A personal knowledge base of things I learn — GPU kernels, hardware architecture, and adjacent systems topics. Each domain lives in its own top-level folder; raw source material is kept under _sources/ for provenance.
| Topic | What it covers |
|---|---|
| hipkitten/ | HipKittens — HazyResearch's AMD CDNA3/4 tile-DSL: schedules, swizzling, chiplet-aware scheduling, and how aiter uses it. |
| ck-dsl-runbook/ | The canonical CK DSL Optimization Runbook: the full lever catalog and method (§1–21 + decision tree): diagnose → lever families → autotune → failure modes → gfx950/CDNA4 reference. Includes the skinny-M decode GEMM worked walkthrough (4.38× slower → rocBLAS parity on MI355X) as a case study. |
| flash-attention-v4/ | FlashAttention-4 — Tri Dao's Blackwell (B200) attention kernel: asymmetric hardware scaling, software-emulated exp, conditional softmax rescaling, the 5-stage warp-specialized pipeline, TMEM & 2-CTA MMA. |
| flydsl-jdbba/ | FlyDSL jagged grouped-GEMM (HSTU jdbba) on MI355X/gfx950: verified varlen layout-API idioms (bounded buffer descriptors, runtime base-offset views, device scalar reads), the memory-bound optimization case study (XCD remap, epilogue vectorization, 16×16×32 atom, XCD-aware persistent visiting order, plus verified dead-ends), and the measurement discipline for benchmarking vs an autotuning baseline. |
| flydsl-dev-env/ | Docker dev environment for aiter FlyDSL + Triton/Gluon kernel work, running non-root: which ROCm PyTorch base image and why (the torch ≥ 2.9.1 floor gates Gluon), a verified Dockerfile + launch scripts, VS Code attach, and running Claude Code inside the container. Centres on the failures that are silent: the render group being mandatory once you drop root (zero GPUs, no error), HOME deciding whether your git/ssh/Claude identity exists at all, setup.py develop downgrading flydsl under you, missing jq disabling Claude hooks, root-owned named volumes, and CK headers needed even with ENABLE_CK=0. Verified on 8×MI355X. |
| fp8-mqa-logits/ | The DeepSeek lightning-indexer FP8 MQA-logits kernel on gfx942/CDNA3: from-first-principles attention→sparse-attention→indexer background (MQA/GQA/MLA, FP8 FN-vs-FNUZ, CDNA3 MFMA tiling, why the head-reduction dominates, DeepGEMM fake-quant validation), the FlyDSL dense-kernel implementation plan, and hard-won gfx942 FlyDSL idioms (fp8-byte→f32 via v_cvt_f32_fp8, runtime scf.for to dodge slow MLIR canonicalization) with the landed scalar correctness-first status. |
| fp8-mqa-logits/ | The DeepSeek lightning-indexer FP8 MQA-logits kernel on gfx942/CDNA3: from-first-principles attention→sparse-attention→indexer background (MQA/GQA/MLA, FP8 FN-vs-FNUZ, CDNA3 MFMA tiling, why the head-reduction dominates, DeepGEMM fake-quant validation), the FlyDSL dense-kernel implementation plan, and hard-won gfx942 FlyDSL idioms (fp8-byte→f32 via v_cvt_f32_fp8, the landed mfma_r<N>_w<M> variant family reaching 1.45–1.65× Triton, plus post-landing optimizations: i64 offset overflow fix, scale hoisting, 128-bit fragment loads, and a grid.y num_splits KV-column split). |
| flydsl-profiling/ | Profiling instruction for FlyDSL kernels. |
| gated-delta-rule/ | The gated delta rule / KDA linear-attention family and the inter-chunk K5 state-scan kernel (SILOTIGER-859): from linear attention through the delta rule, Mamba-2/SSD, Gated DeltaNet (scalar gate — Qwen3-Next, Qwen3.5-MoE, OLMo), KDA (per-channel gate — Kimi K3), the chunkwise parallel K1–K6 stage decomposition, and gfx942 vs gfx950 hardware differences (MFMA K-step, transposing LDS, LDS budget). |
To wire this KB into Claude Code so it's referenceable in every session (and to install the kb-add-topic skill machine-wide), follow SETUP.md. Run it once per machine or after cloning somewhere new.
This repo ships a Claude Code skill, kb-add-topic, that turns raw study material (chat exports, transcripts, paper notes) into a structured, cross-linked topic folder following the conventions below. Invoke it with /kb-add-topic (or just ask Claude to "add a topic to the KB") while working in this repo.
- Numeric prefixes on folders/files indicate reading order (
00-fundamentals first). - Each note has YAML frontmatter (
name,description,source) and a footer linking related notes. - Notes derived from a longer transcript cite the exact source file and line range; the transcript is preserved verbatim in
_sources/. - Factual claims about external papers/models carry a
Sources:footer with primary-source links.
Content is derived from public sources (papers, public repos, vendor docs) and personal study notes. See each topic's README.md for its specific source list.