Add cudnn 5.0.5 to supported versions#321
Add cudnn 5.0.5 to supported versions#321soumith merged 1 commit intopytorch:masterfrom szagoruyko:cudnn-5-0-5
Conversation
|
fused winograd (a.k.a. just winograd) is in cudnn v5, v5.1 added nonfused winograd. Are you seeing convergence problems with either of them? Also, are there reasons to specify patch versions for 5.1 or 5.0, won't major/minor be enough? |
|
is there a way to not use Winograd via benchmark or fastest? cant do just 5.1 or 5.0 because the libcudnn.so.x doesn't come only in major versions. there's a .so.5 and a .so.5.0.5 |
|
Ah, you are right with respect to numbering, my bad. Nonfused winograd is not returned in fastest (cudnnGet*), and benchmark (cudnnFind*) returns a table of results, so they can be filtered in the framework if someone does not want to use Winograd or non-fused winograd. No way to prevent cudnnGet from returning (fused) winograd, AFAIK. |
|
thanks! |
Update upload_binary_sizes.sh to use CUDA 9.2
* Initial implementation of UnspecializedPrimitiveVariable * Update heuristic * Add test for no recompilations for different values
This reverts commit 97b55d1.
I ran tests locally, only rnn do not pass because they were added in 5.1, the rest pass. Having 5.0 is very useful for: