Scaffold and operate an autonomous Triton GPU-kernel optimization loop on FlashInfer-Bench + Modal, modeled on the MLSys-2026 DSA contest winner (Dogacel/auto-gpu-kernel). Provides a ready-to-run template (CLAUDE.md, .claude/commands/{optimize,benchmark,log-experiment}, .claude/agents/{profiler,research,workload-inspector}, Modal benchmark + paired A/B harness, packed-solution scripts), plus customization guidance for retargeting to a different FlashInfer-Bench kernel definition. Use when the user wants to (1) bootstrap an autonomous Triton-kernel optimizer for any FlashInfer-Bench kernel, (2) set up an experiment-logged optimization loop with sub-agents for profiling, workload inspection, and plateau-breaking research, (3) run cloud-only (no local GPU) GPU benchmarking via Modal B200, or (4) reproduce or adapt the auto-gpu-kernel architecture for FlashInfer competition tracks or similar kernel-optimization workflows.
Provides guidance for writing and benchmarking optimized CUDA kernels for NVIDIA GPUs (H100, A100, T4) targeting HuggingFace diffusers and transformers libraries. Supports models like LTX-Video, Stable Diffusion, LLaMA, Mistral, and Qwen. Includes integration with HuggingFace Kernels Hub (get_kernel) for loading pre-compiled kernels. Includes benchmarking scripts to compare kernel performance against baseline implementations. Use for kernel types: attention, rmsnorm, rope, adaln, geglu, benchmark, transformers, diffusers, huggingface-kernels, get_kernel.
Investigate a consistently failing SGLang CI test to find the root cause — whether it is a code regression from a specific PR, a hardware/runner-specific issue, or an environment change. Optionally reproduces the failure on a remote GPU server over SSH + Docker. Use when an SGLang CI test is failing on main, you need to bisect which PR introduced a regression, you suspect a runner-specific or GPU-specific issue, or you want to reproduce a CI failure on a remote server.
Step-by-step tutorial for adding new CUDA kernels to FlashInfer
Write a Statement on Education (Teaching Statement / Teaching Philosophy / Teaching Vision / "onderwijsvisie") for an academic faculty application, especially in Computer Science / AI / Data Science at research universities. Covers structure (1–3 page format), evaluation rubrics, evidence-based pedagogy, supervision frameworks, US vs Dutch/EU framing, TU Delft programme curricula (BSc CSE, MSc CS, MSc DSAIT) with verified course mapping, real-world example anatomies, and a section-by-section drafting workflow that produces a credible, specific, non-generic statement. Use whenever an applicant must write or revise a teaching statement, education statement, teaching vision, teaching philosophy, or onderwijsvisie — particularly for TU Delft, ETH, KTH, Cambridge, Imperial, or any EU/NL CS faculty position.
Generate an Obsidian-format CUDA StudyVault for learning NVIDIA's GPU programming stack. Three modes: (1) Curriculum mode (default, no source files needed) builds a prereq-ordered vault around the 6-topic learning path: CUDA kernels (threads/blocks/warps, memory hierarchy, TMA, WGMMA, cp.async), CUTLASS + CuTe (Layout/Tensor, MMA atoms, GEMM pipelines), cuTile (Python tile DSL), open-gpu-kernel-modules (RM, GSP firmware, UVM, kernel-open), NCCL (collectives, NVLink-SHARP, transports), and NVSHMEM (PGAS, symmetric heap, IBGDA); (2) Codebase mode generates an onboarding vault from a real CUDA/CUTLASS/NCCL/NVSHMEM/driver source tree; (3) Document mode turns NVIDIA whitepapers, GTC slide decks, and PDFs into study notes. Mode is auto-detected from CWD markers. Use when the user wants to learn CUDA from scratch, onboard onto a GPU codebase, build an exercise list for CUDA topics, or preprocess NVIDIA PDFs/docs into structured study material. Pair with the `cuda-tutor` skill to quiz against the resulting vault.
Interactive quiz tutor for a CUDA StudyVault built by `cuda-tutor-setup`. Delivers 4-question rounds with concept-level proficiency tracking (🟥/🟨/🟩/🟦/⬜) across the 6 CUDA learning topics: CUDA kernels (threads/blocks/warps, memory hierarchy, TMA, WGMMA, cp.async), CUTLASS + CuTe (Layout/Stride/Tensor, MMA atoms, GEMM pipelines), cuTile (Python-first tile DSL), open-gpu-kernel-modules (RM, GSP firmware, UVM, kernel-open layout), NCCL (collectives, topology, NVLink-SHARP, transports), and NVSHMEM (PGAS, symmetric heap, IBGDA, on-stream API). Use when the user wants to (1) take a diagnostic CUDA assessment, (2) drill weak GPU concepts, (3) study a specific CUDA topic, (4) review the learning dashboard, or says things like "quiz me on CUDA", "test my CUTLASS knowledge", "drill NCCL", "/cuda-tutor", "퀴즈".
Generate an Obsidian-format Triton StudyVault for learning OpenAI Triton's GPU programming stack. Three modes: (1) Curriculum mode (default, no source files needed) builds a prereq-ordered vault around 6 topics: Triton basics, tiling & autotuning, matmul patterns (tiled GEMM, split-K, FP8 `tl.dot`), attention & reductions (FlashAttention), compiler internals (TTIR/TTGIR/LLIR, NVIDIA/AMD backends, WGMMA + TMA lowering), and ecosystem/production (`torch.compile`/Inductor, AOT, `proton`); (2) Codebase mode generates an onboarding vault from a real Triton/Inductor source tree; (3) Document mode turns Triton docs, blog posts, and tutorial PDFs into study notes. Mode is auto-detected from CWD markers. Use when the user wants to learn Triton from scratch, onboard onto a Triton codebase, build a Triton exercise list, or preprocess Triton PDFs/docs into study material. Pair with the `triton-tutor` skill to quiz against the resulting vault.