变更摘要

本次变更将各 torch_extension 算子的算子加载时机从模块导入期改为首次调用时：删除模块级别的 _op_module = builder.load()（或裸调用 builder.load()），改为在算子执行函数内部按需调用 builder.load()，从而避免整包导入即触发 JIT 编译。涉及 block_sparse_attention、all_gather_quant_matmul、apply_rotary_pos_emb、apply_rotary_pos_emb_grad 等算子模块，其中 block_sparse_attention.py 在正式目录与 experimental 目录下各有一份同步修改。其余改动为导入顺序整理与代码格式规整，不改变对外行为。

主要改动

• 算子加载时机延迟到首次调用: 在 block_sparse_attention.py（含 attention/ 与 experimental/attention/ 两份）中删除模块级 _op_module = _op_builder.load()，并在 block_sparse_attention 函数内改为 op_module = _op_builder.load() 后再调用 op_module.npu_block_sparse_attention(...)。
• all_gather_quant_matmul 同样按需加载: 删除模块级 _op_module = _all_gather_quant_matmul_op_builder.load()，在 _npu_all_gather_quant_matmul 中改为先 op_module = _all_gather_quant_matmul_op_builder.load()，再调用 op_module.npu_all_gather_quant_matmul(...)。
• apply_rotary_pos_emb 去除导入期加载: 删除模块级 apply_rotary_pos_emb_op_builder.load() 调用，ApplyRotaryPosEmbOpBuilder 的构造与调用路径不再在导入时触发加载。
• apply_rotary_pos_emb_grad 去除导入期加载: 删除模块级 apply_rotary_pos_emb_grad_op_builder.load() 调用，加载行为不再于模块导入阶段发生。
• 导入顺序与格式整理: 多个文件统一将 cann_ops_transformer.op_builder 的导入调整到 torch.library 之前，并合并了多行 raise ValueError、张量构造及函数调用表达式为单行。

CLA Signature Pass

weihao18, thanks for your pull request. All authors of the commits have signed the CLA. 👍

Thanks for your pull-request.
The full list of commands accepted by me can be found at here.
You can get sig-info at here.
You can self-configure the PR merge rules for this repository. For more details, please refer to Here.
For more, you also can visit HICANN.

PR Approval Progress

✅ Congratulations! All modules have met the lgtm and approve requirements.

Module Approval Details

💡 Tip:

• Committer can comment /approve or /lgtm
• Commenting /approve implies both code review (lgtm) and intent to merge (approve)

/compile

/compile

/lgtm
/approve

/lgtm
/approve

/lgtm

/lgtm