tortoise-tts/tortoise/api_fast.py attempts to deserialize object on a CUDA device but torch.cuda.is_available() is False

[kentontroy]

BACKGROUND:

The init() function of class TextToSpeech correctly has:

if torch.backends.mps.is_available():
            self.device = torch.device('mps')
.....

The nightly build of PyTorch recognizes the mps device on my Mac M3 Ultra.

PROBLEM:

When running:

time python do_tts.py \
    --output_path results \
    --preset ultra_fast \
    --voice test \
    --text "Frank Herbert is the best science fiction author ever"

An error occurs:

"tortoise-tts/tortoise/api_fast.py attempts to deserialize object on a CUDA device but torch.cuda.is_available() is False"

RESOLUTION THAT WORKED:

To resolve this, I added this to tortoise-tts/tortoise/api_fast.py: 

def __init__(self, autoregressive_batch_size=None, models_dir=MODELS_DIR,
                 enable_redaction=True, kv_cache=False, use_deepspeed=False, half=False, device=None,
                 tokenizer_vocab_file=None, tokenizer_basic=False):


        hifi_model = torch.load(get_model_path('hifidecoder.pth'), map_location=self.device)

The original torch.load method does not specify the map_location.

Please confirm or suggest a better way.

Thanks - very nice project!
kentontroy