convert.py

import torch
import safetensors.torch
from transformers import T5Tokenizer, T5EncoderModel

#https://huggingface.co/Tencent-Hunyuan/HunyuanDiT-v1.2/blob/main/t2i/model/pytorch_model_ema.pt
input_diffusion = "pytorch_model_ema.pt"

#https://huggingface.co/Tencent-Hunyuan/HunyuanDiT-v1.2/tree/main/t2i/clip_text_encoder
input_bert = "./clip_text_encoder/pytorch_model.bin"

#https://huggingface.co/stabilityai/sdxl-vae/blob/main/sdxl_vae.safetensors
# or
#https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors
input_vae = "sdxl_vae.safetensors"

output = "hunyuan_dit_1.2.safetensors"

bert_sd = torch.load(input_bert, weights_only=True)

mt5 = T5EncoderModel.from_pretrained("google/mt5-xl")
tokenizer = T5Tokenizer.from_pretrained("google/mt5-xl")

sp_model = torch.ByteTensor(list(tokenizer.sp_model.serialized_model_proto()))
t5_sd = mt5.state_dict()

out_sd = {}

out_sd["text_encoders.mt5xl.spiece_model"] = sp_model

for k in t5_sd:
    out_sd["text_encoders.mt5xl.transformer.{}".format(k)] = t5_sd[k].half()

for k in bert_sd:
    if not k.startswith("visual."):
        out_sd["text_encoders.hydit_clip.transformer.{}".format(k)] = bert_sd[k].half()

hydit = torch.load(input_diffusion, weights_only=True)
for k in hydit:
    out_sd["model.{}".format(k)] = hydit[k].half()


vae_sd = safetensors.torch.load_file(input_vae)

for k in vae_sd:
    out_sd["vae.{}".format(k)] = vae_sd[k].half()

safetensors.torch.save_file(out_sd, output)