Add upload_to_shared: push results to HuggingFriends generated_data/<model>_<variant>

Fix misleading H100 text in log (actual GPU is GPU_TYPE/L40S)

Add clean_frames: delete video/ frame dirs from volume (keep mp4+prompt)

Default extract_frames=False (judge reads mp4); HF gets mp4+prompt only

Upload: exclude video/ frames from HF dataset (too many small files)

Add --upload: push each source results to HF dataset from cloud

Defaults for production: GPU=L40S, keep_prob=0.4, model_key=genie; per-sample timing

Default keep_prob=0.4

Fix: chdir to repo so GE relative class paths resolve (in-process inference)

Align output layout to partner (task/episode/1 + prompt/ + video frames); --view all|head

Batch pipeline: load-once + parallel shards + resumable + deterministic hash sampling (matches partner run_local.py)

Fix: uninstall deepspeed in image (not needed for inference, breaks import without CUDA_HOME)

Add --n_chunk CLI arg to control video length

Remove README_RUN.md

Add GE-Base Modal inference code (ge_modal_app.py + README_RUN.md)

initial commit