gemma-2-9b-it-untied

Created with:

from transformers import AutoModelForCausalLM
import torch

# 加载模型并解除共享
model = AutoModelForCausalLM.from_pretrained("google/gemma-2-9b-it", tie_word_embeddings=False)

# 解除共享的 lm_head 和 embed_tokens 权重
model.lm_head.weight.data = model.model.embed_tokens.weight.data.clone()

# 将模型转换为 bf16 格式
model = model.to(dtype=torch.bfloat16)

# 指定保存路径
untied_model_dir = "mergekit/output"

# 保存解除共享且为 bf16 格式的模型
model.save_pretrained(untied_model_dir)
model.config.save_pretrained(untied_model_dir)

I didn't copy tokenizer from the original model, do it yourself if you want.