<html> <head> <meta http-equiv="Content-Type" content="text/html; charset=Windows-1252"> <meta name="Generator" content="Microsoft Exchange Server"> <style></style> </head> <body> <div> <div dir="auto">This is when I test it on the accel nodes unfortunately, though the error vanishes in 1.14.0.</div> <div class="x_gmail_extra"><br> <div class="x_gmail_quote">On 25 Oct 2019 09:13, "Henrik R. Nagel via RT" <support@metacenter.no> wrote:<br type="attribution"> </div> </div> </div> <font size="2"><span style="font-size:10pt;"> <div class="PlainText">Hi,<br> <br> > Unfortunately, when I downgrade (again, using conda), I can no longer<br> > import tensorflow, because it can’t find libcuda.so.1.<br> <br> The libcuda.so.1 is located in /usr/lib64, but only on the accel compute nodes and not on the login nodes. If you tested the downgraded version of TensorFlow on a login node, then that is the reason why it didn't work. You must run the downgraded version of TensorFlow as a batch job on the "accel" partition or request an interactive compute node there.<br> <br> Best regards,<br> <br> Henrik<br> </div> </span></font> </body> </html>