tortoise-tts icon indicating copy to clipboard operation
tortoise-tts copied to clipboard

autoregressive sample is 256 at both standard and high_quality preset

Open MagicChakram opened this issue 1 year ago • 1 comments

OS: windows 10 Device: Rtx 3060m (6gb / 3840 cuda cores) Cuda: 12

I have a strange issue, when trying to read a text file - and using --preset standard and --preset high_quality ends up generating both 256 samples per part. While reading same text and same drivers etc but on 3080 all as ok: high quality needs more samples than standard.

I think it has to do with, maybe, cuda count? Or ram? 3080M - has 8gb and 6144 cores 3060M - has 6gb and 3840 cores

edit: using "read.py" mode, not "read_fast" mode. Thanks

MagicChakram avatar Oct 30 '23 12:10 MagicChakram

figured that this is in the code. So the issue is then - 3080 is producing way less samples than intended, according to this, and 3060 is fine?

 presets = {
            'ultra_fast': {'num_autoregressive_samples': 16, 'diffusion_iterations': 30, 'cond_free': False},
            'fast': {'num_autoregressive_samples': 96, 'diffusion_iterations': 80},
            'standard': {'num_autoregressive_samples': 256, 'diffusion_iterations': 200},
            'high_quality': {'num_autoregressive_samples': 256, 'diffusion_iterations': 400},
        }

MagicChakram avatar Nov 01 '23 05:11 MagicChakram