Skip to content

如何将akpt模型保存成mindir #1968

@Pikachu-Yj

Description

@Pikachu-Yj

1.Describe the current behavior / 问题描述

我需要将DeepSeek-R1-Distill-Qwen-1.5B的akpt模型转成mindir模型。模型仓库地址是https://modelers.cn/models/MindSpore-Lab/DeepSeek-R1-Distill-Qwen-1.5B。

2.Environment / 环境信息

  • Hardware Environment / 硬件环境:
后端类型 硬件具体类别
CPU Linux CPU
  • Software Environment / 软件环境:
Software Version
MindSpore MindSpore 2.5.0
Mindnlp 0.4.0
Python Python 3.10.16
OS platform Ubuntu 24.04
GCC/Compiler version 11.4.0

3.Steps to reproduce the issue / 重现步骤

下面是在CPU上的执行代码。我需要将akpt模型转成mindir模型。模型仓库地址是https://modelers.cn/models/MindSpore-Lab/DeepSeek-R1-Distill-Qwen-1.5B。

from mindnlp.transformers import AutoModelForCausalLM, AutoTokenizer
import mindspore as ms

model_name = 'DeepSeek-R1-Distill-Qwen-1.5B'
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "请介绍一下你自己"
inputs = tokenizer(prompt, return_tensors="ms")
outputs = model.generate(
    input_ids=inputs["input_ids"],
    attention_mask=inputs["attention_mask"],
    max_length=512,
    num_return_sequences=1)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print("Generated Response:", response)

在mindnlp文档中,我看到了一个接口mindnlp.engine.export。地址https://mindnlp.cqu.ai/zh/api/engine/export/
但是在使用的时候报错。发现该接口在源码中没有实现。

from mindnlp.transformers import AutoModelForCausalLM, AutoTokenizer
import mindspore as ms

model_name = 'DeepSeek-R1-Distill-Qwen-1.5B'
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "请介绍一下你自己"
inputs = tokenizer(prompt, return_tensors="ms")

mindnlp.engine.export(model, 
                      inputs,
                      file_name="DeepSeek-R1-Distill-Qwen-1.5B",
                      file_format="MINDIR",
                      dynamic_axes={"input_ids": [0], "attention_mask": [0]})

报错:

<class 'mindnlp.transformers.models.qwen2.modeling_qwen2.Qwen2Model'>
Traceback (most recent call last):
  File "/home/pikachu/Project/llm/to_mindir.py", line 15, in <module>
    mindnlp.engine.export(model,
AttributeError: module 'mindnlp' has no attribute 'engine'

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions