Add experimental MPS device support for ASR inference

[artbataev]

What does this PR do ?

Add support for MPS device (Apple Silicon M-series GPU) for ASR inference

Collection: [ASR]

Changelog

• allow to select MPS device in transcribe_speech.py

Usage

MPS device will be automatically selected in transcribe_speech.py or speech_to_text_eval.py if allow_mps=true flag is explicitly set, and MPS device is available.
Will work only with PyTorch >= 2.0, for now env variable PYTORCH_ENABLE_MPS_FALLBACK=1 should be set.

Tested with models:

• https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/stt_en_conformer_ctc_large
• https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/stt_en_conformer_transducer_large

PYTORCH_ENABLE_MPS_FALLBACK=1 python examples/asr/transcribe_speech.py \
    allow_mps=true \
    <other script params>

PYTORCH_ENABLE_MPS_FALLBACK=1 python examples/asr/speech_to_text_eval.py \
    allow_mps=true \
    <other script params>

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

• New Feature
• Bugfix
• Documentation

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

• Related to # (issue)

[titu1994]

I'm surprised ASR works on MPS ! Cool finding. Any idea if its just CTC models or if even RNNT will work ?

[artbataev]

I'm surprised ASR works on MPS ! Cool finding. Any idea if its just CTC models or if even RNNT will work ?

In inference mode Conformer-L CTC and RNNT models work fine with PyTorch 2.0 (no quality degradation, significant speedup on MPS).

Training/finetuning fails, since no implementation for CTC/RNNT for now (CPU will be extremely slow). CTC is not yet planned in PyTorch: pytorch/pytorch#77764

[titu1994]

Well training is not important for such huge models on MPS - the slowdown from moving from cuda to MPS is pretty huge even for inference