rocm-KDA-pilot
rocm-KDA-pilot contient 2 skills collectées depuis jhinpan, avec une couverture métier par dépôt et des pages de détail sur le site.
Skills dans ce dépôt
Orchestrator for ANY FlyDSL GPU kernel optimization loop on AMD ROCm -- FlashAttention, GEMM, MoE (incl. MXFP4/fp8), and other FlyDSL kernels, on gfx942 / gfx950 / gfx1250. Use whenever the task is tuning or optimizing a FlyDSL kernel (raise MFU/TFLOPS, cut latency, fix a profiled bottleneck), not just FlashAttention. Combines ROCmKernelWiki, flyprof (flydsl-rocprof-cli), rocm-report-skill, benchmark discipline, and Humanize RLCR. For FlyDSL kernels profiling MUST go through flyprof + rocm-report-skill -- never hand-roll a rocprofv3 collection script.
Use to turn a ROCm/FlyDSL kernel-optimization issue into a ready-to-run Humanize task -- provision a fresh FlyDSL-lab worktree+branch, snapshot the issue into a draft, generate the plan, and stop at the human review gate. Fixed /sgl-workspace layout. Pairs with rocm-kda-pilot.