fla-ascend-performance

Guidelines for Ascend NPU kernel / Triton-Ascend backend performance work in the FLA repo. Covers profiling with torch_npu, PipeUtilization/MemoryUB CSV analysis, Cube/Vector/MTE/UB bottleneck diagnosis, and kernel optimization (UB tiling, grid splits, fusion/split, varlen, G_T_CONTIG gate loading, constexpr DMA-path split / TAIL_MODE, extract_slice, MTE OOB, int32 address overflow, tl.cast vs constexpr .to, make_block_ptr int32 offsets, correctness gates, tl.dot left-operand clobber). NPU kernels must not use num_warps/num_stages. Per-kernel catalog: references/cases.md (incl. causal_conv1d core-grid). Use when working on NPU profiling, kernel_details/op_statistic, aic_metrics, fla triton_ascend backends (ops or modules), g transpose stride-1, UB overflow, dual-path DCE, grid limits, int64 pointer math, tl.dot reuse, or Ascend performance.

Quality

92%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Passed

No findings from the security scan

Quality

Content

92%Weight 40%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

A thoroughly structured, highly actionable optimization skill with a clear sequenced workflow, explicit validation feedback loops, and clean progressive disclosure into verified reference files. Its one minor gap is occasional repetition of the same reference anchors that could be tightened.

Dimension	Reasoning	Score
Conciseness	Dense and assumes Claude's competence—no padding explaining what an NPU or profiler is—and nearly every token is an actionable specific. A few passages repeat the same cases.md anchors and could be tightened, keeping it just short of fully lean.	4 / 5
Actionability	Provides copy-paste-ready bash commands, Python snippets, exact function names (fla.utils.ascend_ub_manager, compute_row_tile_block_size), env constants (ASCEND_MAX_GRID_DIM=65535), and concrete code patterns covering the common optimization cases.	5 / 5
Workflow Clarity	Clear six-step progress checklist plus sequenced sections 1–5 with explicit validation checkpoints—frozen pytest gate, 'on failure, stop—do not claim speedups', re-profile to confirm metrics, and a reclassify-bottleneck feedback loop.	5 / 5
Progressive Disclosure	SKILL.md is an overview with well-signaled, one-level-deep references (reference.md, cases.md, g-contiguous-loading.md, TRAPS.md) and scripts, all of which exist; detailed case notes and metrics are appropriately split out, with a 'Related files' navigation section.	5 / 5
	Total	19 / 20 Passed

Description

92%Weight 40%Scale 1-5

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

A highly specific, well-scoped description that explicitly covers what the skill does and when to use it, anchored to a narrow NPU performance niche. The only weak spot is trigger-term naturalness, which leans on domain jargon over plain user phrasings.

Dimension	Reasoning	Score
Specificity	Lists multiple concrete actions—profiling with torch_npu, PipeUtilization/MemoryUB CSV analysis, Cube/Vector/MTE/UB bottleneck diagnosis, and named kernel-optimization sub-areas (UB tiling, grid splits, G_T_CONTIG, constexpr DMA split/TAIL_MODE, tl.dot clobber)—giving comprehensive coverage.	5 / 5
Completeness	Explicitly answers both 'what' ('Guidelines for... Covers profiling... and kernel optimization') and 'when' ('Use when working on NPU profiling... or Ascend performance') with concrete trigger phrases.	5 / 5
Trigger Term Quality	The 'Use when working on...' clause covers many relevant triggers (NPU profiling, kernel_details/op_statistic, aic_metrics, triton_ascend backends, g transpose stride-1, int64 pointer math), but the terms skew heavily toward internal jargon rather than natural user phrasings, with a few common variations absent.	4 / 5
Distinctiveness Conflict Risk	Scoped tightly to Ascend NPU / Triton-Ascend / FLA repo triton_ascend directories—a clear niche with distinct triggers and minimal overlap with other skills.	5 / 5
	Total	19 / 20 Passed

Validation

100%

Warnings & errors only

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 16 / 16 Passed

Validation for skill structure

No warnings or errors.

Repository: fla-org/flash-linear-attention
Path: .agents/skills/fla-ascend-performance/SKILL.md
Commit: f470469

Reviewed: 4 days ago

Table of Contents

Discovery Implementation Validation

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.