Buckets:
| # Visual-OPSD (offline) — Trace Collection. | |
| # Cache teacher top-256 logprobs + VT-sensitivity weights for all | |
| # training data so the offline trainer can do cheap KD without | |
| # re-running the teacher every step. | |
| # Expected: ~8 hours on 4× H800. | |
| SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)" | |
| ROOT_DIR="$(cd "${SCRIPT_DIR}/../.." && pwd)" | |
| cd "${ROOT_DIR}" | |
| export PYTHONPATH="${ROOT_DIR}:${PYTHONPATH:-}" | |
| source .venv/bin/activate | |
| MODEL_PATH=${1:-models/ThinkMorph-7B} | |
| OUTPUT_DIR=${2:-traces/visual_opsd_teacher_cache} | |
| NGPU=${3:-4} | |
| MASTER_ADDR=${ARNOLD_WORKER_0_HOST} | |
| MASTER_PORT=(${ARNOLD_WORKER_0_PORT//,/ }) | |
| NPROC_PER_NODE=${ARNOLD_WORKER_GPU} | |
| NNODES=${ARNOLD_WORKER_NUM} | |
| NODE_RANK=${ARNOLD_ID} | |
| echo "=== Visual-OPSD Trace Collection ===" | |
| echo "Model: ${MODEL_PATH}" | |
| echo "Output: ${OUTPUT_DIR}" | |
| echo "GPUs: ${NGPU}" | |
| torchrun \ | |
| --nnodes=${NNODES} \ | |
| --node_rank=${NODE_RANK} \ | |
| --nproc_per_node=${NPROC_PER_NODE} \ | |
| --master_addr=${MASTER_ADDR} \ | |
| --master_port=${MASTER_PORT} \ | |
| scripts/visual_opsd/collect_traces.py \ | |
| --model_path ${MODEL_PATH} \ | |
| --data_root datasets \ | |
| --output_dir ${OUTPUT_DIR} \ | |
| --top_k 256 \ | |
| --batch_save_size 200 | |
Xet Storage Details
- Size:
- 1.19 kB
- Xet hash:
- 5aea0b7587cf3c84ee68181dfc184d4f5461c8069d60b1a5a916ef1dc1de2920
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.