Enabling xla on tensorflow-rocm with Rocm 5.2.3?

[ffleader1]

I have a gfx1030 card l, Rocm 5.2.3 and currently running into xla issues with tensorflow.
My GPU can do pytorch model fine, can do tensorflow fine, without xla. With xla however, I got this error: bitcode module not found at ./opencl.bc

Am I missing something during the install process (because I installed with rocm usecase).

How to get my device running tensorflow training with xla?
Thank you

[aoolmay]

Can confirm, error is reproducible on 6800XT and 6950XT.
Fresh installation: ROCm 5.2.3 + tensorflow-rocm 2.9.2 + Linux 5.11.0-27 (as per installation instructions)
Last working setup, ROCm 5.1.3 + tensorflow 2.9.1

Loss of XLA capability results in ~7% performance loss on most tensorflow workloads.

[Mushoz]

Same issue here. Any idea on when we can expect a fix?

[Mushoz]

I would have expected this issue to get way more attention if it affected everyone. Maybe it doesn't? A question for you @aoolmay and @ffleader1 : What distro are you guys running? I am using Archlinux myself, which is technically unsupported. I am curious if this also affects people that are running supported distributions.

Lastly, my models currently refuse to run at all. How do I explicitly disable XLA so that it at least works again?

[macmv]

I've had the same issue (also running arch with a 6800xt), and setting ROCM_PATH to /opt/rocm fixed it for me.

[Epliz]

hit the same issue with a MI100 GPU, setting ROCM_PATH also helped

[nartmada]

Hi @ffleader1, please check latest ROCm Documentation and ROCm 6.0.0 to see if your query has been resolved. If resolved, please close the ticket. Thanks.

[ffleader1]

Hi @ffleader1, please check latest ROCm Documentation and ROCm 6.0.0 to see if your query has been resolved. If resolved, please close the ticket. Thanks.

I am no longer have access to my original PC
But guessing the complaints were not that much anymore, I guess it's fixed.