Tensorflow mock train

[scaliby]

Description

Create script that resembles tensorflow training but is a no-op.
Use it in docs for tpu7x as maxtext is unrealiable and does not work often.

Preview: https://github.com/scaliby/xpk/blob/tensorflow-mock-train/README.md

Issue

n/a

Testing

Tested commands manually.

[jamOne-]

What's the point of this script actually? Why don't we just use something that actually imports tensorflow and uses it, like: helloworld or simple linear regression?

[scaliby]

What's the point of this script actually? Why don't we just use something that actually imports tensorflow and uses it, like: helloworld or simple linear regression?

Tensorflow doesn't ship by default in python:3 image, that is used by XPK by default so it would have to be installed.
We could explicitly specify the docker image with the tensorflow in this recipe and then write script that actually leverages tensorflow, but will it support the latest and greatest TPU? The official TF version might lack behind, which is actually what is happening with JAX and maxtext (and is the main reason it is unreliable).
This is a zero dependencies script so it will always work. Let's discuss this today during our meeting.

[jamOne-]

What's the point of this script actually? Why don't we just use something that actually imports tensorflow and uses it, like: helloworld or simple linear regression?

Tensorflow doesn't ship by default in python:3 image, that is used by XPK by default so it would have to be installed. We could explicitly specify the docker image with the tensorflow in this recipe and then write script that actually leverages tensorflow, but will it support the latest and greatest TPU? The official TF version might lack behind, which is actually what is happening with JAX and maxtext (and is the main reason it is unreliable). This is a zero dependencies script so it will always work. Let's discuss this today during our meeting.

Got it, thanks!

1. How the customer is going to use the latest TPU if none of the frameworks support it?
2. Why don't we just print "hello world" from this python file then? Due to its current "complexity", user might feel that they are actually testing tensorflow, because the output will look like a real training. It might give them an incorrect feeling of safety, that TF will be correctly working now. (Bonus: I'd suggest changing the name from tensorflow.py to fake_training.py?)

[scaliby]

What's the point of this script actually? Why don't we just use something that actually imports tensorflow and uses it, like: helloworld or simple linear regression?

Tensorflow doesn't ship by default in python:3 image, that is used by XPK by default so it would have to be installed. We could explicitly specify the docker image with the tensorflow in this recipe and then write script that actually leverages tensorflow, but will it support the latest and greatest TPU? The official TF version might lack behind, which is actually what is happening with JAX and maxtext (and is the main reason it is unreliable). This is a zero dependencies script so it will always work. Let's discuss this today during our meeting.

Got it, thanks!

1. How the customer is going to use the latest TPU if none of the frameworks support it?
2. Why don't we just print "hello world" from this python file then? Due to its current "complexity", user might feel that they are actually testing tensorflow, because the output will look like a real training. It might give them an incorrect feeling of safety, that TF will be correctly working now. (Bonus: I'd suggest changing the name from tensorflow.py to fake_training.py?)

1. i think there will be a support, we just need something reliable.
2. updated the script name.

@jamOne- PTAL now