Decouple Filter MP Rules function from cuda imports

[fabianlim]

Unwittingly due to the recent PR #106 we had introduced a filter_mp_rules function and placed it into models.utils. Unfortunately, that file does cuda kernel imports, and will cause pip install failures in machines with no GPU. This is is critical because it precludes building images in CI machines that typically do not have GPUs.

Failure

#33 15.79 File "/usr/lib64/python3.11/ctypes/__init__.py", line 394, in __getitem__ #33 15.79 func = self._FuncPtr((name_or_ordinal, self)) #33 15.79 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ #33 15.79 AttributeError: /home/tuning/.local/lib/python3.11/site-packages/bitsandbytes/libbitsandbytes_cpu.so: undefined symbol: cdequantize_blockwise_fp32 ..#33 ERROR: process "/bin/sh -c if [[ \"${ENABLE_FMS_ACCELERATION}\" == \"true\" ]]; then python -m pip install --user \"$(head bdist_name)[fms-accel]\"; python -m fms_acceleration.cli install fms_acceleration_peft; python -m fms_acceleration.cli install fms_acceleration_foak; python -m fms_acceleration.cli install fms_acceleration_aadp; fi" did not complete successfully: exit code: 1 ------ > importing cache manifest from docker-na-private.artifactory.swg-devops.com/wcp-ai-foundation-team-docker-virtual/sft-trainer-aim:release_ubi9_py311: ------ ------ > importing cache manifest from sft-trainer-aim:release_ubi9_py311: ------ ------

To fix this we move filter_mp_rules out in a different function. We seem to observe that this will solve the problem

[Abhishek-TAMU]

All the changes of moving filter_mp_rules out of .models.utils looks good. Also this change helps fix the issue and passes the image build in GitHub actions here.