commit 6eabbea47ced39111766a043a92a22a11cd45975
author Eddie Yan <eddiey@nvidia.com> Thu Jun 10 22:06:11 2021 -0700
committer Facebook GitHub Bot <facebook-github-bot@users.noreply.github.com> Thu Jun 10 22:07:28 2021 -0700
tree 3b9f7bd570deac58e02c44e41fd30cffb3c9d5b2
parent 455afdf974fc2546cb79f4cc27d307d4b24437ba

Disable cuDNN persistent RNN on A30 (#59830)

Summary:
https://github.com/pytorch/pytorch/issues/59829

cherry-picked from ptrblck 's change CC ngimel xwang233

Pull Request resolved: https://github.com/pytorch/pytorch/pull/59830

Reviewed By: bdhirsh

Differential Revision: D29046145

Pulled By: ngimel

fbshipit-source-id: 270ab3bb6c1c7c759497a15eb38b20a177c94adb

aten/src/ATen/native/cudnn/RNN.cpp

diff --git a/aten/src/ATen/native/cudnn/RNN.cpp b/aten/src/ATen/native/cudnn/RNN.cpp
index da83099..f81de80 100644
--- a/aten/src/ATen/native/cudnn/RNN.cpp
+++ b/aten/src/ATen/native/cudnn/RNN.cpp
@@ -726,7 +726,8 @@
                 (tensors.seq_length >=20 && bsize <=96) ||
                 (tensors.seq_length >=10 && bsize <=32));
       }
-    } else if (prop->major >= 8) {
+    } else if (prop->major >= 8 && prop->multiProcessorCount >= 98) {
+      // SM count check excludes A30 (similar issue to A40)
       if (prop->minor == 6) {
         // Excludes sm_86 GPU devices from using persistent rnn.
         // This is because there are some edge cases that will throw exceptions with cudnn 8.0.5 on Nvidia A40 GPU.