[Skill] build-optimize: Analyze and reduce AITER build time

[sunway513]

Skill

build-optimize

Priority: P1 (High — full build takes 60+ minutes, blocking development velocity)

Motivation

AITER's full build takes 60+ minutes, dominated by CK-Tile kernel compilation (Flash Attention alone generates 2782 HIP files, plus MOE GEMM variants). For development iteration, most developers only need a subset of kernels. A skill that analyzes build time bottlenecks and generates minimal build configurations would dramatically improve development velocity. This also supports the CK-free build strategy where CK kernels shift to FlyDSL as the backend.

What This Skill Should Do

1. Profile build time: Analyze build logs to identify which targets/kernels consume the most compilation time
2. Dependency analysis: Map which Python operators depend on which compiled kernels (CK, ASM, Triton)
3. Suggest minimal configs: Given a target workload (e.g., "DeepSeek-V3 decode only"), output the minimal set of kernels needed
4. Generate build flags: Produce CMake/build flags that skip unnecessary CK kernels for the target configuration
5. CK-free feasibility: For each CK kernel, assess whether a Triton or FlyDSL replacement exists or is planned

Acceptance Criteria

• Can parse build logs and produce a ranked list of compilation time by target
• Can identify CK-Tile kernels that dominate build time (FA, MOE GEMM)
• Can generate minimal build config for a specified model/workload
• Produces CMake flags or environment variables to control kernel compilation scope
• Documents which CK kernels have Triton/FlyDSL replacements available
• Achieves <15 min build time for CK-free configuration