Match multi-video generation settings in Qwen pilot

single_video_567/code/run_qwen3vl8b_batch.sh

@@ -67,7 +67,7 @@ serve_and_eval() {
         --skip "$EVAL_SKIP" --limit "$EVAL_LIMIT" \
         --run-id "$run_id" --model "$model" --thinking auto \
         --max-tokens "$max_tokens" --max-tokens-cap "$max_tokens_cap" \
-        --workers 1 --rps 0.2 \
+        --workers 6 --rps 0.45 \
         --base-url "http://127.0.0.1:$PORT/v1"
     "$PY" "$ROOT/unified_eval.py" score --dataset "$EVAL_DATASET" --run-id "$run_id"
     kill "$SERVER_PID" 2>/dev/null || true
@@ -85,5 +85,5 @@ serve_and_eval() {
 
 "$PY" -m pip show vllm transformers torch | grep -E 'Name:|Version:'
 serve_and_eval Qwen/Qwen3-VL-8B-Instruct "qwen3vl8b-instruct-$RUN_TAG_SUFFIX" 2048 8192
-serve_and_eval Qwen/Qwen3-VL-8B-Thinking "qwen3vl8b-thinking-$RUN_TAG_SUFFIX" 8192 32768
+serve_and_eval Qwen/Qwen3-VL-8B-Thinking "qwen3vl8b-thinking-$RUN_TAG_SUFFIX" 16384 32768
 "$PY" "$ROOT/unified_eval.py" compare --dataset single_567_human