Add ROMA frozen backbone extraction support

run_extract_easycom_direct.sh

@@ -8,6 +8,7 @@ set -euo pipefail
 
 PROJECT_ROOT="${PROJECT_ROOT:-/mnt/58462F4C462F29E8/Code/icassp2027-experiment}"
 MODEL="${MODEL:-/mnt/58462F4C462F29E8/Models/Qwen2.5-Omni-3B}"
+BACKBONE_LOADER="${BACKBONE_LOADER:-qwen}"
 VIDEO_ROOT="${VIDEO_ROOT:-/mnt/58462F4C462F29E8/Datasets/easycom/Main/Video_Compressed}"
 TARGET_ROOT="${TARGET_ROOT:-/mnt/58462F4C462F29E8/Code/EgoSpeak/dataset/EasyCom/target_perframe}"
 FEATURE_ROOT="${FEATURE_ROOT:-${PROJECT_ROOT}/features/easycom_direct}"
@@ -23,6 +24,9 @@ DTYPE="${DTYPE:-bfloat16}"
 DECODER="${DECODER:-decord}"
 DECODER_GPU_ID="${DECODER_GPU_ID:-0}"
 FRAME_CACHE_SIZE="${FRAME_CACHE_SIZE:-64}"
+# Delay worker startup by GPU index to avoid eight processes reading a large
+# full ROMA checkpoint at the same instant. Zero preserves the old behavior.
+MODEL_LOAD_STAGGER_SECONDS="${MODEL_LOAD_STAGGER_SECONDS:-0}"
 # Space-separated host GPU ids, e.g. "0" or "0 1 2 3". One session per GPU at a time.
 GPUS="${GPUS:-0}"
 # Optional: only run selected session ids, e.g. SESSIONS="1 2 3"
@@ -70,6 +74,7 @@ run_one() {
   {
     echo "$(date): GPU ${gpu} starting ${session}"
     echo "model=${MODEL}"
+    echo "backbone_loader=${BACKBONE_LOADER}"
     echo "session_dir=${session_dir}"
     echo "target=${target}"
     echo "output=${output}"
@@ -81,6 +86,7 @@ run_one() {
     if ! CUDA_VISIBLE_DEVICES="${gpu}" PYTHONUNBUFFERED=1 \
       python streaming_turn_model_directly.py \
         --model-id "${MODEL}" \
+        --backbone-loader "${BACKBONE_LOADER}" \
         --easycom-session-dir "${session_dir}" \
         --target "${target}" \
         --use-audio-in-video \
@@ -124,15 +130,30 @@ done
 
 pids=()
 status=0
+worker_index=0
 for gpu in "${GPU_ARR[@]}"; do
   sids="${WORKERS[${gpu}]}"
   [[ -z "${sids// }" ]] && continue
+  delay_seconds="$(python - "${MODEL_LOAD_STAGGER_SECONDS}" "${worker_index}" <<'PY'
+import sys
+print(float(sys.argv[1]) * int(sys.argv[2]))
+PY
+)"
   (
+    if python - "${delay_seconds}" <<'PY'
+import sys
+raise SystemExit(0 if float(sys.argv[1]) > 0 else 1)
+PY
+    then
+      echo "GPU ${gpu} worker waiting ${delay_seconds}s before model load"
+      sleep "${delay_seconds}"
+    fi
     for sid in ${sids}; do
       run_one "${gpu}" "${sid}" || exit 1
     done
   ) &
   pids+=("$!")
+  worker_index=$((worker_index + 1))
 done
 
 for pid in "${pids[@]}"; do