Can ROCm be supported?

[iszotic]

System information

• TensorFlow version (you are using): tensorflow-rocm 2.0.0b3, 15/07/19
• TensorFlow Addons version: 0.4.0
• Is it in the tf.contrib (if so, where): Not applicable
• Are you willing to contribute it (yes/no): testing, I guess, but No
• Are you willing to maintain it going forward? (yes/no): No, same as above

Describe the feature and the current behavior/state. When I try to import the module it throws this in console, using the cpu version of tensorflow 2.0.0b1 it imports succesfully `

import tensorflow import tensorflow_addons Traceback (most recent call last): File "", line 1, in File "/home/voltam/.local/lib/python2.7/site-packages/tensorflow_addons/init.py", line 75, in from tensorflow_addons import image File "/home/voltam/.local/lib/python2.7/site-packages/tensorflow_addons/image/init.py", line 22, in from tensorflow_addons.image.distance_transform import euclidean_dist_transform File "/home/voltam/.local/lib/python2.7/site-packages/tensorflow_addons/image/distance_transform.py", line 24, in get_path_to_datafile("custom_ops/image/_image_ops.so")) File "/home/voltam/.local/lib/python2.7/site-packages/tensorflow/python/framework/load_library.py", line 61, in load_op_library lib_handle = py_tf.TF_LoadLibrary(library_filename) tensorflow.python.framework.errors_impl.NotFoundError: /home/voltam/.local/lib/python2.7/site-packages/tensorflow_addons/custom_ops/image/_image_ops.so: undefined symbol: _ZN10tensorflow8internal21CheckOpMessageBuilder9NewStringEv

` Will this change the current api? How? No Who will benefit with this feature? AMD ROCm supported GPUs

[facaiy]

Hi, would you mind giving v0.5.0 (the latest release) a try? Thanks for reporting!

[iszotic]

I tried tfa-nightly and the same keeps happening, this is the version pip show tfa-nightly Name: tfa-nightly Version: 0.5.0.dev20190811

[seanpmorgan]

So tensorflow-rocm is still packaged as manylinux1 meaning it is not built with the same toolchain as TFA or TF2-RC. https://pypi.org/project/tensorflow-rocm/2.0.0b3/#files https://pypi.org/project/tensorflow/2.0.0rc0/#files

Notice we now publish as manylinux2010 spec compliant. Do you happen to know a point of contact who builds/publishes tensorflow-rocm?

[iszotic]

Here I guess https://github.com/ROCmSoftwarePlatform/tensorflow-upstream, but a specific person, nop, :(

[seanpmorgan]

@whchung Would you happen to know if there are plans to publish tensorflow-rocm as a manylinux2010 package as will be done for TF2?

EDIT ---- Just saw there was an issue filed on the ROCm repo

[whchung]

@seanpmorgan , @sunway513 and @parallelo would provide updates on this.

[parallelo]

We'll take a look into this.

[seanpmorgan]

Great, thanks! Closing this issue... but happy to re-open it if problems persist.

[mbonaker]

Well, the problem does persist as I get undefined symbol: _ZNK10tensorflow15shape_inference16InferenceContext11DebugStringEv if I try using tfa.image.rotate with tfa 0.9.1 and tf-rocm 2.1.1 – same result with tfa-nightly 0.10.0. Was this originally closed because it was thought to be solved or because the responsibility lies with the tensorflow-rocm project? In case of the latter just ignore this comment.

[fsx950223]

@seanpmorgan @bhack I could add rocm support for addons and I have passed custom layer ops tests in my local environment. Could tensorflow team provides a CI/CD environment?

[bhack]

@fsx950223 My opinion is that we could try to upstream the few custom ops we have in TF as we are not trying to accumulate other custom ops if we don't have a quite clear roadmap on the topic we have opened at: https://discuss.tensorflow.org/t/deformable-convolution-and-other-custom-ops/1951

If these PRs will be accepted in TF we have already the ROCM CI available there.

[DanielWicz]

Is ROCm dead for TFA ? Or I can install and then tensorflow-rocm without a problem (except using some cuda-kernelized functions) ?

[seanpmorgan]

TensorFlow Addons is transitioning to a minimal maintenance and release mode. New features will not be added to this repository. For more information, please see our public messaging on this decision: TensorFlow Addons Wind Down

Please consider sending feature requests / contributions to other repositories in the TF community with a similar charters to TFA: Keras Keras-CV Keras-NLP