🐢 This is a fork of the tortoise-tts-fast fork

This fork is exactly the same as the original tortoise-tts-fast but adds deepspeed for even faster inference speed

The original tortoise repo has many of these changes merged in but it does not allow you to import finetuned tortoise models

---

Speeding up TorToiSe inference 5x

This is a working project to drastically boost the performance of TorToiSe, without modifying the base models. Expect speedups of 5~10x

This repo adds the following config options for TorToiSe for faster inference:

• New September 2023 - Deepspeed
• (--kv_cache) enabling of KV cache for MUCH faster GPT sampling
• (--half) half precision inference where possible
• (--sampler dpm++2m) DPM-Solver samplers for better diffusion
• (disable with --low_vram) option to toggle cpu offloading, for high vram users

All changes in this fork are licensed under the AGPL. For avoidance beyond all doubt, the following statement is added as a comment to all changed code files:

AGPL: a notification must be added stating that changes have been made to that file.

Installation

pure python install

The installation process is identical to the original tortoise-tts repo.

⚠️ Unfortunately, deepspeed is not supported on Windows (which is ironic because the repo is operated by Microsoft). Luckily, linux with WSL is not too painful to setup and integrates pretty well with VS code.

conda create --name tortoise python=3.9 numba inflect
conda activate tortoise
conda install pytorch==2.0.0 torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia
conda install transformers=4.29.2
pip install -r requirements.txt 
pip3 install git+https://github.com/152334H/BigVGAN.git
pip install deepspeed=0.10.2 # linux/WSL only

If you are on windows, you will also need to install pysoundfile: conda install -c conda-forge pysoundfile

For more information please reference the original tortoise-tts-fast README or original tortoise repo README