Can HiFiC compression be used on multiple GPUs?

[gregg-p]

Is the HiFiC compression/decompression (tfci.py) able to utilize multiple GPUs when running? I have 8 GPUs, but noticed it is currently only using 1. Using all 8 would provide a significant speed increase. Any help is appreciated. Thanks.

[gtoderici]

Hi there, If you want to use more GPUs you need to write code to achieve that. There's no explicit device placement so you should be able to do that, but it's unclear what you'd achieve with more GPUs. George

…

On Thu, Nov 17, 2022 at 12:53 PM Malocch1o ***@***.***> wrote: Is the HiFiC compression/decompression (tfci.py) able to utilize multiple GPUs when running? I have 8 GPUs, but noticed it is currently only using 1. Using all 8 would provide a significant speed increase. Any help is appreciated. Thanks. — Reply to this email directly, view it on GitHub <#162>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACGUEMZ7JLMFMHXBQXGEXXLWI2LNPANCNFSM6AAAAAASDZOWLY> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[gregg-p]

os.environ["CUDA_VISIBLE_DEVICES"] = "" allows you to specify which CUDA device to run the script on.
At the top of tfci.py, I add: os.environ["CUDA_VISIBLE_DEVICES"] = "0, 1, 2, 3, 4, 5, 6, 7", it recognizes all 8 GPUs, but it still only uses GPU 0. Would modifications need to be made to how the actual model/code functions to utilize more than 1 GPU?
I've noticed this is also the same scenario when training the untrained models.
I appreciate your help & guidance thus far.

[gtoderici]

I think you'll need to write code to do the distribution between GPUs. In particular you'd probably want to load a batch on each GPU explicitly (i.e., use with tf.device("/gpu:X") for X in 0->7 and only then execute the graph). George

…

On Tue, Dec 6, 2022 at 11:47 AM Malocch1o ***@***.***> wrote: *os.environ["CUDA_VISIBLE_DEVICES"] = ""* allows you to specify which CUDA device to run the script on. When I add: *os.environ["CUDA_VISIBLE_DEVICES"] = "0, 1, 2, 3, 4, 5, 6, 7"*, it recognizes all 8 GPUs, but it still only uses GPU 0. Would modifications need to be made to how the actual model/code functions to utilize more than 1 GPU? — Reply to this email directly, view it on GitHub <#162 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACGUEMZISRRQJAPCPRJIJOLWL6J4BANCNFSM6AAAAAASDZOWLY> . You are receiving this because you commented.Message ID: ***@***.***>

[gpatti-sabre]

Would using TensorFlow's tf.distribute.MirroredStrategy() be a feasible option for solving this issue?