faster-whisper: cpu and gpu recognite result differently，gpu recognition all the way '!' and long video cuda out of memory

model: large-v3 question:

1. when I use cpu to transctibe, it can recognite the speech correctly, but when I use gpu to transcribe, it output all ‘!’, and it detected language ‘en’ with probability nan detected language:

2. and when I use gpu, it loaded the model into gpu memory, but when I look at the gpu usage in task manager I see that it is 0% from time to time, is this correct? the gpu usage:

recognite result: cpu: gpu:

this is my code: gpu version:

from faster_whisper import WhisperModel
from changevideospeed import SearchVideoFiles
import pysubs2

def GetSubtitleResult(results : list, audio_path: str) -> tuple:
    subs: pysubs2.SSAFile = pysubs2.load_from_whisper(results)
    file_name = audio_path[:audio_path.rfind('.')]
    
    return (subs, file_name)

def SaveSRT(results : list, audio_path: str) -> None:
    subs, file_name = GetSubtitleResult(results, audio_path)
    subs.save(file_name+'.srt')

def SaveASS(results : list, audio_path: str) -> None:
    subs, file_name = GetSubtitleResult(results, audio_path)
    subs.save(file_name+'.ass')

def LoadModel(model_size: str, download_root: str = 'models/') -> WhisperModel:
    model: WhisperModel = WhisperModel(download_root = download_root, model_size_or_path = model_size,
                            device='cuda', # cpu_threads=12,
                            compute_type='float16',
                            local_files_only = True)
    return model

def Transcribe(model : WhisperModel, audio_path: str, language:str = None) -> list:
    segments, info = model.transcribe(audio = audio_path, language = language)
    segments = list(segments)

    print("Detected language '%s' with probability %f" % (info.language, info.language_probability))
    
    results= []
    for s in segments:
        segment_dict = {'start':s.start,'end':s.end,'text':s.text}
        results.append(segment_dict)
    
    return results

if __name__ == '__main__':
    audio_dir = 'audio'
    file_paths = SearchVideoFiles(audio_dir)
    model_size = 'large-v3'

    model: WhisperModel = LoadModel(model_size)
    for path in file_paths:
        results: list = Transcribe(model, path)
        SaveSRT(results, path)

cpu version: just change the LoadModel function

def LoadModel(model_size: str, download_root: str = 'models/') -> WhisperModel:
    model: WhisperModel = WhisperModel(download_root = download_root, model_size_or_path = model_size,
                            device='cpu', cpu_threads=12,
                            # compute_type='float16',
                            local_files_only = True)
    return model

if anyone can help me, thank you very much!