Support HPU on ASR/TTS

[Spycsh]

Type of Change

hpu hardware enabling

Description

Now ASR/TTS should support HPU/CPU/GPU

Expected Behavior & Potential Risk

Now ASR/TTS should support HPU/CPU/GPU

How has this PR been tested?

Run the two UT on HPU

/intel-extension-for-transformers/intel_extension_for_transformers/neural_chat/tests/audio# python test_asr.py
/intel-extension-for-transformers/intel_extension_for_transformers/neural_chat/tests/audio# python test_tts.py

Dependency Change?

None

[hshen14]

@hshen14 without LD_PRELOAD, ipex.cpu.runtime.CPUPool will also not be used anymore in the main code, otherwise it will cause an error that expect that env variable to exist. Please re-check that.

Yes, we don't need ipex.cpu.runtime neither. We can just set the core binding using numactl -m 0 -C cores to get the similar effect.

[Spycsh]

@hshen14 without LD_PRELOAD, ipex.cpu.runtime.CPUPool will also not be used anymore in the main code, otherwise it will cause an error that expect that env variable to exist. Please re-check that.

Yes, we don't need ipex.cpu.runtime neither. We can just set the core binding using numactl -m 0 -C cores to get the similar effect.

Absolutely. I'm just concerning that running the neuralchat service shell script will normally bind to one socket, which may be different cores than 24. But normally I believe it is slight performance difference.

[hshen14]

As we discussed, there might be no additional perf optimizations on HPU at this moment. Can you move forward this PR?