Your browser does not seem to support JavaScript. As a result, your viewing experience will be diminished, and you have been placed in read-only mode.
Please download a browser that supports JavaScript, or enable it if it's disabled (i.e. NoScript).
@exllm 并行改为1后，预填充性能有接近5~10倍的提升（不同的任务变化很大），但生成token的性能提升不到10%。我也将“--override-tensor blk.\d+.ffn_.*_exps.=CPU \”注释掉进行过测试，性能反而有略微下降，可能跟我的硬件和跑的任务也有一定关系。
