Skip to content

generate_llm_qnn.py / generateLlmIO 在 MAI-UI-2B(qwen3-vl-2b) 上崩溃(deepstack_embeds missing) #4141

@brelance

Description

@brelance

generate_llm_qnn.py / generateLlmIO 在 MAI-UI-2B 上崩溃(deepstack_embeds missing)

Summary

在将 MAI-UI-2B/qwen3-vl 导出为 MNN 后,执行 generate_llm_qnn.py(QNN 转换流程)时 Step1/Step2 发生 Segmentation fault。手动执行 generateLlmIO 时报错 deepstack_embeds is input but not found
随后崩溃。

Environment

  • OS: Linux (x86_64)
  • MNN: 本地编译 build_qnn
  • QNN SDK: 2.42.0.251225(已下载)
  • 模型:Tongyi-MAI/MAI-UI-2B(qwen3-vl-2b)

Model export (Success)

python3 llmexport.py
--path /data1/models/Tongyi-MAI/MAI-UI-2B
--export mnn
--generate_for_npu --seperate_embed --sym
--act_bit 16
--quant_bit 4 --quant_block 64
--omni --hqq
--calib_data /home/junjie/edge-server/calib_prompts.txt
--dst_path /data1/models/MNN/MAI-UI-2B_ForQNN

导出为 mnn 步骤执行成功。


Steps to reproduce

1) QNN 流程(脚本)

python3 MNN/transformers/llm/export/npu/generate_llm_qnn.py \
--model /data1/models/MNN/MAI-UI-2B_ForQNN \
--soc_id 69 --dsp_arch v79 \
--mnn_path /home/junjie/edge-server/MNN/build

Step1: Make IO
Segmentation fault (core dumped)

Cost:  0.3413975238800049  s
Step2: Seperate Model
model: /home/junjie/edge-server/MNN/transformers/llm/export/model/llm.mnn
Segmentation fault (core dumped)
Cost:  0.34470558166503906  s
Step3: Compile to QNN
/home/junjie/edge-server/qairt/2.42.0.251225
Traceback (most recent call last):
File "/home/junjie/edge-server/MNN/build_qnn/../source/backend/qnn/npu_convert.py", line 10, in <module>
    with open(sys.argv[1]) as f:
        ^^^^^^^^^^^^^^^^^
FileNotFoundError: [Errno 2] No such file or directory: 'npu_postreat.json'
Cost:  0.04529619216918945  s
Step4: Move result file to  /home/junjie/edge-server/MNN/transformers/llm/export/model
End

2) 手动执行 generateLlmIO

/home/junjie/edge-server/MNN/build/generateLlmIO \
/data1/models/MNN/MAI-UI-2B_ForQNN \
/home/junjie/edge-server/MNN/transformers/llm/export/tmp/testdir \
128


blockSize=128 in main, 149 
modelPath.c_str()=s /data1/models/MNN/MAI-UI-2B_ForQNN/llm.mnn in main, 153 
llmConfigPath.c_str()=s /data1/models/MNN/MAI-UI-2B_ForQNN/llm_config.json in main, 154 
Failed to create dir /home/junjie/edge-server/MNN/transformers/llm/export/tmp/testdir.
Can't open file:/sys/devices/system/cpu/cpufreq/schedutil/affected_cpus
Can't open file:/sys/devices/system/cpu/cpufreq/boost/affected_cpus
CPU Group: [ 345  65  5  128  194  250  307  373  37  166  222  289  335  75  138  260  317  383  47  176  232  299  355  19  269  381  45  108  174  230  297  353  17  83  146  202  85  325  55  118  184  240  363  27  93  156  212  279  20  290  347  10  77  262  319  49  111  178  234  357  224  87  206  272  329  59  121  188  244  300  367  30  214  148  204  270  327  57  186  242  365  29  95  158  315  280  337  67  7  196  252  309  375  39  101  168  180  41  104  170  227  293  13  142  265  321  51  114  378  237  23  152  209  275  331  61  1  124  190  247  245  112  179  235  358  21  88  150  207  273  122  189  303  301  368  31  98  160  217  283  132  199  255  311  192  239  361  25  91  154  210  277  333  63  3  126  182  249  305  371  35  164  220  287  343  73  136  259  172  33  162  219  285  341  71  134  257  313  43  106  97  229  295  351  15  81  144  200  267  323  53  116  56  46  109  175  231  298  354  18  84  147  203  326  382  119  185  241  364  28  94  157  213  336  66  6  127  117  183  362  26  92  155  211  278  334  64  4  129  193  306  372  36  165  221  288  344  74  137  316  130  120  187  243  366  96  159  215  281  338  68  8  58  197  253  376  102  169  225  291  348  11  78  140  261  195  251  308  374  38  100  167  223  346  76  139  54  318  48  110  177  233  356  86  149  205  271  328  369  22  89  151  208  274  330  60  0  123  246  302  359  32  99  161  218  284  340  70  133  256  312  379  103  216  282  339  69  9  131  198  254  310  377  40  42  226  292  349  12  79  141  264  320  50  113  236  380  248  304  370  34  163  286  342  72  135  258  314  191  44  107  173  296  352  16  82  145  201  268  324  115  105  171  228  294  350  14  80  143  266  322  52  263  181  238  360  24  90  153  276  332  62  2  125 ], 1500000 - 2400000
The device supports: i8sdot:0, fp16:0, i8mm: 0, sve2: 0, sme2: 0
170 tensor [ deepstack_embeds ] is input but not found
Create module error

这里报了Failed to create dir /home/junjie/edge-server/MNN/transformers/llm/export/tmp/testdir,手动mkdir -p /home/junjie/edge-server/MNN/transformers/llm/export/tmp/testdir后报错如下:

blockSize=128 in main, 149 
modelPath.c_str()=s /data1/models/MNN/MAI-UI-2B_ForQNN/llm.mnn in main, 153 
llmConfigPath.c_str()=s /data1/models/MNN/MAI-UI-2B_ForQNN/llm_config.json in main, 154 
Can't open file:/sys/devices/system/cpu/cpufreq/schedutil/affected_cpus
Can't open file:/sys/devices/system/cpu/cpufreq/boost/affected_cpus
CPU Group: [ 345  65  5  128  194  250  307  373  37  166  222  289  335  75  138  260  317  383  47  176  232  299  355  19  269  381  45  108  174  230  297  353  17  83  146  202  85  325  55  118  184  240  363  27  93  156  212  279  20  290  347  10  77  262  319  49  111  178  234  357  224  87  206  272  329  59  121  188  244  300  367  30  214  148  204  270  327  57  186  242  365  29  95  158  315  280  337  67  7  196  252  309  375  39  101  168  180  41  104  170  227  293  13  142  265  321  51  114  378  237  23  152  209  275  331  61  1  124  190  247  245  112  179  235  358  21  88  150  207  273  122  189  303  301  368  31  98  160  217  283  132  199  255  311  192  239  361  25  91  154  210  277  333  63  3  126  182  249  305  371  35  164  220  287  343  73  136  259  172  33  162  219  285  341  71  134  257  313  43  106  97  229  295  351  15  81  144  200  267  323  53  116  56  46  109  175  231  298  354  18  84  147  203  326  382  119  185  241  364  28  94  157  213  336  66  6  127  117  183  362  26  92  155  211  278  334  64  4  129  193  306  372  36  165  221  288  344  74  137  316  130  120  187  243  366  96  159  215  281  338  68  8  58  197  253  376  102  169  225  291  348  11  78  140  261  195  251  308  374  38  100  167  223  346  76  139  54  318  48  110  177  233  356  86  149  205  271  328  369  22  89  151  208  274  330  60  0  123  246  302  359  32  99  161  218  284  340  70  133  256  312  379  103  216  282  339  69  9  131  198  254  310  377  40  42  226  292  349  12  79  141  264  320  50  113  236  380  248  304  370  34  163  286  342  72  135  258  314  191  44  107  173  296  352  16  82  145  201  268  324  115  105  171  228  294  350  14  80  143  266  322  52  263  181  238  360  24  90  153  276  332  62  2  125 ], 1500000 - 2400000
The device supports: i8sdot:0, fp16:0, i8mm: 0, sve2: 0, sme2: 0
170 tensor [ deepstack_embeds ] is input but not found
Create module error

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