Skip to content

issue/1519 - fix Ascend paged attention block index name collision - #1520

Open
ETOwang wants to merge 1 commit into
InfiniTensor:mainfrom
ETOwang:issue/1519
Open

issue/1519 - fix Ascend paged attention block index name collision#1520
ETOwang wants to merge 1 commit into
InfiniTensor:mainfrom
ETOwang:issue/1519

Conversation

@ETOwang

@ETOwang ETOwang commented Aug 22, 2026

Copy link
Copy Markdown

Related issue

Closes #1519

Summary

  • Rename the logical paged-attention block index in the Ascend decode kernel.
  • Rename the logical paged-attention block index in the Ascend prefill kernel.
  • Avoid collision with the Ascend C built-in block_idx identifier.

Root cause

Ascend C reserves block_idx as a compiler built-in variable.
The affected variables are logical page-block indices, not hardware
block IDs. The compiler therefore expands the declarations and reports
illegal initializer.

Implementation

Renamed the local variables from block_idx to page_block_idx.

No algorithm, operator interface, indexing formula, memory layout, or
output behavior was changed.

Testing

compile success

Environment

  • Hardware: Huawei Ascend 910B3
  • CANN: 8.1.RC1
  • OS: Linux

@ETOwang
ETOwang requested a review from a team August 22, 2026 12:54
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

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

1 participant