Timing with torch Events instead of time.perf_counter gives pretty much the same times

Baseline speed when running question_answering_components.py

{'language model': 1610.3536376953125,
'model name': 'deepset/bert-base-cased-squad2',
'prediction head': 8769.41357421875,
'preproc': 802.9445190429688,
'question': 'When were the first traces of Human life found in France?',
'total': 11183.6533203125}

{'language model': 595.620849609375,
'model name': 'deepset/minilm-uncased-squad2',
'prediction head': 3255.84228515625,
'preproc': 1075.1549682617188,
'question': 'When were the first traces of Human life found in France?',
'total': 4922.5927734375}

{'language model': 1586.070556640625,
'model name': 'deepset/bert-base-cased-squad2',
'prediction head': 108.00390625,
'preproc': 795.5881042480469,
'question': 'How many pretrained models are available in Transformers?',
'total': 2490.778076171875}

{'language model': 596.1339111328125,
'model name': 'deepset/minilm-uncased-squad2',
'prediction head': 89.099609375,
'preproc': 1066.5256958007812,
'question': 'How many pretrained models are available in Transformers?',
'total': 1755.6483154296875}