[Installation]: 安装编译后的 MindIE-SD 后，运行测试，出现libopapi.so的问题-MindIE-SD-AtomGit

已关闭

[Installation]: 安装编译后的 MindIE-SD 后，运行测试，出现libopapi.so的问题 #62

zhangtao_work创建于 3月12日关闭于 3月18日

zhangtao_work

3月12日创建

在提交新问题之前，请确保您已经在社区中搜索过相关问题，并使用了社区中提供的资源/工具后，仍未找到满意的解决方式。

环境信息

操作系统 OpenEuler
昇腾硬件信息 910B3
CANN软件版本 8.5.0
安装的对应软件版本
- MindIESD: Main分支（5c574baec9f911747d4906d8d40748eda97772cb）
- Torch 2.8.0

您是如何安装操作的

编译

python3 setup.py bdist_wheel

安装

pip install dist/*.whl

测试 Laser Attention

(torch280_py310_diffusion) [root@hostname-xxi4w tests]# python3 layers/flash_attn/test_attn_forward.py
....2026-03-12 05:01:25.286+00:00 UTC [354762] [281473298419744] [MindIE-SD] [ERROR] exception.py:19: [MIE06E000001] Parameters invalid. Unsupported FA type: 'test'. Supported values: {'prompt_flash_attn', 'fused_attn_score', 'ascend_laser_attention'}
.EE......
======================================================================
ERROR: test_attn_forward_manual_la_bnsd (__main__.TestAttentionFunc)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "./MindIE-SD/tests/layers/flash_attn/test_attn_forward.py", line 229, in test_attn_forward_manual_la_bnsd
    out_fused = attention_forward(
  File "/root/miniconda3/envs/torch280_py310_diffusion/lib/python3.10/site-packages/mindiesd/layers/flash_attn/attention_forward.py", line 85, in attention_forward
    return attn_func(query, key, value, attn_mask, scale)
  File "/root/miniconda3/envs/torch280_py310_diffusion/lib/python3.10/site-packages/mindiesd/layers/flash_attn/ascend_laser_attention.py", line 111, in forward_attn_bnsd
    _, output1 = ops.laser_attention(
  File "/root/miniconda3/envs/torch280_py310_diffusion/lib/python3.10/site-packages/mindiesd/layers/_custom_ops.py", line 34, in laser_attention
    return getattr(torch.ops.mindiesd, "la")(
  File "/root/miniconda3/envs/torch280_py310_diffusion/lib/python3.10/site-packages/torch/_ops.py", line 1243, in __call__
    return self._op(*args, **kwargs)
RuntimeError: aclnnLaserAttention or aclnnLaserAttentionGetWorkspaceSize not in libopapi.so, or libopapi.so not found.

======================================================================
ERROR: test_attn_forward_manual_la_bsnd (__main__.TestAttentionFunc)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "./02_MindIE-SD/MindIE-SD/tests/layers/flash_attn/test_attn_forward.py", line 128, in test_attn_forward_manual_la_bsnd
    out_fused = attention_forward(
  File "/root/miniconda3/envs/torch280_py310_diffusion/lib/python3.10/site-packages/mindiesd/layers/flash_attn/attention_forward.py", line 85, in attention_forward
    return attn_func(query, key, value, attn_mask, scale)
  File "/root/miniconda3/envs/torch280_py310_diffusion/lib/python3.10/site-packages/mindiesd/layers/flash_attn/ascend_laser_attention.py", line 111, in forward_attn_bnsd
    _, output1 = ops.laser_attention(
  File "/root/miniconda3/envs/torch280_py310_diffusion/lib/python3.10/site-packages/mindiesd/layers/_custom_ops.py", line 34, in laser_attention
    return getattr(torch.ops.mindiesd, "la")(
  File "/root/miniconda3/envs/torch280_py310_diffusion/lib/python3.10/site-packages/torch/_ops.py", line 1243, in __call__
    return self._op(*args, **kwargs)
RuntimeError: aclnnLaserAttention or aclnnLaserAttentionGetWorkspaceSize not in libopapi.so, or libopapi.so not found.

欢迎加入社区，感谢您对社区的贡献 🎉!

ascend-robot成员

3月12日添加了label：installation

zhangtao_work

3月12日修改了issue 的描述

pyf

3月13日评论：

我也遇到一样的问题：

python ada_test.py 
Traceback (most recent call last):
  File "/home/MindIE-SD/ada_test.py", line 77, in <module>
    sct, smask = get_mask(q, k, v, scale, head_num, is_causal, input_layout, keep_sink, keep_recent, sparsity,
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/MindIE-SD/ada_test.py", line 41, in get_mask
    smask, sct = get_estimate_mask(
                 ^^^^^^^^^^^^^^^^^^
  File "/home/MindIE-SD/mindiesd/layers/flash_attn/sparse_flash_attn_ada_bsa.py", line 31, in get_estimate_mask
    smask, sct = ops.sparse_block_estimate(
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/MindIE-SD/mindiesd/layers/_custom_ops.py", line 209, in sparse_block_estimate
    return getattr(torch.ops.mindiesd, "sparse_block_estimate")(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/root/miniconda3/lib/python3.11/site-packages/torch/_ops.py", line 1243, in __call__
    return self._op(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: aclnnSparseBlockEstimate or aclnnSparseBlockEstimateGetWorkspaceSize not in libopapi.so, or libopapi.so not found.

环境信息如下：
操作系统 OpenEuler
昇腾硬件信息 910B4
CANN软件版本 8.5.0
安装的对应软件版本
MindIESD: Main分支（5c574baec9f911747d4906d8d40748eda97772cb）
Torch 2.8.0

guowenna1成员

3月13日评论：

export ASCEND_CUSTOM_OPP_PATH=/usr/local/python3.11.13/lib/python3.11/site-packages/mindiesd/ops/vendors/aie_ascendc/
先参考这个命令设置算子自定义路径，正在同步排查原因

M mazhixin00_00成员

3月13日将 guowenna1 设为负责人

M mazhixin00_00成员

3月13日关联了看板：MindIE-SD

guowenna1成员

3月13日评论：

https://gitcode.com/Ascend/MindIE-SD/pull/207 已修复

M mazhixin00_00成员

3月18日 issue状态由 TODO 改变为 DONE

M mazhixin00_00成员

3月18日关闭了 issue

ascend-robot成员

3月18日添加了label：resolved

Canlin Guo

3月19日评论：

I can reproduce this issue in the latest main and dev:

 File "/root/vllm-workspace/vllm-omni/vllm_omni/diffusion/attention/backends/abstract.py", line 99, in forward
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715]     return self.forward_npu(query, key, value, attn_metadata)
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715]            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715]   File "/root/vllm-workspace/vllm-omni/vllm_omni/diffusion/attention/backends/flash_attn.py", line 203, in forward_npu
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715]     output = attention_forward(
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715]              ^^^^^^^^^^^^^^^^^^
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715]   File "/root/vllm-workspace/.venv/lib/python3.11/site-packages/mindiesd/layers/flash_attn/attention_forward.py", line 85, in attention_forward
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715]     return attn_func(query, key, value, attn_mask, scale)
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715]            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715]   File "/root/vllm-workspace/.venv/lib/python3.11/site-packages/mindiesd/layers/flash_attn/ascend_laser_attention.py", line 111, in forward_attn_bnsd
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715]     _, output1 = ops.laser_attention(
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715]                  ^^^^^^^^^^^^^^^^^^^^
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715]   File "/root/vllm-workspace/.venv/lib/python3.11/site-packages/mindiesd/layers/_custom_ops.py", line 34, in laser_attention
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715]     return getattr(torch.ops.mindiesd, "la")(
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715]            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715]   File "/root/vllm-workspace/.venv/lib/python3.11/site-packages/torch/_ops.py", line 1255, in __call__
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715]     return self._op(*args, **kwargs)
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715]            ^^^^^^^^^^^^^^^^^^^^^^^^^
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:715] RuntimeError: aclnnLaserAttention or aclnnLaserAttentionGetWorkspaceSize not in libopapi.so, or libopapi.so not found.
[Stage-0] ERROR 03-19 08:01:45 [diffusion_worker.py:427] Error executing RPC: aclnnLaserAttention or aclnnLaserAttentionGetWorkspaceSize not in libopapi.so, or libopapi.so not found.

C changetheway成员

7月14日移除了看板：MindIE-SD

lijinxi成员

8月31日关联了看板：MindIE-SD