Printing debug information in __repr__ only when a debugger has set a trace

[brownbaerchen]

Due Diligence

• General:
  • title of the PR is suitable to appear in the Release Notes
• Implementation:
  • unit tests: all split configurations tested
  • unit tests: multiple dtypes tested
  • NEW unit tests: MPS tested (1 MPI process, 1 GPU)
  • benchmarks: created for new functionality
  • benchmarks: performance improved or maintained
  • documentation updated where needed

Description

Python uses __repr__ and __str__ to give textual information about objects. The intention is that __repr__ gives output for developers and __str__ gives output for users. In heat arrays, this means __repr__ tells you, for each MPI rank, what kind of data is contained in the array, while __str__ tells you the actual data.

This is all well and good but the problem is when people are using heat interactively, in which case they get the __repr__ if they just print the array, but if they are users, they want to know the data and not the local shape.
For example:

python
import heat as ht
a = ht.arange(4)
a  # prints a.__repr__()
f"{a}"  # prints a.__str__()

when doing something like this, you tend to want to know the contents of the array in both cases, i.e. the result of __str__.

This PR is not a very nice solution and I would be fine if we reject this, but I don't know a better way.
What this PR proposes is to return the normal __str__ output in __repr__, unless the interactive session is a debugging session. This is determined via sys.gettrace().

Issue/s resolved: #2018

Changes proposed:

• Print debug information in __repr__ only if the interactive session is a debugging session

[github-actions]

Thank you for the PR!

[brownbaerchen]

I am kind of lost why this fails. On Jureca I get

(heat_venv) [me@jrlogin02 heat]$ python
Python 3.12.3 (main, Jul 26 2024, 17:40:49) [GCC 13.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import heat as ht
/path/heat/heat/core/_config.py:108: UserWarning: Heat has CUDA GPU-support (PyTorch version 2.5.1+cu124 and `torch.cuda.is_available() = True`), but CUDA-awareness of MPI could not be detected. This may lead to performance degradation as direct MPI-communication between GPUs is not possible.
  warnings.warn(
>>> a = ht.arange(4, device='gpu')
>>> a
DNDarray([0, 1, 2, 3], dtype=ht.int32, device=gpu:0, split=None)
>>> a.__repr__(force_debug_output=True)
'<DNDarray(MPI-rank: 0, Shape: (4,), Split: None, Local Shape: (4,), Device: gpu:0, Dtype: int32)>'
>>> a.__str__() == a.__repr__()
True

Same result with cpu array and the tests pass both in serial as well as in parallel.

So it seems to work also on GPUs. Anyone any ideas? I will not spend time on this until we decide that we even want this behavior.