I am a Research Student doing profiling on a pytorch code related to NLP. During Profiling I realized that a certain function “embedding_dense_backward” does have slightly more Cpu time than Cuda time. I went into the documentation here
however it wasn’t helpful. Thus could anyone tell what’s this function really do, to my understanding it might be related to updation on an embedding vector during backpropogation. Apologies if question is trivial