Q: is it possible (any examples?) to train two different models on two different GPUs, in different threads of the same process?

[mw66]

Hi,

Is it possible to train two different models on two different GPUs, in different threads of the same process? (Mainly because the training data itself is the same, but takes lots of memory; does keras support such use case?)

If yes, and thread safe (BTW with Tensorflow backend), can you give minimal clean example to train two different MNIST models in two different threads in keras?

Especially, with the new Python 3.14 no GIL ?

Thanks!

[maitry63]

Hi @mw66 , Tested with Python 3.14 using the TensorFlow backend with Keras 3. At the moment, this configuration doesn’t appear to be supported, it is still in progress. You may want to try Python 3.12 or use the Torch or JAX backend instead, as those currently work more reliably.

[mw66]

@maitry63

Thanks for the info. Can you share your tried working code with Torch or JAX backend? Along with the Python version and freezed requirements.txt of all the library versions?

Thank you!

[maitry63]

Hi @mw66 , I am adding a simple example with Mnist dataset with multigpu parallel training with jax & Torch using multiprocessing. Find the gist here.