【合入来源】

https://gitcode.com/Ascend/pytorch/issues/3682

• 需求
• 问题单
• issue/工单
• 重构优化
• 资料更新

【修改方案】

移除 torch_npu 对 6 个 PyTorch 模块方法（nn.Module.share_memory 与 torch.jit.ScriptModule.add_module / bfloat16 / register_buffer / register_parameter / register_module）的 monkey-patch 拦截机制，并整体删除torch_npu/utils/npu_intercept.py。其中仍被使用的 CANN 包检测逻辑（_cann_package_check() 与 cann_pytorch_version_map）迁移至 torch_npu/_init/registry/registry_manager.py（其唯一调用点所在）。- test/npu/test_unsupport_api.py：删除 6 项 *_runtimeerror 拦截测试；test_Module_share_memory_runtimeerror 改为断言 NPU 上正常执行；test/jit/test_script_module.py：TestScriptModuleShareMemory.test_share_memory_on_npu_raises改为断言 NPU 上正常执行；TestScriptModuleMetadata 的 2 项 raise 测试删除。

【资料变更】

不涉及

【接口变更】

不涉及

【功能验证】

说明测试场景，测试方法。如果本次测试方式与常规单元测试不同，请详细说明您的测试步骤
新增/变更内容是否已新增/适配UT测试用例看护，并补充测试自验证截图

candidate 的 9 项失败全部为原拦截 raise 测试：

• share_memory / bfloat16：不再抛错（功能可用）——预期行为变更；
• add_module / register_module：原生 RuntimeError（消息不同）；
• register_parameter / register_buffer：原生 AttributeError（类型不同）。

官方用例对照

对进行削减的6个api在npu和gpu环境下进行调用，获取其输出结果，观察削减后两个环境中代码运行结果是否一致

npu运行结果 （Ascend 910B4，PyTorch: 2.13.0）

结果说明：这里主要考虑行为的一致，pass/fail仅表示是否会报错，若fail，则看报错信息是否一致。

削减后

============================================================
1. 构造时调用 (ScriptModule __init__ 中)
============================================================
  PASS | add_module/register_buffer/register_parameter/register_module in __init__

============================================================
2. traced ScriptModule 构造后调用
============================================================
  FAIL | add_module on traced
       | RuntimeError: Cannot re-assign modules in a ScriptModule with non-scripted module, tried to replace existing modul
  PASS | bfloat16 on traced
  FAIL | register_buffer on traced
       | AttributeError: cannot assign buffer before Module.__init__() call
  FAIL | register_parameter on traced
       | AttributeError: cannot assign parameter before Module.__init__() call
  FAIL | register_module on traced
       | RuntimeError: Cannot re-assign modules in a ScriptModule with non-scripted module, tried to replace existing modul

============================================================
3. scripted ScriptModule 构造后调用
============================================================
  FAIL | add_module on scripted
       | RuntimeError: Cannot re-assign modules in a ScriptModule with non-scripted module, tried to replace existing modul
  FAIL | bfloat16 on scripted
       | RuntimeError: bfloat16 is not supported on ScriptModules
  FAIL | register_buffer on scripted
       | RuntimeError: Can't add a new parameter after ScriptModule construction. Tried to add 'b
  FAIL | register_parameter on scripted
       | RuntimeError: Can't add a new parameter after ScriptModule construction. Tried to add 'p
  FAIL | register_module on scripted
       | RuntimeError: Cannot re-assign modules in a ScriptModule with non-scripted module, tried to replace existing modul

============================================================
4. share_memory on nn.Module
============================================================
  PASS | share_memory (no-op, return self, device stays npu)

cuda运行结果（torch==2.13.0+cu126）

============================================================
1. 构造时调用 (ScriptModule __init__ 中)
============================================================
  PASS | add_module/register_buffer/register_parameter/register_module in __init__

============================================================
2. traced ScriptModule 构造后调用
============================================================
  FAIL | add_module on traced
       | RuntimeError: Cannot re-assign modules in a ScriptModule with non-scripted module, tried to replace existing modul
  PASS | bfloat16 on traced
  FAIL | register_buffer on traced
       | AttributeError: cannot assign buffer before Module.__init__() call
  FAIL | register_parameter on traced
       | AttributeError: cannot assign parameter before Module.__init__() call
  FAIL | register_module on traced
       | RuntimeError: Cannot re-assign modules in a ScriptModule with non-scripted module, tried to replace existing modul

============================================================
3. scripted ScriptModule 构造后调用
============================================================
  FAIL | add_module on scripted
       | RuntimeError: Cannot re-assign modules in a ScriptModule with non-scripted module, tried to replace existing modul
  FAIL | bfloat16 on scripted
       | RuntimeError: bfloat16 is not supported on ScriptModules
  FAIL | register_buffer on scripted
       | RuntimeError: Can't add a new parameter after ScriptModule construction. Tried to add 'b
  FAIL | register_parameter on scripted
       | RuntimeError: Can't add a new parameter after ScriptModule construction. Tried to add 'p
  FAIL | register_module on scripted
       | RuntimeError: Cannot re-assign modules in a ScriptModule with non-scripted module, tried to replace existing modul

============================================================
4. share_memory on nn.Module
============================================================
  PASS | share_memory (no-op, return self, device stays cuda)

削减后api调用的测试结果与cuda上一致

【CheckList】

PR提交人对以下CheckList自检项进行全量自检，自检通过或不涉及，均修改 [ ] 为 [x]

• 代码注释完备，正********确记录错误日志
• 代码实现进行了返回值、空指针等校验
• PR标题正确使用类型标签，如：feat、fix、refactor、docs、test等
• PR持续集成流水线（CI）执行通过，代码检查无异常