convert.py

import re
import torch

# load the original XTTS weights (requires coqui installed for the ['config'] entry)
src = torch.load("./models/xtts/model.pth", map_location="cpu")['model']

dst = {
	"ar": "./models/tortoise/autoregressive.pth",
	"df": "./models/tortoise/diffusion_decoder.pth",
}

for model, path in dst.items():
	dst[model] = torch.load(path, map_location="cpu")
	torch.save( dst[model], f'{path}.bkp' )

# copy
regexes = {
	"ar": r'^gpt\.',
	"df": r'^diffusion_decoder\.',
}
for k, v in src.items():
	for model, regex in regexes.items():
		if re.match(regex, k):
			key = re.sub(regex, "", k)
			if key not in dst[model]:
				continue
			print(f"Writing {k} into {key}")
			dst[model][key] = v
			break

# save
torch.save(dst['ar'], "./models/tortoise/autoregressive.xtts.pth")
torch.save(dst['df'], "./models/tortoise/diffusion_decoder.xtts.pth")