Individual skill
optimize-triton-ascend-grouped-attention-access
Optimize irregular KV-cache access in Triton-Ascend grouped decode attention by vector-loading contiguous inner dimensions, assembling discrete outer rows with CANN slice extensions, and transposing UB tiles for tl.dot. Use when an agent ports grouped-query decode attention whose K or V cache addresses are indirect and naive low-axis-discrete loads degrade to scalar memory access.
skills/optimize-triton-ascend-grouped-attention-access/SKILL.md
Source commit 4ab5ee7adf47755ecd42723f55f4c58e38be95a7
What was checked
- Matches the observed repository HEAD
- Yes
- Scanner rules are current
- Yes
- Repository scan complete
- No
- Skill scan complete
- No
- Observed behavior
- Instructions only
- Last checked
- Nov 4, 2026
These are observations at the recorded commit, not a guarantee of safety. Review the source before running it.
Install this skill
Choose an agent in your Stack or copy a command below. Commands fetch the pinned commit and copy only this skill directory. Existing destinations are refused. Copying a command does not install anything.
An install command is unavailable: current rules, complete repository and skill scans, an eligible behavior level and a permitted license are required.
Repository context
- Repository stars (not skill ratings)
- 189
- Repository last updated
- Aug 15, 2026
- License
- Apache-2.0
Observed agent integration: No specific integration evidence. Standard SKILL.md installation does not imply verified compatibility.
- ai Repository topic
- ascend Repository topic
- ascend-npu Repository topic
- npu Repository topic
- triton Repository topic
- triton-ascend Repository topic
- claude-code Repository topic
- codex Repository topic
- opencode Repository topic
- engineering Derived category
Files and excerpt
File metadata and permitted, redacted excerpts come from the recorded scan. Source links are pinned to that commit.
Read excerpt
# Optimize Triton-Ascend Grouped Attention Access
## Goal
Preserve vectorized contiguous loads while gathering KV-cache rows selected by an indirect token-location vector.
## Workflow
1. Classify each target tile by which axis is discrete and which axis is contiguous.
2. Load a discrete outer row and its contiguous head dimension as one vector.
3. Assemble BLOCK_N rows into an on-chip tensor with extension.insert_slice.
4. Transpose the assembled K tile when tl.dot requires HEAD_DIM by BLOCK_N.
5. Load V directly as BLOCK_N by VALUE_DIM when its inner value dimension is contiguous.
6. Feed the tiles into numerically stable online softmax across KV splits.