scripts/run_worker.sh

#!/usr/bin/env bash
# One persistent worker for a mixed-method batch.
#
# The CLI --method only has to be a valid starting point: every request carries
# its own method, cache window and method-specific knobs, which the worker
# re-validates per request.  fullkv is used here because it is the one method
# that constrains nothing (no head map, no ring window).
set -euo pipefail
requests=$(realpath "${1:?usage: $0 REQUESTS_JSONL LOG}")
log=$(realpath -m "${2:?usage: $0 REQUESTS_JSONL LOG}")
gpus=${LINGBOT_GPUS:-1,3,4,5}
nproc=${LINGBOT_NPROC:-4}
ulysses=${LINGBOT_ULYSSES:-4}
port=${LINGBOT_PORT:-29771}
root=/home/n2501108a/KV/experiments/rope_reindex_allmethods_20260824
repo=/home/n2501108a/KV/forcingkv_private_worktrees/Lingbot-World-2
python=/home/n2501108a/miniconda3/envs/diff/bin/python
checkpoint=/mnt/storage/n2501108a/context_memory_matched_prune_20260716/models/lingbot-world-v2-14b-causal-fast
mkdir -p "$(dirname "$log")" "$root/worker_placeholder"
cd "$repo"
CUDA_VISIBLE_DEVICES="$gpus" "$python" -m torch.distributed.run \
  --nproc_per_node "$nproc" --master_port "$port" \
  vendor/lingbot_world2/run_context_memory_lingbot_sp.py \
  --checkpoint-dir "$checkpoint" \
  --output-root "$root/worker_placeholder" \
  --profile-name persistent_worker --method fullkv \
  --local-attn-size -1 --sink-size 0 \
  --generation-kv-policy method-native \
  --num-output-latent-frames 16 --num-output-pixel-frames 60 \
  --chunk-size 4 --height 480 --width 832 --seed 1 \
  --sampling-shift 5.0 --ulysses-size "$ulysses" --t5-cpu \
  --camera-condition tail_motion --prompt-adapter strict \
  --anchor-selection none \
  --requests-jsonl "$requests" --skip-existing \
  2>&1 | tee -a "$log"