YYbear~

YES, we support multi-thread in our JIT code，Convolution，LSTM, and so on, so you have to set OMP_NUM_THREADS. We also use MKL_GEMM , Tanh to build our FC and RNN op....

We have use MKLDNN in our latest code, but MKLDNN OP will NOT be used for running model. We will use our JIT code. MKLDNN OP will be used in...

LRN and other OP have been added in https://github.com/PaddlePaddle/Anakin/pull/515, wait for merge

可以先参考DTR教程试一试 https://megengine.org.cn/doc/stable/zh/user-guide/dtr.html

可以贴一下复现的参考链接...如models的下载地址，readme的链接

you can rewrite bn module from pytorch with megengine functional, simply replace `import torch.nn.functional as F` with `import megengine.functional as F`

显卡型号是啥?

用nvidia-smi看起来更直接, 查了一下2204应该是3090的卡https://devicehunt.com/view/type/pci/vendor/10DE/device/2204 卡住的原因可能是用了cuda10的megengine , 导致触发了ptx的jit, 这会非常慢且无法 如果megengine要支持30系列卡需要使用cuda11.x的版本. 可以尝试的方案 1、用cu11系列的megengine https://github.com/Qsingle/MegEngine_CU11 2、使用cpu完成计算