| # Encode worker: launch_encode_v2.sh <gpu> <mem_gb> <name>. Ranks share shards via claims/ (--dynamic). | |
| set -euo pipefail | |
| export OMP_NUM_THREADS=16 MKL_NUM_THREADS=16 NUMBA_NUM_THREADS=8 CUDA_VISIBLE_DEVICES=$1 HF_HOME=/data/mukil/hf HF_HUB_OFFLINE=1 PYTORCH_ALLOC_CONF=expandable_segments:True | |
| cd /data/mukil/datasets/fused/tts | |
| /data/mukil/mininconda3/envs/moss-tts/bin/python moss_tts_local_v1.5/finetuning/encode_single_speaker.py \ | |
| --dataset-dir /data/mukil/datasets/seamless-interaction-s2s-stereo-24k/data \ | |
| --data-dir /data/mukil/datasets/fused \ | |
| --model-path OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5 \ | |
| --codec-path OpenMOSS-Team/MOSS-Audio-Tokenizer-v2 \ | |
| --speaker-model microsoft/wavlm-base-plus-sv \ | |
| --device cuda:0 --gpu-mem-gb "$2" --dynamic 2>&1 | tee -a "/data/mukil/datasets/fused/logs/encode_v2_$3.log" | |