Qwen-Drive Resume Runs May Skip New Predictions
A static review of Qwen-Drive’s official batch runner found that it reads scene tokens from existing output files and skips scenes already marked complete. The token check does not verify whether the previous run used the same weights, planning mode, seed or candidate count, so changing configuration and reusing an output directory can skip new trajectory predictions or leave results from different configurations mixed together. The review recommends separate output directories for new configurations, configuration hashes and run identifiers, records of requested and newly predicted scene counts, and explicit timing boundaries. Model loading may still use the GPU even when all scenes are skipped. The review did not download model weights, run GPU work, train the model or conduct a driving evaluation, leaving latency, memory use, success rates and trajectory quality unknown.
Reusing an output directory after changing Qwen-Drive’s weights, planning mode or other settings can produce incomplete or mixed results. Counts of new predictions and configuration-linked output directories are therefore necessary for meaningful timing and output comparisons.
单卡 A6000 跑 Qwen -Drive:怎样留下可比较的第一条基线?
把 direct_planning 改成 reasoning_planning,再运行一次同一条命令,终端很快显示任务完成。能把这次耗时当作“加入推理之后的速度”吗?
Qwen-Drive 官方批处理脚本给出了一个容易漏掉的答案:它默认读取输出文件中的场景 token,跳过已经完成的场景。脚本先加载模型,再筛选尚未完成的 token;如果全部存在,就输出 nothing to do 并返回。改变模式不会自动使旧记录失效。你可能真的加载了权重,却根本没有执行新的轨迹预测。1
这也是单卡研究的第一个任务:留下能够证明“这一次究竟执行了什么”的基线。本文以 RTX A6000 为目标设备,核对截至 2026-10-18 的官方源码与权重元数据,给出待执行的研究方法。本文没有下载模型权重,没有运行 GPU、训练或驾驶评测;文中的硬件能力不是本机兼容性结论,全部延迟、峰值显存和成功率保持未知。