kali20gakki/msAgent / skills/profiler/ascend-computation-analysis
ascend-computation-analysis
VerifiedAnalyze Ascend NPU computation-side profiling data for single-card runs or a selected rank from multi-card runs. Use this skill when the user asks to diagnose computation bottlenecks, AI Core / AI Vector / AICPU hotspots, dynamic shape overhead, block dim issues, redundant TransData/Transpose/Cast, cross-stream waits, frequency/runtime-state issues, or fusion opportunities.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 11.2 KB |
| references/failure_handbook.md | markdown | 11.5 KB |
| references/fusion_rules.md | markdown | 3.6 KB |
references/fusion_rules.md
View on GitHub# Ascend Computation Fusion Rules
Use this reference only when the main workflow finds repeated operator sequences, small-operator overhead, vector-heavy expressions, redundant data movement, or advisor fusion findings.
Do not recommend fusion by name alone. Fusion is useful only when it changes a meaningful bottleneck.
## Core Principles
### Generality Principle
Prioritize fusion patterns that appear across multiple mainstream models or repeated hot paths in the current workload.
Avoid recommending a custom fusion just because a one-off sequence exists.
### Benefit Principle
Evaluate whether fusion changes the bottleneck.
Possible benefit layers:
- Host benefit: fewer operators, lower dispatch overhead, less graph scheduling overhead.
- Device benefit: fewer intermediate tensors, less load/store traffic, better producer-consumer pipeline behavior.
If the fused operators remain bound by the same pipeline with little movement or dispatch overhead removed, expected benefit may be low.
## Deep Fusion Vs Shallow Fusion
Deep fusion:
- Producer and consumer can work in block-level or tile-level handoff.
- Intermediate data may avoid full materialization.
- More likely to improve device-side efficiency.
Shallow fusion:
- Multiple logical operations are packaged together.
- Intermediate data or pipeline behavior may not change much.
- More likely to help host dispatch than device execution.
Use this distinction when explaining expected benefit.
## Fusion Opportunity Signals
Look for:
- Repeated short operator sequences with meaningful cumulative time.
- Many small elementwise or vector operators around AI Core kernels.
- Repeated `Cast`, `TransData`, or `Transpose` around a stable compute pattern.
- Host-bound behavior from many small operators.
Do not rely on:
- A single occurrence.
- Low cumulative time.
- Pattern name without measured impact.
## Fusion Patterns To Check
When these appear in hot repeated paths, check whether the official fused operator or replacement path is applicable:
- `RotaryMul` / `RotaryMulGrad`
- `RmsNorm` / `RmsNormGrad`
- `ScaledMaskedSoftmax` / `ScaledMaskedSoftmaxGrad`
- `MatmulAllReduce`
- `FlashAttentionScore`
- `SwiGlu`
Treat the list as a prompt for investigation, not as an automatic recommendation.
## Cube-Vector, Vector-Vector, And Small-Operator Fusion
Cube-Vector fusion:
- Combines compute and vector/data-movement style work.
- Benefit is more likely when it reduces materialization or improves cache/locality.
- Low data volume may limit benefit.
Vector-Vector fusion:
- Combines vector-style operations.
- Often helps when it removes repeated load, store, or cast overhead.
Small-operator fusion:
- Useful when launch/dispatch overhead dominates.
- Especially relevant for high-frequency tiny operators.
Always tie the recommendation to measured count and cumulative time.
## When Not To Recommend Fusion
Avoid strong fusion recommendations when:
- The sequence has low cumulative duration.
- The sequence is not on the critical path.
- The fused pattern would still be bound by the same dominant pipeline.
- The input data lacks enough operator sequence evidence.
- The recommendation would require model-code changes that are not justified by measured impact.
## Output Guidance
When recommending fusion, state:
- The repeated sequence or pattern.
- Its count and cumulative time when available.
- The expected benefit layer: host dispatch, memory movement, pipeline handoff, or dtype/format cleanup.
- The concrete experiment to try.
- The metric that should improve.
Do not generate replacement code without concrete model code context.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |