037.5 KiB  (list<element: float>)
  timestamp: 6027.5 KiB  (float)
  frame_index: 11872.4 KiB  (int64)
  episode_index: 11872.4 KiB  (int64)
  index: 11872.4 KiB  (int64)
  task_index: 11872.4 KiB  (int64)

Full-table read time (1,496,285 rows):
  All columns (7):    0.0874s  (mean of 5)
  Subset (4 cols): 0.0080s  (mean of 5)
  Subset is 0.09x of full
  => Subset reads less data; loader column selection could help.

Chunk read (76 rows) at random offsets (in-memory slice, mean of 5): 0.0000s

Note: LeRobot loads the full HF dataset (all columns). Column selection
only helps when the omitted columns are a large share of the table.

=== Profiling video ===
Root: /lustre/fsw/portfolios/dir/projects/dir_cosmos_base_lustre/cosmos3_action_datasets/bridge_raw
Video key: observation.images.main
Profiling 5 video(s), 76 frames each, warmup=1 repeat=10 access=random

/usr/local/lib/python3.12/dist-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
/usr/local/lib/python3.12/dist-packages/torch_automated_profiler/utils/misc.py:13: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
  from pkg_resources import parse_version
  videos/observation.images.main/chunk-000/file-000.mp4
    open+create=0.0018s  decode=0.2420s  total=0.2438s  frames=76
  videos/observation.images.main/chunk-000/file-001.mp4
    open+create=0.0024s  decode=0.2358s  total=0.2382s  frames=76
  videos/observation.images.main/chunk-000/file-002.mp4
    open+create=0.0025s  decode=0.2687s  total=0.2712s  frames=76
  videos/observation.images.main/chunk-000/file-003.mp4
    open+create=0.0023s  decode=0.2687s  total=0.2710s  frames=76
  videos/observation.images.main/chunk-000/file-004.mp4
    open+create=0.0020s  decode=0.2589s  total=0.2609s  frames=76

Summary (mean over videos):
  open+create decoder: 0.0022s
  decode (get_frames_at): 0.2548s
  total per video load: 0.2570s
DONE_PROFILE
RESULT: bridge_raw_profile PASSED