[QUESTION]flux/src/moe_ag_scatter/ths_op/gemm_grouped_v3_ag_scatter.cc里的写stream很耗时，有优化方法吗？CU_CHECK(CUStreamWriteValue(this->cp_stream_intra_node,(CUdeviceptr)(ptr_offset(barrier_block.get(), src_rank * sizeof(int))),1,CU_STREAM_WRITE_VALUE_DEFAULT));

[jinchen89]

@wenlei-bao

这里触发了大页内存到gpu内存的拷贝，gpu和cpu的同步阻碍了其他kernel的预加载，能解决吗？ 有没有替代的方法，比如写个flag?或者用回调函数避免等待？

[houqi]

It can be solved, but not a very easy one. have to

• put the shape fully into device side
• run no CUDA runtime API and use kernels for us
• make sure memory is pre-allocated

For large-batch size, the sync is not so bad. but we will record this issue.