Download runs/scripts/worker.sh from BlueSourceJY/ERCBench-LongLive-400: direct link, hf CLI and curl.
- Browser
- Download file 1 kB
-
https://huggingface.co/datasets/BlueSourceJY/ERCBench-LongLive-400/resolve/main/runs/scripts/worker.sh
- Command line
-
hf download hf://datasets/BlueSourceJY/ERCBench-LongLive-400/runs/scripts/worker.sh
-
curl -L -o worker.sh https://huggingface.co/datasets/BlueSourceJY/ERCBench-LongLive-400/resolve/main/runs/scripts/worker.sh
1 kB
| # Multi-GPU LongLive worker (used only by ~/gen_common/ll_boost.sh once the ShotStream eval has finished). | |
| # usage: worker.sh <gpu>. Runs 4 -> 6 -> 8 shot configs; episodes are claimed via mkdir in ~/longlive_run/claims. | |
| # Each process seeds torch once and draws the noise for every episode (also skipped/claimed ones), so each episode | |
| # gets exactly the same noise as in the single-GPU run. | |
| g=$1 | |
| cd ~/longlive_run/LongLive/LongLive1.0; source ~/longlive_run/venv/bin/activate | |
| export LL_CLAIM_DIR=~/longlive_run/claims; mkdir -p $LL_CLAIM_DIR | |
| for n in 4 6 8; do | |
| echo "$(date -u +%FT%T.%6NZ) [CONFIG] ${n}shot start gpu=$g" | |
| CUDA_VISIBLE_DEVICES=$g python ~/gen_common/tok_hook_run.py wan.modules.tokenizers ~/longlive_run/logs/tokens_gpu$g.jsonl interactive_episode.py --config_path ~/longlive_run/inputs/longlive_${n}shot.yaml | |
| rc=$?; echo "$(date -u +%FT%T.%6NZ) [CONFIG] ${n}shot EXIT_CODE=$rc gpu=$g"; [ $rc -ne 0 ] && exit $rc | |
| done | |
| echo "$(date -u +%FT%T.%6NZ) [WORKER] gpu=$g DONE" | |