确定性问题描述：[功能问题-EZ1001][Real]Real算子存在 功能问题-EZ1001 通路：torchair golden版本：20261109 golden链接：https://gitcode.com/opencann/operator-artifacts/blob/master/Real/golden文件/real.py cann版本：20261109 ttk版本：20261109 用例链接：https://gitcode.com/sleeveen/operator_test_case/tree/main/LongTail/Cases/Functional_Accuracy/torchair/torch.real 用例名：aclnnReal_L1_003 错误码：EZ1001 复现命令：python3 -m ttk e2e -i 失败用例集_Real_torchair_功能失败-EZ1001.csv --dl=1 --plugin /home/CI/golden/torch.real/ --device-whitelist=0 -c -d --aclgraph --xpu-perf -t aclnnReal_L1_003 问题现象：
【用例 aclnnReal_L1_003】 2026-11-10 20:03:33,564 [ERROR] [1786289 aclnnReal_L1_003 OnGraphAclgraph] [graph_runner.py:540]: Graph aclgraph execution failed: copy_d2d_baseformat_opapi:../torch_npu/csrc/aten/ops/op_api/CopyKernelOpApi.cpp:125 NPU function error: call aclnnInplaceCopy failed, error code is 161002 [ERROR] 2026-11-10-20:03:27 (PID:1786289, Device:0, RankID:-1) ERR00100 PTA call acl api failed. [ERROR] 2026-11-10-20:03:27 (PID:1786289, Device:0, RankID:-1) ERR00100 PTA call acl api failed. 2026-11-10 20:03:34,219 [ERROR] [1786289 aclnnReal_L1_003 OnGraphCst] [graph_runner.py:496]: Failed to get TorchAir NPU backend: cannot import name 'op_broadcast' from 'torchair._ge_concrete_graph.ge_converter.experimental.hcom_broadcast' (/npu/miniconda3/envs/CI_py310_aclnn/lib/python3.10/site-packages/torch_npu/dynamo/torchair/_ge_concrete_graph/ge_converter/experimental/hcom_broadcast.py)
CANNBot实际失败分析： 1) 失败根因与类别：框架/底层库支持缺失问题，非Real算子缺陷。aclgraph图模式在捕获阶段对输入做静态拷贝（dst_tensors[0].copy_(src_tensors[0])，acl_graph.py:1535）时调用aclnnInplaceCopy，该算子不支持DT_COMPLEX32（错误码161002/EZ1001）；EAGER模式已PASS，说明aclnnReal本身无问题。CST/DYN模式失败是另一独立问题：torchair与torch_npu版本不匹配（cannot import name 'op_broadcast'），属环境问题。 2) 关键日志证据：
dst_tensors[0].copy_(src_tensors[0])
aclnnInplaceCopy
cannot import name 'op_broadcast'
AclNN_Parameter_Error(EZ1001): Tensor self not implemented for DT_COMPLEX32, should be in dtype support list [...]
acl_graph.py:1535 dst_tensors[0].copy_(src_tensors[0])
Failed to get TorchAir NPU backend: cannot import name 'op_broadcast'
EAGER ... PASS
Golden Dtypes: ('float16',)
('complex32',)
迁移来源：https://gitcode.com/sleeveen/operator_test_case/issues/252
附件下载链接：
确定性问题描述：[功能问题-EZ1001][Real]Real算子存在 功能问题-EZ1001
通路：torchair
golden版本：20261109
golden链接：https://gitcode.com/opencann/operator-artifacts/blob/master/Real/golden文件/real.py
cann版本：20261109
ttk版本：20261109
用例链接：https://gitcode.com/sleeveen/operator_test_case/tree/main/LongTail/Cases/Functional_Accuracy/torchair/torch.real
用例名：aclnnReal_L1_003
错误码：EZ1001
复现命令：python3 -m ttk e2e -i 失败用例集_Real_torchair_功能失败-EZ1001.csv --dl=1 --plugin /home/CI/golden/torch.real/ --device-whitelist=0 -c -d --aclgraph --xpu-perf -t aclnnReal_L1_003
问题现象：
CANNBot实际失败分析：
1) 失败根因与类别：框架/底层库支持缺失问题，非Real算子缺陷。aclgraph图模式在捕获阶段对输入做静态拷贝（
dst_tensors[0].copy_(src_tensors[0])，acl_graph.py:1535）时调用aclnnInplaceCopy，该算子不支持DT_COMPLEX32（错误码161002/EZ1001）；EAGER模式已PASS，说明aclnnReal本身无问题。CST/DYN模式失败是另一独立问题：torchair与torch_npu版本不匹配（cannot import name 'op_broadcast'），属环境问题。2) 关键日志证据：
AclNN_Parameter_Error(EZ1001): Tensor self not implemented for DT_COMPLEX32, should be in dtype support list [...]acl_graph.py:1535 dst_tensors[0].copy_(src_tensors[0])触发aclnnInplaceCopy失败Failed to get TorchAir NPU backend: cannot import name 'op_broadcast'EAGER ... PASS，Golden Dtypes: ('float16',)，输入dtype('complex32',)3) 定位建议：①确认CANN/torch_npu版本对complex32 d2d copy的支持，推动补齐aclnnInplaceCopy的DT_COMPLEX32支持；②排查torchair/torch_npu版本一致性，修复op_broadcast导入错误；③短期可在用例集规避complex32图模式，或图模式输入拷贝前先转float16。
开发责任人：叶钊
PL：叶钊
迁移来源：https://gitcode.com/sleeveen/operator_test_case/issues/252