Avoid nested `InheritFromOtherOptimizer`

[NouamaneTazi]

We shouldn't pass a InheritFromOtherOptimizer to another InheritFromOtherOptimizer InheritFromOtherOptimizer(optimizer=InheritFromOtherOptimizer(), ...)

But when adding an assert and launch a training we do hit it (so we should fix init_optimizer_and_grad_accumulator)

Shouldn't be a big issue, but we can avoid the additional nesting

[IvanKirchev]

Hi, I would like to take on this issue. Could you please assgin me if help is still needed?

[xrsrke]

assigned :)

[Wodlfvllf]

is any one working on this? If not I can pick it up

[IvanKirchev]

is any one working on this? If not I can pick it up

No progress, feel free to take over

[Wodlfvllf]

@xrsrke
When setting up the development environment on a fresh Docker container by following the CONTRIBUTING.md guide, the test suite fails with a ModuleNotFoundError. Even after creating a self-contained conda environment with Python 3.11 and correctly installing dependencies using pip install -e ".[dev]" and pip install -e ".[test]", essential packages like transformers are not being installed.

The issue is only resolved by manually running pip install transformers after the standard setup commands have been completed. But then running tests is giving "tests module not found"

The Exact Error Message

The error encountered when running pytest is:

ImportError while importing test module '/workspace/nanotron/tests/test_optimizer.py'.
Traceback:
...
src/nanotron/config/config.py:13: in
from transformers import AutoTokenizer
ModuleNotFoundError: No module named 'transformers'