[BUG] Ascend C compilation fails because paged attention uses reserved block_idx

[ETOwang]

版本

• InfiniCore commit: 46ca684
• Hardware: Huawei Ascend 910B3
• CANN: 8.1.RC1
• OS: Linux
• Python: Python 3.10 or later
• Build command: python scripts/install.py --ascend-npu=y

问题描述

InfiniCore cannot compile the Ascend paged attention kernels because
block_idx is used as a local variable name.

Ascend C defines block_idx as a compiler built-in variable/macro.
As a result, the local variable declaration is expanded by the
preprocessor and rejected by the Ascend C compiler.

复现步骤

1. git clone --recursive https://github.com/InfiniTensor/InfiniCore.git
2. cd InfiniCore
3. source scripts/set_env_linux.sh
4. python scripts/install.py --ascend-npu=y

实际结果

The build fails with errors :

The errors occur in:

• src/infiniop/ops/paged_attention/ascend/paged_attention_ascend_kernel.cpp
• src/infiniop/ops/paged_attention_prefill/ascend/paged_attention_prefill_ascend_kernel.cpp

预期结果

The Ascend paged attention kernels should compile successfully.

根因分析

Ascend C reserves block_idx as a built-in variable.
The affected variables in these kernels represent logical paged
attention block indices, rather than Ascend hardware block IDs.

Therefore, they should not be replaced with AscendC::GetBlockIdx().
They should be renamed to a non-reserved identifier while preserving
the existing indexing logic.

建议修复

Rename the logical page-block variables from block_idx to
page_block_idx in the affected Ascend kernels.

This is a source-level naming fix and does not change the algorithm,
operator interface, memory layout, or output semantics.