[Bugfix] Fix the bug of torch_npu that raising segment fault when enable pin_memory while creating a tensor

[shen-shanshan]

What this PR does / why we need it?

Fix the bug in torch 2.5.1 that raising segment fault when enable pin_memory while creating a tensor using torch.tensor.

Does this PR introduce any user-facing change?

How was this patch tested?

[wangxiyuan]

please update the commit message to make it readable.

[wangxiyuan]

when ACL_OP_INIT_MODE can be set to 0? If it's not hardcode to 1, please add it to env.py. Otherwise it's fine adding here.

[shen-shanshan]

@wangxiyuan I have asked @mengwei805 , this value can be fixed at 1 directly.

[shen-shanshan]

please update the commit message to make it readable.

ok.

[wangxiyuan]

Thanks

[wangxiyuan]

@ganyi1996ppo please double check this PR as well. If it's fine. Feel free to merge it.

[ganyi1996ppo]

Have no clue about this, why we need ACL_OP_INIT_MODE to ensure pin_memory's functionality, why its not open by default? can we find a runtime expert to explain this option? @wangxiyuan @shen-shanshan @wuhuikx

[shen-shanshan]

Have no clue about this, why we need ACL_OP_INIT_MODE to ensure pin_memory's functionality, why its not open by default? can we find a runtime expert to explain this option? @wangxiyuan @shen-shanshan @wuhuikx

@ganyi1996ppo ACL_OP_INIT_MODE is default as 0. When torch_npu initialize, it will read this env var and use this to init ACL OP. Can you supplement more details? @mengwei805

[wangxiyuan]

@shen-shanshan The question form @ganyi1996ppo is that you should add more comment about the bug. For example is there any related issue from torch-npu? how torch-npu deal with this bug? what and how ACL_OP_INIT_MODE works in torch-npu. We can't just add this ENV for a quite simple reason that torch-npu has a bug

[shen-shanshan]

@shen-shanshan The question form @ganyi1996ppo is that you should add more comment about the bug. For example is there any related issue from torch-npu? how torch-npu deal with this bug? what and how ACL_OP_INIT_MODE works in torch-npu. We can't just add this ENV for a quite simple reason that torch-npu has a bug

That is complex, give me some time dive into the source code of torch_npu to find the answer.