google / google/gemma.cpp

How to convert safetensors into sbs?

未关闭
#533 4 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看
主要语言
C++
星标
7k
派生
660
平均合并
20 小时 43 分钟
30 天内合并 PR
33

描述

I know it's an old question, but I still have no idea how to finish it.

I got some files after fine-tune and `merge_and_unload`, like below:

```
added_tokens.json config.json generation_config.json model.safetensors special_tokens_map.json
tokenizer.json tokenizer.model tokenizer_config.json
```

I only find the script `python/convert_from_safetensors.py`.
It seems to be created for PaliGemma only, but I still tried it with the error below:

```
python3 python/convert_from_safetensors.py --load_path /model/gemma-3-1b-finetune/merged_gemma/
I0401 15:06:34.645604 140119943138176 convert_from_safetensors.py:423] Restoring pure python logging
I0401 15:06:34.645855 140119943138176 convert_from_safetensors.py:429]
====
Reading from /model/gemma-3-1b-finetune/merged_gemma/ and writing to /tmp/gemmacpp.sbs
====
Traceback (most recent call last):
File "/model/gemma-3-1b-finetune/gemma.cpp/python/convert_from_safetensors.py", line 436, in
app.run(main)
File "/usr/local/lib/python3.11/site-packages/absl/app.py", line 316, in run
_run_main(main, args)
File "/usr/local/lib/python3.11/site-packages/absl/app.py", line 261, in _run_main
sys.exit(main(argv))
^^^^^^^^^^
File "/model/gemma-3-1b-finetune/gemma.cpp/python/convert_from_safetensors.py", line 432, in main
export_paligemma_sbs(load_path, metadata_file, sbs_file)
File "/model/gemma-3-1b-finetune/gemma.cpp/python/convert_from_safetensors.py", line 212, in export_paligemma_sbs
with safetensors.safe_open(file, framework="pt") as f:
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
OSError: No such device (os error 19)
```

Any suggest for the convert scripts?
Or maybe there is another fine-tune guide?

贡献指南

打开贡献指南

调研方向

从 python/convert_from_safetensors.py 以及 traceback 中显示的 export_paligemma_sbs 调用开始。检查脚本如何处理所提供的 fine-tuning 文件,以及为什么 safetensors.safe_open 会报告“No such device.”。Done 应确定适用于这些文件的受支持转换路径,或记录所需的输入和限制。

由索引模型根据 Issue 内容生成。

评估

技术栈
cpp, python
领域
machine-learning, tooling
Issue 类型
缺陷
难度
4/5
预计耗时
3-5 天
活跃度
停滞
描述清晰度
需要澄清
新手友好度
25/100

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。