MindIR推理模型导出时不支持列表/字典推导式

sjtu-weimang · 2025 年9 月 1 日 03:00

我在导出MindIR格式的推理模型时，遇到了如下错误：

该错误表明将代码编译为计算图时，mindspore框架无法处理代码中的列表推导式和字典推导式，根据错误报告中建议将jit_syntax_level 设置为LAX也无法解决。请问有没有在不展开推导式的情况下导出推理模型的办法？mindspore版本：2.6.0

chengxiaoli · 2025 年9 月 1 日 03:48

用户您好，欢迎使用mindspore，已经收到您的问题，会尽快分析答复~

llc · 2025 年9 月 2 日 09:10

用户您好，可以麻烦提供下完整的报错截图，以及脚本目录下的rank0/om目录下的analyze_fail.ir文件吗？

sjtu-weimang · 2025 年9 月 2 日 11:18

报错如下：
[WARNING] CORE(1196275,7b46ffae3440,python):2025-09-02-18:56:47.031.197 [mindspore/core/utils/ms_context.cc:528] GetJitLevel] Pynative mode can not set jit_level to O2, use O0 instead.
[WARNING] UTILS(1196275,7b46ffae3440,python):2025-09-02-18:56:48.183.876 [mindspore/ccsrc/utils/comm_manager.cc:80] GetInstance] CommManager instance for CPU not found, return default instance.
Traceback (most recent call last):
File “/root/DPFT-mindspore/model.py”, line 46, in
ms.export(
File “/root/miniconda3/envs/mindspore_env/lib/python3.9/site-packages/mindspore/train/serialization.py”, line 2077, in export
_export(net, file_name, file_format, *inputs, **kwargs)
File “/root/miniconda3/envs/mindspore_env/lib/python3.9/site-packages/mindspore/train/serialization.py”, line 2133, in _export
_save_mindir(net, file_name, *inputs, **kwargs)
File “/root/miniconda3/envs/mindspore_env/lib/python3.9/site-packages/mindspore/train/serialization.py”, line 2373, in _save_mindir
mindir_stream, net_dict = _cell_info(net, incremental, *inputs)
File “/root/miniconda3/envs/mindspore_env/lib/python3.9/site-packages/mindspore/train/serialization.py”, line 2354, in _cell_info
graph_id, _ = _executor.compile(net, *inputs, phase=phase_name, do_convert=False)
File “/root/miniconda3/envs/mindspore_env/lib/python3.9/site-packages/mindspore/common/api.py”, line 1967, in compile
result = self._graph_executor.compile(obj, args, kwargs, phase)
ValueError: When handling script ‘call_func_str(input_0,) in graph mode’, the inputs should be constant, but found variable ‘input_0’ to be nonconstant. Try to set jit_syntax_level to LAX.

C++ Call Stack: (For framework developers)

mindspore/ccsrc/pipeline/jit/ps/static_analysis/prim.cc:3625 CheckInterpretInput

The Traceback of Net Construct Code:

0 In file /root/DPFT-mindspore/models/dpft.py:250, 18~60

        batch=[features[input] for input in self.inputs],
              ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

1 In file /root/DPFT-mindspore/models/dpft.py:240, 19~92

    features = {input: self.embeddings[input](features[input]) for input in self.inputs}
               ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

2 In file /root/DPFT-mindspore/models/dpft.py:236, 19~87

    features = {input: self.necks[input](features[input]) for input in self.inputs}
               ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

3 In file /root/DPFT-mindspore/models/dpft.py:230~233, 19~9

    features = {

4 In file /root/DPFT-mindspore/models/dpft.py:224, 19~88

    features = {input: self.backbones[input](batch[input]) for input in self.inputs}
               ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

5 In file /root/DPFT-mindspore/models/dpft.py:224, 27~62

    features = {input: self.backbones[input](batch[input]) for input in self.inputs}
                       ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

6 In file /root/DPFT-mindspore/models/backbones/resnet.py:80~81, 8~46

    if self.channel_last:

7 In file /root/DPFT-mindspore/models/backbones/resnet.py:81, 12~46

        x = ops.transpose(x, (0, 3, 1, 2))  # (B, H, W, C) -> (B, C, H, W)
        ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

8 In file /root/DPFT-mindspore/models/dpft.py:224, 27~41

    features = {input: self.backbones[input](batch[input]) for input in self.inputs}
                       ^~~~~~~~~~~~~~

9 In file /root/DPFT-mindspore/models/backbones/resnet.py:88~90, 8~27

    for  i,layer in enumerate(self.body):
    ^

10 In file /root/DPFT-mindspore/models/dpft.py:224, 27~41

    features = {input: self.backbones[input](batch[input]) for input in self.inputs}
                       ^~~~~~~~~~~~~~

11 In file /root/DPFT-mindspore/models/backbones/resnet.py:88~90, 8~27

    for  i,layer in enumerate(self.body):
    ^

12 In file /root/DPFT-mindspore/models/dpft.py:224, 27~41

    features = {input: self.backbones[input](batch[input]) for input in self.inputs}
                       ^~~~~~~~~~~~~~

13 In file /root/DPFT-mindspore/models/backbones/resnet.py:88~90, 8~27

    for  i,layer in enumerate(self.body):
    ^

14 In file /root/DPFT-mindspore/models/dpft.py:224, 27~41

    features = {input: self.backbones[input](batch[input]) for input in self.inputs}
                       ^~~~~~~~~~~~~~

15 In file /root/DPFT-mindspore/models/backbones/resnet.py:88~90, 8~27

    for  i,layer in enumerate(self.body):
    ^

16 In file /root/DPFT-mindspore/models/dpft.py:224, 27~41

    features = {input: self.backbones[input](batch[input]) for input in self.inputs}
                       ^~~~~~~~~~~~~~

17 In file /root/DPFT-mindspore/models/backbones/resnet.py:88~90, 8~27

    for  i,layer in enumerate(self.body):
    ^

18 In file /root/DPFT-mindspore/models/dpft.py:224, 27~41

    features = {input: self.backbones[input](batch[input]) for input in self.inputs}
                       ^~~~~~~~~~~~~~

19 In file /root/DPFT-mindspore/models/backbones/resnet.py:88~90, 8~27

    for  i,layer in enumerate(self.body):
    ^

20 In file /root/DPFT-mindspore/models/backbones/resnet.py:93~94, 8~54

    if self.channel_last:

21 In file /root/DPFT-mindspore/models/backbones/resnet.py:94, 12~54

        features = self._to_channel_last(features)
        ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

22 In file /root/DPFT-mindspore/models/backbones/resnet.py:94, 23~54

        features = self._to_channel_last(features)
                   ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

23 In file /root/DPFT-mindspore/models/backbones/resnet.py:75, 15~92

    return OrderedDict({k: ops.transpose(v, (0, 2, 3, 1)) for k, v in features.items()})
           ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

(See file ‘/root/DPFT-mindspore/rank_0/om/analyze_fail.ir’ for more details. Get instructions about analyze_fail.ir at 搜索 | 昇思MindSpore社区)
ir文件如下：
analyze_fail.zip (16.5 KB)

llc · 2025 年9 月 3 日 11:20

用户您好，根据日志和ir图可以看到，

您这里call了self._to_channel_last，但是这个属性实际上是个dict

不可以被call。您可以检查下您的代码逻辑。

zhouyifengCode · 2025 年9 月 3 日 11:20

你这边的那个features是个字典吗？我记得是静态图语法不支持这种写法的，而导出mindir需要支持静态图语法；有没有可以直接复现这个问题的简单代码，发出来一起看看？

sjtu-weimang · 2025 年9 月 3 日 12:20

feature是一个通过列表推导式生成的字典，在mindspore中可以通过load_param_into_net 将权重加载到网络中，并在graph mode下推理，但是无法导出为mindIR， When handling script ‘call_func_str(input_0,) in graph mode’, the inputs should be constant, but found variable ‘input_0’ to be nonconstant. Try to set jit_syntax_level to LAX.但是inputs的形状是确定的，所以features的形状也就确定了

sjtu-weimang · 2025 年9 月 3 日 12:20

这是一个类的方法，OrderedDict 是返回值

llc · 2025 年9 月 4 日 01:33

用户您好，ordereddict属于第三方库，无法在图内表达，不支持mindir导出。

llc · 2025 年9 月 4 日 12:15

用户您好，有生成之前提到的analyzefail.ir文件吗

sjtu-weimang · 2025 年9 月 5 日 00:17

analyze_fail.zip (27.1 KB)

llc · 2025 年9 月 5 日 01:38

用户您好，您这里shape图内表达有误是因为静态图内不支持使用dict()创建dict。代码位置在 /root/DPFT-mindspore/models/fusers/mpfusion.py:748, 19~64/ layerDict =dict(zip(self.mpfusion.values(), self.heads))/

如果您要导出mindir，建议您设置jit_syntax_level为STRICT以获取更准确的报错信息。

sjtu-weimang · 2025 年9 月 5 日 02:39

zip()和set() 也不能用吗？

llc · 2025 年9 月 5 日 02:43

用户您好，zip可以，set不行

chengxiaoli · 2025 年9 月 11 日 01:07

用户您好，MindSpore支撑人已经分析并给出了问题的原因，由于较长时间未看到您采纳回答，这里版主将进行采纳回答的结帖操作，如果还其他疑问请发新帖子提问，谢谢支持~

system · 2025 年9 月 11 日 02:07

此话题已在最后回复的 60 分钟后被自动关闭。不再允许新回复。

话题	回复	浏览量
MindSpore报错The graph generated form MindIR is not support to execute in the PynativeMode,please convert to the GraphMode 模型训练-Model Training	4	2025 年8 月 20 日
mindyolo在ckpt模型转为ONNX模型时报错推理经验-Inference Experience	3	2025 年8 月 11 日
MindSpore应用系列：树莓派上运行YOLOv5进行实时目标检测（五）应用实践-Application Cases	30	2025 年6 月 10 日
mindspore推理报错NameError:The name 'LTM' is not defined, or not supported in graph mode.及解决推理经验-Inference Experience	4	2025 年8 月 1 日
MindSpore应用系列：树莓派上运行YOLOv5进行实时目标检测（二）应用实践-Application Cases	18	2025 年6 月 10 日