Awesome-Kernel-Workflows

KO | EN

gitlite — search

#markdown #list #generative-ai #animation #android #tailwindcss #devops #redis #react-native #nlp

Awesome-Kernel-Workflows

(★ 9)

No description available.

File Explorer

ZIP Download

- - ci.yml
- - agent-retry.js
  - arg-guard.js
  - backend-axis.js
  - embedded-eval.js
  - runtime-safe-point.js
  - substrate-invocation.js
  - turn-timeout.js
  - typed-args.js
- - iterative-loop.js
  - search-based.js
  - single-pass.js
  - tree-exploration.js
- - generate-workflow.js
  - print-workflow-prompts.js
  - validate-workflow.js
- genome-trajectory-schema.md
- README.md
- README.zh-CN.md
- workflow-tool-storage-and-observability.md
- - __init__.py
  - _evidence_amd.py
  - _evidence_ascend.py
  - _evidence_metax.py
  - _evidence_nvidia.py
  - _gen_flag_parser.py
  - REGISTRY.md
  - validate_backend.py
- - ADAPTER_CONTRACT.md
  - embedded_eval.js
  - sol_execbench_eval.js
- - integration_registry.json
  - integration_strategist.py
  - pack_sol_candidate.py
  - parse_sol_bench.py
  - README.md
  - ROLLOUT.md
- - sm100-blackwell.md
- - perf_to_evidence.py
  - profiler_registry.json
  - profiling_strategist.py
  - README.md
- - test_anti_cheat.py
  - test_code_integrity.py
  - test_diagnose.py
  - test_driver_conformance.py
  - test_driver_scripts.py
  - test_evidence_nvidia.py
  - test_manifest_parse_v11.py
  - test_method_gate.py
  - test_smoke.py
  - test_validate_backend.py
- - README.md
  - verification_registry.json
  - verification_strategist.py
- anti_cheat.py
- ARCHITECTURE.md
- arg_guard.js
- BACKEND-DRIVER-SDK.md
- CHANGELOG.md
- code_integrity.py
- diagnose.py
- evidence_schema.py
- EXPERIMENT-SOLVER-SET.md
- grounding.js
- IMPROVEMENTS.md
- memory_store.py
- method_gate.py
- README.md
- SOLVER-AUDIT.md
- SOLVER-SDK.md
- verify_insight.py
- iterative-loop.js
- search-based.js
- single-pass.js
- tree-exploration.js
- generate-workflow.js
- validate-workflow.js
- accelopt-kernel-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- adaexplore-kernel-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- ako4x-kernel-optimizer.js
- manifest.yaml
- README.md
- README.zh-CN.md
- argus-kernel-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- ascendc-kernel-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- astra-kernel-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- atrex-kernel-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- automegakernel-megakernel-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- workflows.json
- cuda-agent-kernel-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- cudallm-fsr-kernel-generation.js
- manifest.yaml
- README.md
- README.zh-CN.md
- cupilot-kernel-optimization.js
- manifest.yaml
- README.md
- cutlass-gemm-optimization.js
- manifest.yaml
- optimized_gemm_kernel.py
- README.md
- README.zh-CN.md
- solution_example.json
- - 20260620-llama-cpp-mmq-awk.md
- - triton-run-sh-generic.md
- akw-modification-plan.md
- manifest-schema.yaml
- workflow-fidelity-audit-2026-06-03.md
- fact-kernel-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- gemmptx-gemm-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- generalist-kernel-optimization.js
- manifest.yaml
- README.md
- gpuforecasters-kernel-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- harness-engineering-kernel-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- in-place-patch-optimization.js
- manifest.yaml
- external-resources.md
- kda-kernel-workflow.js
- manifest.yaml
- README.md
- README.zh-CN.md
- keet-kernel-explanation.js
- manifest.yaml
- README.md
- README.zh-CN.md
- kernelagent-triton-synthesis.js
- manifest.yaml
- README.md
- kernelband-kernel-optimization.js
- manifest.yaml
- README.md
- kernelblaster-kernel-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- kernelfoundry-kernel-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- kernelfoundrydx-kernel-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- kernelskill-kernel-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- ksearch-kernel-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- llamacpp-embedded-search.js
- llamacpp-metal-embedded-search.js
- manifest.yaml
- manifest.yaml
- README.md
- README.zh-CN.md
- regrapht-kernel-optimization.js
- add-agent-retry-scaffolding.js
- check-agent-retry-guards.js
- check-canonical-args.js
- check-fidelity-contracts.js
- count-workflows.sh
- install-global-workflows.sh
- llamacpp_metal_register_variant.py
- llamacpp_register_variant.py
- merge-routing-into-manifests.py
- patch-anti-cheat-flags.js
- patch-arg-guard.js
- patch-backend-axis.js
- patch-embedded-eval.js
- patch-genome-report.js
- patch-model-tier.js
- patch-runtime-safe-point.js
- patch-sol-execbench-eval.js
- patch-turn-timeout.js
- patch-typed-args.js
- patch_arg_guard.py
- sgl_register_variant.py
- stage1-roundtrip.sh
- validate-manifests.sh
- manifest.yaml
- README.md
- stark-kernel-optimization.js
- manifest.yaml
- README.md
- README.zh-CN.md
- stitchcuda-kernel-optimization.js
- manifest.yaml
- README.md
- tritorx-operator-generation.js
- - config.defaults.json
  - hardware-facts-sm100.md
  - hardware-facts-sm90.md
- - correctness.py
  - problem_shapes.json
  - README.md
  - tolerances.json
- - bench_confirm.sh
  - bench_screen.sh
  - calibrate.sh
  - gpu_run
  - ncu_profile.sh
  - wslock.py
- - ncu_fingerprint.py
  - ncu_parse.py
  - render_config.py
  - schema.sql
  - wsdb.py
- - cluster-dsmem.md
  - fp8-register-accumulation.md
  - INDEX.md
  - l2-policy-and-prefetch.md
  - persistent-kernel-scheduling.md
  - register-pressure-occupancy.md
  - smem-swizzle-bank-conflicts.md
  - tcgen05-tmem-basics.md
  - tma-bulk-async-copy.md
  - warp-specialization-pingpong.md
  - wgmma-smem-pipelines.md
- manifest.yaml
- README.md
- warpspeed-kernel-search.js
- manifest.yaml
- README.md
- README.zh-CN.md
- xe-forge-kernel-optimization.js
.gitignore
Agent.md
AGENTS.md
CHANGELOG.md
CHANGELOG.zh-CN.md
CONTRIBUTING.md
README.md
README.zh-CN.md
VERSION

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<h1 align="center">Awesome Kernel Workflows</h1>

Curated GPU kernel optimization methods from academia and industry, packaged as reusable <a href="https://docs.anthropic.com/en/docs/claude-code">Claude Code</a> Workflows

English · <a href="README.zh-CN.md">简体中文</a>

---

## What is this?

GPU kernel optimization produces a steady stream of strong work: self-improving loops (e.g. AccelOpt), agentic frameworks (e.g. AKO), and engineering playbooks behind industrial SOTA (FlashInfer, CUTLASS, etc.). These contributions usually ship as **papers plus code repos**. Handing them to a coding agent as-is rarely reproduces the **decision process**—when to profile, how to read NCU, how to accumulate experience, and so on.

This repository aims to:

1. **Collect** — Methods and systems from academia and industry that are suitable for agent-driven kernel optimization.
2. **Distill** — Abstract each method’s core loop (plan / execute / profile / learn, etc.) into explicit phases.
3. **Ship** — Implement them as [Claude Code Workflows](https://docs.anthropic.com/en/docs/claude-code/workflows) (`.js`) you can drop into your kernel project and invoke.

Each subdirectory is **one method or one workflow**, with the workflow definition, parameter notes, and (over time) source links and examples.

---

## Repository layout

```
Awesome-Kernel-Workflows/
├── README.md # English (this file)
├── README.zh-CN.md # 简体中文
├── AccelOpt/ # Self-improving loop with NCU profiling
│ ├── accelopt-kernel-optimization.js
│ └── README.md
├── KEET/ # NCU profile → explanation pipeline
│ ├── keet-kernel-explanation.js
│ └── README.md
├── ARGUS/ # Data-flow invariant guided optimization
│ ├── argus-kernel-optimization.js
│ └── README.md
├── AKO4X/ # Multi-round closed-loop with experience accumulation
│ ├── ako4x-kernel-optimizer.js
│ └── README.md
├── KDA/ # Evidence-driven draft-plan-implement-validate-decide
│ ├── kda-kernel-workflow.js
│ └── README.md
├── AdaExplore/ # MCTS tree search + failure-driven skill memory
│ ├── adaexplore-kernel-optimization.js
│ └── README.md
├── KernelFoundry/ # MAP-Elites evolutionary + meta-prompt co-evolution
│ ├── kernelfoundry-kernel-optimization.js
│ └── README.md
├── KSearch/ # World-model-guided tree search
│ └── ksearch-kernel-optimization.js
├── KernelBand/ # MAB with hardware-aware pruning + clustering
│ ├── kernelband-kernel-optimization.js
│ └── README.md
├── KernelAgent/ # Multi-agent Triton synthesis with parallel verification
│ ├── kernelagent-triton-synthesis.js
│ ├── manifest.yaml
│ └── README.md
├── STARK/ # Multi-agent collaboration + tree search + grounded instruction
│ ├── stark-kernel-optimization.js
│ ├── manifest.yaml
│ └── README.md
├── ReGraphT/ # CUDA reasoning graph + Monte Carlo Graph Search
│ ├── regrapht-kernel-optimization.js
│ ├── manifest.yaml
│ └── README.md
├── Astra/ # Multi-agent optimization for existing CUDA kernels
│ ├── astra-kernel-optimization.js
│ ├── manifest.yaml
│ └── README.md
├── AutoMegaKernel/ # Adapter to AMK megakernel schedule search
│ ├── automegakernel-megakernel-optimization.js
│ ├── manifest.yaml
│ └── README.md
├── HarnessEngineering/ # Frozen-harness profile/implement/official-gate loop
│ ├── harness-engineering-kernel-optimization.js
│ ├── manifest.yaml
│ └── README.md
├── Atrex/ # Strict adapter to Atrex Kernel Agent
│ ├── atrex-kernel-optimization.js
│ ├── manifest.yaml
│ └── README.md
├── CUDALLM/ # Feature Search and Reinforcement for CUDA generation
│ ├── cudallm-fsr-kernel-generation.js
│ ├── manifest.yaml
│ └── README.md
├── CutlassGEMM/ # CUTLASS GEMM multi-config dispatch tuning
│ ├── cutlass-gemm-optimization.js
│ └── README.md
├── GemmPTX/ # GEMM instruction-path optimization via PTX/SASS evidence
│ ├── gemmptx-gemm-optimization.js
│ ├── manifest.yaml
│ └── README.md
├── FACT/ # Compositional CUTLASS pattern synthesis
│ ├── fact-kernel-optimization.js
│ └── README.md
├── GPUForecasters/ # PUCT search with learned speedup forecasting
│ ├── gpuforecasters-kernel-optimization.js
│ └── README.md
├── KernelBlaster/ # Memory-augmented in-context RL for CUDA kernels
│ ├── kernelblaster-kernel-optimization.js
│ └── README.md
├── KernelFoundryDx/ # Diagnosis-driven multi-island Triton evolution
│ ├── kernelfoundrydx-kernel-optimization.js
│ └── README.md
├── KernelSkill/ # Dual-memory multi-agent CUDA optimization
│ ├── kernelskill-kernel-optimization.js
│ └── README.md
├── StitchCUDA/ # Planner/Coder/Verifier CUDA synthesis
│ ├── stitchcuda-kernel-optimization.js
│ └── README.md
├── WarpSpeed/ # Rewindable parallel kernel-search on a multi-GPU node
│ ├── warpspeed-kernel-search.js
│ ├── infra/ # gpu_run mutex launcher, bench tiers, NCU profiling
│ ├── tools/ # wsdb CLI (SQLite tree), config renderer, NCU parse
│ ├── wiki/ # technique pages (tag-indexed, arch-scoped)
│ ├── harness-template/ # project-owned correctness harness scaffold
│ ├── tests/ # mock GPU suite + vm dry-run + node acceptance
│ └── README.md
├── Xe-Forge/ # Multi-stage CoVeR optimization for Intel XPU
│ ├── xe-forge-kernel-optimization.js
│ └── README.md
├── _meta/ # Meta-workflow: paper → workflow generation
│ ├── README.md
│ ├── tools/
│ │ ├── generate-workflow.js
│ │ └── validate-workflow.js
│ ├── templates/
│ │ ├── iterative-loop.js
│ │ ├── search-based.js
│ │ ├── single-pass.js
│ │ └── tree-exploration.js
│ └── manifests/
│ ├── schema.yaml
│ ├── accelopt.yaml
│ ├── keet.yaml
│ ├── argus.yaml
│ └── ksearch.yaml
└── <MoreMethods>/ # PRs welcome
 └── *.js
```

Workflow files follow Claude Code conventions: export `meta` (name, description, phases) and orchestrate multi-step agent work via `phase()` and `agent()`.

---

## Workflow suitability

Every top-level workflow declares its backend capability in `<Workflow>/manifest.yaml` (the source of truth). The table below shows the **driver-path capability** — the backend(s) the workflow supports when invoked with `args.backend` + `args.backend_dir`. When invoked without a driver directory (legacy path), the `.js` `WORKFLOW_SUITABILITY` guard may accept a narrower subset; see each workflow's `manifest.yaml` `notes` field for legacy/driver differences. The check is deliberately conservative: workflows do not infer language or problem type from natural-language `problem_definition`.

| Workflow | Supported language/backend | Supported `problem_type` values | Good fit | Avoid when |
|----------|----------------------------|---------------------------------|----------|------------|
| [AscendC](AscendC/) | AscendC / Ascend (msprof + `ascendc_direct_launch`) | `ascend-kernel-optimization`, `ascend-kernel-generation` | Ascend 910B NPU kernel optimization/generation via msprof and the substrate ascend backend | CUDA/Triton/ROCm/Metal targets — use their own workflows |
| [AccelOpt](AccelOpt/) | CUDA (default) · Triton via driver (vendor-locked: ncu) | `cuda-kernel-optimization`, `cuda-kernel-generation` | Existing CUDA kernels, or CUDA seed generation followed by NCU/benchmark-driven optimization | Triton/SYCL/XPU tasks or missing benchmark/profile contract |
| [KEET](KEET/) | CUDA (vendor-locked: ncu) | `performance-explanation` | Explaining CUDA source plus Nsight Compute profile evidence | Kernel generation or optimization without profile artifacts |
| [ARGUS](ARGUS/) | ROCm/CUDA/Triton/ARGUS-DSL (legacy; driver pending) | `invariant-guided-kernel-optimization`, `gpu-kernel-optimization` | GPU kernel optimization with invariant checker/test feedback | Tasks without invariant or validation evidence |
| [AKO4X](AKO4X/) | Triton (default) · CUDA/CuTe/TileLang/C++/PyTorch via driver | `gpu-kernel-optimization`, `kernel-generation` | Multi-round benchmark-driven GPU kernel optimization across supported DSLs | Non-kernel application code or unsupported backend toolchains |
| [KDA](KDA/) | CUDA (default) · Triton via driver (experimental) · Ascend via substrate (faithful but simplified) | `cuda-kernel-optimization`, `cuda-kernel-generation`, `ascend-kernel-optimization`, `ascend-kernel-generation` | Evidence-driven CUDA implementation, validation, and optimization loops; Ascend routes through the substrate ascend backend | Backends without a substrate driver or validation contract |
| [K-Search](KSearch/) | Triton (default) · CUDA/Python via driver | `gpu-kernel-optimization`, `kernel-search` | World-model tree search with an evaluator/benchmark contract | Tasks without executable evaluator feedback |
| [AdaExplore](AdaExplore/) | CUDA · Triton via driver (experimental) | `triton-kernel-optimization`, `triton-kernel-generation` | PyTorch operator spec to Triton using MCTS and failure memory | Direct CUDA/CUTLASS/SYCL optimization |
| [KernelFoundry](KernelFoundry/) | CUDA (default) · Triton via driver (experimental) | `gpu-kernel-optimization`, `kernel-generation`, `kernel-search` | MAP-Elites quality-diversity search with descriptor/archive feedback | Single deterministic patch workflows without archive state |
| [CUDA Agent](CUDAAgent/) | CUDA (vendor-locked: ncu) | `cuda-kernel-generation`, `cuda-kernel-optimization` | PyTorch model/operator to custom CUDA ops and bindings | Triton/SYCL/CUTLASS-only tasks |
| [cuPilot](cuPilot/) | CUDA (vendor-locked: ncu) | `cuda-kernel-optimization` | Strategy-level CUDA evolution with roofline/profiler evidence | Non-CUDA kernels |
| [TritorX](TritorX/) | Triton (vendor-locked: linter) | `aten-triton-operator-generation`, `operator-generation` | ATen/Triton operator coverage generation and compile/lint/test loops | Performance-first CUDA tuning |
| [KernelBand](KernelBand/) | Triton (default) · CUDA via driver | `gpu-kernel-optimization`, `kernel-search` | Bandit-guided search using hardware signatures and profiling | Backends without comparable profiling/evaluator evidence |
| [KernelAgent](KernelAgent/) | Triton (default) · CUDA via driver (experimental) | `triton-kernel-generation`, `operator-generation` | Triton synthesis with PyTorch-style verification harnesses | CUDA/C++/CUTLASS kernels |
| [STARK](STARK/) | CUDA (default) · Triton via driver (experimental) | `cuda-kernel-optimization`, `kernel-search` | CUDA tree-search refinement with multi-agent planning/debugging | Non-CUDA backends until code-context adapters exist |
| [ReGraphT](ReGraphT/) | CUDA (default) · Triton via driver (experimental) | `cuda-kernel-optimization`, `kernel-search` | CUDA reasoning graph search and Monte Carlo graph search | Non-CUDA optimization traces |
| [Astra](Astra/) | CUDA (default) · Triton via driver (experimental) | `cuda-kernel-optimization` | Existing production CUDA/PyBind kernels with tests and profiling | From-scratch Triton/SYCL generation |
| [AutoMegaKernel](AutoMegaKernel/) | CUDA via AutoMegaKernel (method-intrinsic; external harness) | `amk-schedule-search`, `megakernel-synthesis`, `llama-megakernel-optimization` | Existing AMK checkout; Llama-family CUDA megakernel schedule search over `ScheduleConfig` + `kernel_knobs` | Ordinary standalone CUDA/Triton/CUTLASS kernels; missing AMK checkout; high-batch serving throughput |
| [Harness Engineering](HarnessEngineering/) | CUDA/Triton/CuTe via caller-owned external harness | `gpu-kernel-optimization`, `harness-constrained-kernel-optimization`, `flashinfer-kernel-optimization` | Existing kernel plus frozen compile/correctness/timing commands; profile-backed bounded edits with complete artifact retention | Missing executable harness contract; generation without a baseline; treating prose as verification |
| [Atrex](Atrex/) | CUDA/ROCm through the official Atrex orchestrator | `atrex-kernel-optimization`, `production-kernel-optimization`, `sol-execbench-operator-optimization` | Existing Atrex checkout and GPU gateway; SOL-ExecBench or native Atrex-Bench campaign | Missing official checkout/platform/command; ordinary project harnesses; bypassing Atrex's supervisor |
| [CUDA-LLM](CUDALLM/) | CUDA (default) · Triton via driver (experimental) | `cuda-kernel-generation`, `cuda-kernel-optimization` | CUDA feature search/reinforcement from task specs | Non-CUDA output languages |
| [CutlassGEMM](CutlassGEMM/) | CUTLASS / C++ (method-intrinsic) | `cutlass-gemm-optimization` | CUTLASS GEMM/SOL-ExecBench dispatch tuning | General elementwise/attention kernels outside CUTLASS GEMM |
| [GemmPTX](GemmPTX/) | CUDA/CuTe/CUTLASS (vendor-locked: PTX/SASS disassembly) | `gemm-ptx-optimization`, `cuda-gemm-ptx-optimization`, `gemm-instruction-optimization` | Existing GEMM kernels where instruction-path selection must be proven by PTX/SASS evidence | Generic compute-bound kernels; missing disassembly/benchmark/correctness contract |
| [FACT](FACT/) | CUTLASS / C++ (method-intrinsic) | `cutlass-pattern-synthesis`, `cutlass-gemm-optimization` | CUTLASS pattern discovery, realization, composition, and ablation | Standalone Triton/SYCL kernels |
| [GPU Forecasters](GPUForecasters/) | CUDA (vendor-locked: ncu) | `cuda-kernel-optimization`, `kernel-search` | CUDA/GPU search with speedup forecaster and execute/abstain feedback | Tasks without GPU execution or forecast calibration |
| [InPlacePatch](InPlacePatch/) | CUDA/ROCm (vendor-locked: nvcc/hipcc) | `embedded-kernel-optimization` | Byte-exact in-place kernel patches with project-native build/test/benchmark | Standalone kernels with exp_dir-based workflows |
| [KernelBlaster](KernelBlaster/) | CUDA (vendor-locked: ncu) | `cuda-kernel-optimization` | NCU elapsed-cycle CUDA optimization with persistent memory | Non-CUDA backends |
| [KernelFoundryDx](KernelFoundryDx/) | Triton (method-intrinsic) | `triton-kernel-optimization`, `triton-kernel-generation` | PyTorch reference to Triton multi-island evolution | CUDA/CUTLASS/SYCL kernels |
| [LlamacppEmbeddedSearch](LlamacppEmbeddedSearch/) | llama.cpp ggml-cuda (method-intrinsic) | `embedded-kernel-search` | Multi-variant fan-out search for kernels embedded in llama.cpp ggml-cuda | Standalone CUDA/Triton kernel optimization |
| [KernelSkill](KernelSkill/) | CUDA (default) · Triton via driver (experimental) | `cuda-kernel-optimization`, `cuda-kernel-generation` | CUDA custom kernels from PyTorch reference with memory/profiler guidance | Triton/SYCL/Metal tasks |
| [StitchCUDA](StitchCUDA/) | CUDA (default) · Triton via driver (experimental) | `cuda-kernel-generation`, `cuda-kernel-optimization` | Planner/Coder/Verifier CUDA synthesis and replanning | Non-CUDA backends |
| [WarpSpeed](WarpSpeed/) | CUDA (vendor-locked: requires NCU/compute-sanitizer) | `cuda-kernel-optimization`, `kernel-search` | Sustained campaigns on a dedicated multi-GPU node: checkpoint tree, GPU mutex, two-tier benchmarking, rewind | One-shot patches, shared GPUs, hosts without NCU/sanitizer |
| [Xe-Forge](Xe-Forge/) | Intel XPU (vendor-locked: xpu) | `xpu-kernel-optimization`, `triton-kernel-optimization` | Intel XPU CoVeR staged refinement | NVIDIA CUDA-only tuning |
| [Generalist](Generalist/) | CUDA/Triton/MetaX (default) · Ascend via substrate (faithful but simplified) | `cuda-kernel-generation`, `cuda-kernel-optimization`, `ascend-kernel-generation`, `ascend-kernel-optimization` | Backend-clean substrate solver (beam); the default single-solver baseline; Ascend via the substrate ascend backend | Tasks needing a non-substrate measurement signal |
| [Meta-Workflow](_meta/) | Tooling | N/A | Generating and validating workflow definitions | Direct kernel optimization |

---

## Catalog {#catalog}

| Method | Tags | Core loop | Paper / project |
|--------|------|-----------|-----------------|
| [AscendC](AscendC/) | ![AscendC](https://img.shields.io/badge/AscendC-0066CC?style=flat) ![msprof](https://img.shields.io/badge/msprof-555?style=flat) ![Iterative](https://img.shields.io/badge/iterative-blue?style=flat) | Setup → Generate → Evaluate → Optimize (msprof) → Report | Proven session-local variant ([AscendC/README.md](https://raw.githubusercontent.com/qhy991/Awesome-Kernel-Workflows/main/AscendC/README.md)); promoted upstream |
| [AccelOpt](AccelOpt/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![NCU](https://img.shields.io/badge/NCU-555?style=flat) ![Iterative](https://img.shields.io/badge/iterative-blue?style=flat) ![Experience](https://img.shields.io/badge/experience--memory-orange?style=flat) | Plan → Execute → Profile → Learn → Iterate | [arXiv:2511.15915](https://arxiv.org/abs/2511.15915) (MLSys 2026) |
| [KEET](KEET/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![NCU](https://img.shields.io/badge/NCU-555?style=flat) ![Pipeline](https://img.shields.io/badge/pipeline-purple?style=flat) ![Explanation](https://img.shields.io/badge/explanation-teal?style=flat) | Source Inspection → Profile Inspection → Aggregation → Review | [arXiv:2605.04467](https://arxiv.org/abs/2605.04467) (UMD/NVIDIA/LLNL 2026) |
| [ARGUS](ARGUS/) | ![ROCm](https://img.shields.io/badge/ROCm-ED1C24?style=flat&logo=amd&logoColor=white) ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![DSL](https://img.shields.io/badge/DSL-darkgreen?style=flat) ![ICRL](https://img.shields.io/badge/ICRL-blue?style=flat) ![Invariants](https://img.shields.io/badge/invariants-red?style=flat) | Plan → Select → Lower → Validate → Learn | [arXiv:2604.18616](https://arxiv.org/abs/2604.18616) (CausalFlow/HKUST/Stanford 2026) |
| [AKO4X](AKO4X/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![Triton](https://img.shields.io/badge/Triton-6C3483?style=flat) ![CuTe](https://img.shields.io/badge/CuTe-darkgreen?style=flat) ![TileLang](https://img.shields.io/badge/TileLang-darkgreen?style=flat) ![Iterative](https://img.shields.io/badge/iterative-blue?style=flat) ![Experience](https://img.shields.io/badge/experience--memory-orange?style=flat) | Round-Init → Iterate → Archive → Retrospect | [AKO Project](https://tongminglaic.github.io/AKO) |
| [KDA](KDA/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![Iterative](https://img.shields.io/badge/iterative-blue?style=flat) ![Evidence](https://img.shields.io/badge/evidence--driven-green?style=flat) | Inspect → Plan → Implement → Validate → Decide | [MIT HAN Lab](https://github.com/mit-han-lab/kernel-design-agents) |
| [K-Search](KSearch/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![Triton](https://img.shields.io/badge/Triton-6C3483?style=flat) ![Tree](https://img.shields.io/badge/tree--search-darkblue?style=flat) ![WorldModel](https://img.shields.io/badge/world--model-orange?style=flat) | Init Tree → Select → Generate → Evaluate → Refine/Backtrack | [arXiv:2602.19128](https://arxiv.org/abs/2602.19128) |
| [AdaExplore](AdaExplore/) | ![Triton](https://img.shields.io/badge/Triton-6C3483?style=flat) ![MCTS](https://img.shields.io/badge/MCTS-darkblue?style=flat) ![SkillMemory](https://img.shields.io/badge/skill--memory-orange?style=flat) | Select → Expand → Evaluate → Backpropagate | [arXiv:2604.16625](https://arxiv.org/abs/2604.16625) |
| [KernelFoundry](KernelFoundry/) | ![SYCL](https://img.shields.io/badge/SYCL-0071C5?style=flat&logo=intel&logoColor=white) ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![Triton](https://img.shields.io/badge/Triton-6C3483?style=flat) ![Evolutionary](https://img.shields.io/badge/MAP--Elites-darkblue?style=flat) ![MetaPrompt](https://img.shields.io/badge/meta--prompt-orange?style=flat) | Select → Vary → Evaluate → Insert → Evolve-Prompts | [arXiv:2603.12440](https://arxiv.org/abs/2603.12440) (Intel 2026) |
| [CUDA Agent](CUDAAgent/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![Iterative](https://img.shields.io/badge/iterative-blue?style=flat) ![RL-trained](https://img.shields.io/badge/RL--trained-red?style=flat) | Profile → Implement → Verify → Refine | [arXiv:2602.24286](https://arxiv.org/abs/2602.24286) (ByteDance/Tsinghua 2026) |
| [cuPilot](cuPilot/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![NCU](https://img.shields.io/badge/NCU-555?style=flat) ![Evolutionary](https://img.shields.io/badge/evolutionary-darkblue?style=flat) ![Roofline](https://img.shields.io/badge/roofline-teal?style=flat) ![RAG](https://img.shields.io/badge/RAG-orange?style=flat) | Strategize → Translate → Revise → Evolve | [arXiv:2512.16465](https://arxiv.org/abs/2512.16465) (SEU/Tsinghua 2025) |
| [TritorX](TritorX/) | ![ASIC](https://img.shields.io/badge/ASIC%2FNPU-333?style=flat) ![Triton](https://img.shields.io/badge/Triton-6C3483?style=flat) ![FSM](https://img.shields.io/badge/FSM-blue?style=flat) ![Linter](https://img.shields.io/badge/linter-green?style=flat) ![Coverage](https://img.shields.io/badge/coverage--first-teal?style=flat) | Generate → Lint → Compile/Test → Debug (loop) | [arXiv:2512.10977](https://arxiv.org/abs/2512.10977) (Meta 2025) |
| [KernelBand](KernelBand/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![Triton](https://img.shields.io/badge/Triton-6C3483?style=flat) ![NCU](https://img.shields.io/badge/NCU-555?style=flat) ![MAB](https://img.shields.io/badge/MAB--UCB-darkblue?style=flat) ![Clustering](https://img.shields.io/badge/clustering-orange?style=flat) ![HW-Pruning](https://img.shields.io/badge/HW--pruning-red?style=flat) | Profile → Cluster → Select(UCB) → Generate → Evaluate → Update | [arXiv:2511.18868](https://arxiv.org/abs/2511.18868) (PKU 2026) |
| [KernelAgent](KernelAgent/) | ![Triton](https://img.shields.io/badge/Triton-6C3483?style=flat) ![PyTorch](https://img.shields.io/badge/PyTorch-EE4C2C?style=flat&logo=pytorch&logoColor=white) ![Pipeline](https://img.shields.io/badge/pipeline-purple?style=flat) ![MultiAgent](https://img.shields.io/badge/multi--agent-teal?style=flat) ![Verification](https://img.shields.io/badge/verification-green?style=flat) | Route → Generate(parallel) → Verify → Refine → Compose | [PyTorch Blog](https://pytorch.org/blog/kernelfalcon-autonomous-gpu-kernel-generation-via-deep-agents/) (PyTorch Labs 2025) |
| [STARK](STARK/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![Tree](https://img.shields.io/badge/tree--search-darkblue?style=flat) ![MultiAgent](https://img.shields.io/badge/multi--agent-teal?style=flat) ![Grounded](https://img.shields.io/badge/grounded--instruction-green?style=flat) ![DynamicContext](https://img.shields.io/badge/dynamic--context-orange?style=flat) | Setup → Select(ε-greedy) → Plan/Code/Debug → Evaluate → Update | [arXiv:2510.16996](https://arxiv.org/abs/2510.16996) (Meta/Duke 2025) |
| [ReGraphT](ReGraphT/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![Tree](https://img.shields.io/badge/tree--search-darkblue?style=flat) ![RAG](https://img.shields.io/badge/RAG-orange?style=flat) ![ReasoningGraph](https://img.shields.io/badge/reasoning--graph-teal?style=flat) ![MCGS](https://img.shields.io/badge/MCGS-darkblue?style=flat) | BuildGraph → Select(MCGS) → Generate → Evaluate → UpdateGraph | [arXiv:2510.19873](https://arxiv.org/abs/2510.19873) (CAS/SCUT 2025) |
| [Astra](Astra/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![MultiAgent](https://img.shields.io/badge/multi--agent-teal?style=flat) ![Profiling](https://img.shields.io/badge/profiling-green?style=flat) ![SGLang](https://img.shields.io/badge/SGLang-orange?style=flat) | Setup → Test/Profile → Plan → Code → Evaluate → Record | [arXiv:2509.07506](https://arxiv.org/abs/2509.07506) (Stanford/SJTU/NJU 2025) |
| [AutoMegaKernel](AutoMegaKernel/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![Megakernel](https://img.shields.io/badge/megakernel-333?style=flat) ![ScheduleSearch](https://img.shields.io/badge/schedule--search-darkblue?style=flat) ![ExternalHarness](https://img.shields.io/badge/external--harness-teal?style=flat) ![Verification](https://img.shields.io/badge/verification-green?style=flat) | Doctor → ProposeSurface → EvalBaseline → Loop/Autoresearch → Audit | [arXiv:2606.09682](https://arxiv.org/abs/2606.09682) / [AMK](https://github.com/RightNow-AI/AutoMegaKernel) |
| [Harness Engineering](HarnessEngineering/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![Triton](https://img.shields.io/badge/Triton-6C3483?style=flat) ![ExternalHarness](https://img.shields.io/badge/external--harness-teal?style=flat) ![Verification](https://img.shields.io/badge/verification-green?style=flat) | FreezeContract → Baseline → Profile/Decide → Implement → OfficialGate → Audit | [arXiv:2607.17979](https://arxiv.org/abs/2607.17979) / [artifact](https://github.com/syhya/mlsys26-flashinfer-contest) |
| [Atrex](Atrex/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![ROCm](https://img.shields.io/badge/ROCm-ED1C24?style=flat&logo=amd&logoColor=white) ![Profiling](https://img.shields.io/badge/profiling-green?style=flat) ![ExternalHarness](https://img.shields.io/badge/external--harness-teal?style=flat) ![ABBA](https://img.shields.io/badge/ABBA--verification-darkblue?style=flat) | Doctor → LaunchOfficialCampaign → AuditCanonicalEvidence → Report | [arXiv:2607.14541](https://arxiv.org/abs/2607.14541) / [Atrex Kernel Agent](https://github.com/alibaba/atrex-kernel-agent) |
| [CUDA-LLM](CUDALLM/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![Iterative](https://img.shields.io/badge/iterative-blue?style=flat) ![FeatureSearch](https://img.shields.io/badge/feature--search-teal?style=flat) ![Reinforcement](https://img.shields.io/badge/reinforcement-red?style=flat) | Catalog → SelectFeatures → Generate → Evaluate → Reinforce | [arXiv:2506.09092](https://arxiv.org/abs/2506.09092) (2025) |
| [CutlassGEMM](CutlassGEMM/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![C++](https://img.shields.io/badge/C%2B%2B-00599C?style=flat&logo=c%2B%2B&logoColor=white) ![CUTLASS](https://img.shields.io/badge/CUTLASS-76B900?style=flat) ![NCU](https://img.shields.io/badge/NCU-555?style=flat) ![Iterative](https://img.shields.io/badge/iterative-blue?style=flat) ![Dispatch](https://img.shields.io/badge/multi--config--dispatch-teal?style=flat) | Analyze → GenerateConfigs → Profile(NCU) → TuneDispatch → Validate | [CUTLASS](https://github.com/NVIDIA/cutlass) / [SOL-ExecBench](https://github.com/NVIDIA/SOL-ExecBench) |
| [GemmPTX](GemmPTX/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![PTX](https://img.shields.io/badge/PTX%2FSASS-555?style=flat) ![GEMM](https://img.shields.io/badge/GEMM-333?style=flat) ![Evidence](https://img.shields.io/badge/disassembly--evidence-green?style=flat) ![Iterative](https://img.shields.io/badge/iterative-blue?style=flat) | HardwareCensus → GEMMSignature → InstructionPlan → DisassembleVerify → Profile → Decide | AKW engineering workflow |
| [FACT](FACT/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![C++](https://img.shields.io/badge/C%2B%2B-00599C?style=flat&logo=c%2B%2B&logoColor=white) ![CUTLASS](https://img.shields.io/badge/CUTLASS-76B900?style=flat) ![Pipeline](https://img.shields.io/badge/pipeline-purple?style=flat) ![Ablation](https://img.shields.io/badge/ablation-green?style=flat) | Discover → Realize → Compose → Ablate | [FACT Project](https://github.com/Project-FACT/FACT) |
| [GPU Forecasters](GPUForecasters/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![Tree](https://img.shields.io/badge/tree--search-darkblue?style=flat) ![PUCT](https://img.shields.io/badge/PUCT-darkblue?style=flat) ![Forecasting](https://img.shields.io/badge/speedup--forecaster-teal?style=flat) ![Abstain](https://img.shields.io/badge/abstain-orange?style=flat) | TrainForecaster → Select(PUCT) → Forecast/Execute → Update | [arXiv:2605.31464](https://arxiv.org/abs/2605.31464) |
| [KernelBlaster](KernelBlaster/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![NCU](https://img.shields.io/badge/NCU-555?style=flat) ![Iterative](https://img.shields.io/badge/iterative-blue?style=flat) ![Experience](https://img.shields.io/badge/experience--memory-orange?style=flat) ![ICRL](https://img.shields.io/badge/in--context--RL-red?style=flat) | Profile/Classify → Retrieve → Apply → Evaluate → Reward/Update | [arXiv:2602.14293](https://arxiv.org/abs/2602.14293) |
| [KernelFoundryDx](KernelFoundryDx/) | ![Triton](https://img.shields.io/badge/Triton-6C3483?style=flat) ![Evolutionary](https://img.shields.io/badge/evolutionary-darkblue?style=flat) ![MultiAgent](https://img.shields.io/badge/multi--agent-teal?style=flat) ![RAG](https://img.shields.io/badge/RAG-orange?style=flat) ![Diagnosis](https://img.shields.io/badge/diagnosis-teal?style=flat) | RAG-Seed → Evolve(Islands) → Evaluate → Diagnose → Migrate | [arXiv:2605.30359](https://arxiv.org/abs/2605.30359) (CUHK/Huawei 2026) |
| [KernelSkill](KernelSkill/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![NCU](https://img.shields.io/badge/NCU-555?style=flat) ![MultiAgent](https://img.shields.io/badge/multi--agent-teal?style=flat) ![SkillMemory](https://img.shields.io/badge/skill--memory-orange?style=flat) ![Verification](https://img.shields.io/badge/verification-green?style=flat) | Seed → Review → Repair/Optimize → Profile → UpdateMemory | [arXiv:2603.10085](https://arxiv.org/abs/2603.10085) |
| [StitchCUDA](StitchCUDA/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![MultiAgent](https://img.shields.io/badge/multi--agent-teal?style=flat) ![Pipeline](https://img.shields.io/badge/pipeline-purple?style=flat) ![Verification](https://img.shields.io/badge/verification-green?style=flat) ![Replanning](https://img.shields.io/badge/adaptive--replanning-orange?style=flat) | Plan → Code → Verify → Replan → Iterate | [arXiv:2603.02637](https://arxiv.org/abs/2603.02637) |
| [WarpSpeed](WarpSpeed/) | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![Tree](https://img.shields.io/badge/tree--search-darkblue?style=flat) ![NCU](https://img.shields.io/badge/NCU-555?style=flat) ![Experience](https://img.shields.io/badge/experience--memory-orange?style=flat) ![Rewind](https://img.shields.io/badge/rewind-red?style=flat) ![CrossReview](https://img.shields.io/badge/cross--model--review-green?style=flat) | Plan → Generate(∥) → Screen(A/B) → Confirm → Profile → Record → Postmortem/Rewind | [Design note](https://raw.githubusercontent.com/qhy991/Awesome-Kernel-Workflows/main/WarpSpeed/README.md) |
| [Xe-Forge](Xe-Forge/) | ![XPU](https://img.shields.io/badge/Intel--XPU-0071C5?style=flat&logo=intel&logoColor=white) ![Triton](https://img.shields.io/badge/Triton-6C3483?style=flat) ![Pipeline](https://img.shields.io/badge/pipeline-purple?style=flat) ![CoVeR](https://img.shields.io/badge/CoVeR-purple?style=flat) ![VTune](https://img.shields.io/badge/VTune-555?style=flat) | Stage → Generate → Verify → Refine → Promote | [Xe-Forge Project](https://github.com/intel/Xe-Forge) |
| [Meta-Workflow](_meta/) | ![Tooling](https://img.shields.io/badge/tooling-gray?style=flat) | Research → Model → Assemble → Generate → Validate | — |

### Tag Legend

| Category | Tags |
|----------|------|
| **Backend** | ![CUDA](https://img.shields.io/badge/CUDA-76B900?style=flat&logo=nvidia&logoColor=white) ![ROCm](https://img.shields.io/badge/ROCm-ED1C24?style=flat&logo=amd&logoColor=white) ![SYCL](https://img.shields.io/badge/SYCL-0071C5?style=flat&logo=intel&logoColor=white) ![XPU](https://img.shields.io/badge/Intel--XPU-0071C5?style=flat&logo=intel&logoColor=white) |
| **Kernel Language** | ![Triton](https://img.shields.io/badge/Triton-6C3483?style=flat) ![DSL](https://img.shields.io/badge/DSL-darkgreen?style=flat) ![CuTe](https://img.shields.io/badge/CuTe-darkgreen?style=flat) ![TileLang](https://img.shields.io/badge/TileLang-darkgreen?style=flat) ![CUTLASS](https://img.shields.io/badge/CUTLASS-76B900?style=flat) ![C++](https://img.shields.io/badge/C%2B%2B-00599C?style=flat&logo=c%2B%2B&logoColor=white) |
| **Search Strategy** | ![Iterative](https://img.shields.io/badge/iterative-blue?style=flat) ![Tree](https://img.shields.io/badge/tree--search-darkblue?style=flat) ![MCTS](https://img.shields.io/badge/MCTS-darkblue?style=flat) ![PUCT](https://img.shields.io/badge/PUCT-darkblue?style=flat) ![Evolutionary](https://img.shields.io/badge/MAP--Elites-darkblue?style=flat) ![MAB](https://img.shields.io/badge/MAB--UCB-darkblue?style=flat) ![Pipeline](https://img.shields.io/badge/pipeline-purple?style=flat) ![CoVeR](https://img.shields.io/badge/CoVeR-purple?style=flat) |
| **Profiler** | ![NCU](https://img.shields.io/badge/NCU-555?style=flat) |
| **Learning Mechanism** | ![Experience](https://img.shields.io/badge/experience--memory-orange?style=flat) ![SkillMemory](https://img.shields.io/badge/skill--memory-orange?style=flat) ![WorldModel](https://img.shields.io/badge/world--model-orange?style=flat) ![MetaPrompt](https://img.shields.io/badge/meta--prompt-orange?style=flat) ![Clustering](https://img.shields.io/badge/clustering-orange?style=flat) ![ICRL](https://img.shields.io/badge/ICRL-blue?style=flat) ![RL-trained](https://img.shields.io/badge/RL--trained-red?style=flat) |
| **Special** | ![Invariants](https://img.shields.io/badge/invariants-red?style=flat) ![Evidence](https://img.shields.io/badge/evidence--driven-green?style=flat) ![Explanation](https://img.shields.io/badge/explanation-teal?style=flat) ![HW-Pruning](https://img.shields.io/badge/HW--pruning-red?style=flat) |

### Methodology Classification Matrix

Use this matrix as the primary taxonomy when adding or reviewing workflows. Backend and language badges are useful, but the more important dimensions are **loop topology**, **authoritative feedback**, and **state carried across attempts**.

| Workflow | Primary category | Search topology | Authoritative feedback | Carried state / memory | Fidelity boundary |
|----------|------------------|-----------------|------------------------|------------------------|-------------------|
| [AscendC](AscendC/) | `iterative_self_improving` | plan→generate→evaluate→optimize (msprof) loop | msprof bottleneck class + eval_single_runner timing (ascendc_direct_launch) | candidate submissions, best path, stagnation counter | Substrate reference — the substrate ascend backend owns the measurement signal |
| [AccelOpt](AccelOpt/) | `iterative_self_improving` | Beam-style Plan/Execute/Profile/Learn loop | NCU metrics, latency, slow/fast pairs | `experienceMemory`, candidate beam | Faithful CUDA/NCU adaptation of a broader accelerator method |
| [KEET](KEET/) | `single_pass_pipeline` | Source/profile analysis pipeline | NCU profile plus source-grounded hypotheses | Hypothesis verdicts and explanation report | Explanation workflow, not an optimizer |
| [ARGUS](ARGUS/) | `iterative_self_improving` | ICRL Plan/Select/Lower/Validate/Learn loop | Invariant violations, tests, throughput | Planner policy, invariant violation log, candidate beam | Strict use requires executable invariant checker artifacts |
| [AKO4X](AKO4X/) | `iterative_self_improving` | Round-level loop with inner iterations | Smoke/full benchmark, optional NCU | Experience headers, `TRAPS.md`, archive | High-fidelity to AKO4X round/archive protocol |
| [KDA](KDA/) | `iterative_self_improving` | One-candidate-at-a-time evidence loop | Validation command and target metric | Draft/plan docs, candidate records | High-fidelity to KDA agent-flow contract |
| [K-Search](KSearch/) | `tree_exploration` | Co-evolving world-model tree with backtracking | Evaluator speedup and pass/fail result | Decision tree, solution DB, best solution | High-fidelity inference/search translation |
| [AdaExplore](AdaExplore/) | `tree_exploration` | MCTS with large/small expansion iterations | Compile/correctness/performance evaluator | MCTS statistics, diversity pool, skill memory | High-fidelity standalone method translation |
| [KernelFoundry](KernelFoundry/) | `search_based` | MAP-Elites quality-diversity evolution | Compile, correctness, benchmark, descriptor evidence | Elite archive, descriptors, meta-prompts | Needs deterministic archive/descriptor artifacts for strictness |
| [CUDA Agent](CUDAAgent/) | `iterative_self_improving` | Profile/Implement/Verify/Refine loop | Compile, correctness, speedup reward | Interaction history and skill-loop notes | Inference-time loop only; no RL/PPO training reproduction |
| [cuPilot](cuPilot/) | `search_based` | Strategy-level evolutionary loop | NVCC/function check, NCU, roofline evidence | Strategy pool, population, RAG corpus | Strictness depends on concrete roofline/RAG/NCU artifacts |
| [TritorX](TritorX/) | `multi_stage_refinement` | FSM over Generate/Lint/Compile/Test/Debug | Linter, compiler, OpInfo correctness tests | Failure summaries, operator coverage stats | Coverage-first FSM unless real target harness is present |
| [KernelBand](KernelBand/) | `search_based` | Clustered multi-armed bandit with masked UCB | NCU feature vectors, latency reward | Cluster assignments, bandit statistics | Strictness requires real feature-vector and mask artifacts |
| [KernelAgent](KernelAgent/) | `multi_stage_refinement` | Route, parallel seeds, verify, refine, compose | Sandboxed verification pass/fail | Candidate pool, refinement history | Faithful but simplified runtime/process model |
| [STARK](STARK/) | `tree_exploration` | Epsilon-greedy tree with Plan/Code/Debug roles | Compile, correctness, runtime | Tree memory, leaderboard, dynamic contexts | Faithful but runtime role separation is prompt-mediated |
| [ReGraphT](ReGraphT/) | `tree_exploration` | Monte Carlo Graph Search over reasoning graph | Evaluator JSON with speedup/correctness | CUDA reasoning graph, selected paths | Training-free inference/search phase only |
| [Astra](Astra/) | `multi_stage_refinement` | Multi-agent production-kernel optimization loop | Tests, profiling, speedup | Run log, reintegration notes, best result | Idea-preserving unless source repo/runtime is available |
| [AutoMegaKernel](AutoMegaKernel/) | `search_based` | Structured ScheduleConfig + kernel_knobs search delegated to AMK loop/autoresearch | AMK validate-before-launch, full-model correctness, latency/roofline verdict | Best config, rows/results.tsv, flywheel corpus | Strict adapter to external AMK harness; not standalone reimplementation |
| [Harness Engineering](HarnessEngineering/) | `iterative_self_improving` | Frozen contract, profile-backed single-hypothesis rounds, official gate | Caller-owned compile/correctness/verification/timing commands | Incumbent evidence, accepted/rejected strategy ids, raw artifacts | Workflow adaptation; contest infrastructure and hidden tests are external |
| [Atrex](Atrex/) | `iterative_self_improving` | Official Atrex Long Horizon campaign delegated to its sole orchestrator | Complete-workload terminal validation plus same-allocation ABBA | Canonical memory/vN, episode journals, worktrees, promoted version | Strict adapter; no duplicate Atrex supervisor or optimization loop |
| [CUDA-LLM](CUDALLM/) | `iterative_self_improving` | Feature Search and Reinforcement loop | Compile/correctness/latency reward | Feature catalog, feature scores, candidates | Workflow adaptation; no model training reproduction |
| [CutlassGEMM](CutlassGEMM/) | `iterative_self_improving` | NCU-guided multi-config CUTLASS dispatch tuning loop | SOL-ExecBench correctness/speedup, NCU metrics, MFU | Tile configs, dispatch thresholds, per-M performance regimes | Engineering workflow adaptation for CUTLASS GEMM tuning |
| [GemmPTX](GemmPTX/) | `iterative_self_improving` | One-candidate-at-a-time instruction hypothesis loop with PTX/SASS verification before profiling | compile/test/benchmark/disassemble commands, optional NCU/profile metrics | candidate history, verified instruction paths, hypothesis_not_realized dead ends | Original AKW workflow adaptation for GEMM instruction-path tuning |
| [FACT](FACT/) | `multi_stage_refinement` | Pattern discovery, realization, composition, and ablation | CUTLASS compile/correctness/performance and ablation speedup | Pattern registry, dependency graph, composed candidates | Faithful but simplified compositional synthesis workflow |
| [GPU Forecasters](GPUForecasters/) | `tree_exploration` | PUCT tree search guided by learned forecaster with abstain | GPU speedup evaluator plus forecast/abstain calibration | Forecaster model, search tree, GPU iterations ledger | Workflow adaptation; surrogate training/calibration must be concrete for strictness |
| [KernelBlaster](KernelBlaster/) | `iterative_self_improving` | MAIC-RL rollouts keyed by hardware performance state | NCU Elapsed Cycles, correctness, reward | Optimization database, trajectories, replay buffer | Faithful in-context RL adaptation with persistent memory |
| [KernelFoundryDx](KernelFoundryDx/) | `search_based` | Multi-island evolutionary Triton search with diagnosis hints | Compile/correctness/speedup plus anti-cheating checks | Island populations, elite archives, hint library | Faithful paper adaptation; no public runtime/source repo available |
| [KernelSkill](KernelSkill/) | `iterative_self_improving` | Seed/review plus repair-or-optimize refinement loop | Compiler/verifier/profiler, speedup, NCU/nsys evidence | Long-term skill library, optimize history, repair chain | Faithful decision-process adaptation; gate is prompt/workflow mediated |
| [StitchCUDA](StitchCUDA/) | `multi_stage_refinement` | Planner/Coder/Verifier with adaptive replanning | Compile, correctness, benchmark speedup | Plan history, failure counters, best candidate | Faithful but simplified three-agent orchestration |
| [WarpSpeed](WarpSpeed/) | `tree_exploration` | Pre-registered experiments over an assumption-tagged checkpoint tree; postmortem ablation then frontier rewind | Significance-gated A/B screen + locked-clock confirm, curated NCU, sanitizers, codex cross-review | SQLite tree + GPU-minutes ledger, append-only BitLessons, NCU cache, evidence branches | Original AKW engineering workflow (no paper); GPU discipline is mutex-enforced |
| [Xe-Forge](Xe-Forge/) | `multi_stage_refinement` | Hard-ordered 11-stage CoVeR loops | Intel Triton compile, correctness, speedup, optional VTune | Best-in-stage kernels, promotion history | Workflow adaptation for Intel XPU project pipeline |
| [Meta-Workflow](_meta/) | `tooling` | Research/Model/Assemble/Generate/Validate | Manifest schema and static/semantic checks | Templates, manifests, validation reports | Repository infrastructure, not a paper method |

> Industrial practices (e.g. tiled attention as in FlashInfer, CUTLASS tuning workflows) will be added when they are concrete enough to agentize. Open an issue or PR to nominate the next entry.

---

## Requirements

- [Claude Code](https://docs.anthropic.com/en/docs/claude-code) (Workflows-enabled build)
- Per-workflow backend/toolchain. CUDA workflows need NVIDIA GPU + CUDA; Triton/SYCL/XPU/C++ workflows need their corresponding runtime and compiler stack.
- **Recommended for CUDA profiling workflows:** NVIDIA Nsight Compute (`ncu`) — workflows such as AccelOpt are evidence-driven; they avoid guessing bottlenecks
- Per-workflow extras may include: standalone benchmark harness, compiler/build command, experiment output directory (see each file’s header comments)

Run inside a container or dedicated virtual environment when possible; agents may compile code or run `pip install`.

---

## Quick start

### 1. Clone

```bash
git clone https://github.com/qhy991/Awesome-Kernel-Workflows.git
```

### 2. Install as global Claude Code workflows (recommended)

Install once; available in **every project** via `~/.claude/workflows/`:

```bash
git clone https://github.com/qhy991/Awesome-Kernel-Workflows.git
cd Awesome-Kernel-Workflows
./scripts/install-global-workflows.sh
```

Restart Claude Code, browse with `/workflows`, or invoke by `meta.name` (e.g. `/ako4x-kernel-optimizer`).

### 3. Or install into a single kernel project only

Copy into the project’s `.claude/workflows/` directory:

```bash
mkdir -p /path/to/your-kernel-project/.claude/workflows
cp Awesome-Kernel-Workflows/AccelOpt/accelopt-kernel-optimization.js \
   /path/to/your-kernel-project/.claude/workflows/
```

### 4. Invoke from Claude Code

From your kernel project root, start Claude Code and call the workflow (parameters are documented in each `.js` file). AccelOpt example:

Problem-definition entry:

```javascript
Workflow({
 name: 'accelopt-kernel-optimization',
 args: {
 problem_definition: 'Implement y = gelu(x) for a contiguous fp32 tensor',
 language: 'cuda',
 problem_type: 'cuda-kernel-generation',
 target_gpu: 'H100',
 test_command: '<user-provided correctness command with {kernel_path}/{result_path}>',
 benchmark_command: '<user-provided benchmark command with {kernel_path}/{result_path}>',
 iterations: 5,
 seed_candidates: 3,
 exp_dir: '/tmp/kernel_workflow_exp',
 },
})
```

Existing-kernel entry:

```javascript
Workflow({
 name: 'accelopt-kernel-optimization',
 args: {
 kernel_path: '/path/to/kernel.cu',
 op_description: 'Quantized GEMM Q4_0 weight × FP32 activation',
 language: 'cuda',
 problem_type: 'cuda-kernel-optimization',
 target_gpu: 'H100',
 test_command: '<user-provided correctness command with {kernel_path}/{result_path}>',
 benchmark_command: '<user-provided benchmark command with {kernel_path}/{result_path}>',
 harness_path: '/path/to/harness.cu',
 harness_build_cmd: '<user-provided harness build command>',
 harness_run_args: '',
 kernel_name_regex: 'forward_kernel',
 ncu_binary: '<user-provided ncu binary path>',
 exp_dir: '/path/to/experiment/output',
 iterations: 3,
 breadth: 3,
 samples_per_plan: 2,
 },
})
```

**AccelOpt highlights** (see [`AccelOpt/accelopt-kernel-optimization.js`](AccelOpt/accelopt-kernel-optimization.js)):

- Implements **Plan → Execute → Profile → Summarize → Accumulate Experience → Repeat** from the paper
- Embeds **NCU** in Setup / Evaluate / Learn (`--set full`, source-level analysis, etc.) — profile first, then diagnose, then edit code
- Maintains `experienceMemory`: extracts reusable patterns from slow/fast kernel pairs and NCU metric diffs

Without NCU, pass `test_command` / `benchmark_command` for custom test and benchmark fallbacks.

---

## Related projects

This repo does **not** replace full optimization frameworks. It provides **pluggable Claude Code workflow snippets** that complement:

| Project | Relationship |
|---------|----------------|
| [AKO4ALL](https://github.com/TongmingLAIC/AKO4ALL) | General agentic kernel optimization (TASK.md + skills); complementary to workflows here |
| [AccelOpt](https://arxiv.org/abs/2511.15915) | Source paper for the first catalogued workflow |
| [SOL-ExecBench](https://github.com/NVIDIA/SOL-ExecBench) | Standardized kernel optimization benchmark suite |
| [flashinfer-bench](https://github.com/flashinfer-ai/flashinfer-bench) | Industrial operator-level benchmarking and scoring |

---

## Contributing

PRs and issues are welcome, especially:

- **New workflows** tied to a paper, open-source system, or reproducible industrial process — include the source link and a short note on why it fits agents
- **Improvements** to existing workflows: NCU parsing, error handling, defaults, comments
- **Comparisons** on the same kernel across workflows (speedup, iteration count) — `examples/` may be added later

Checklist for contributors:

1. Add `<MethodName>/` with the `*.js` workflow
2. Document **source paper/repo** and **required args** in the file header
3. Update the catalog table in **both** `README.md` and `README.zh-CN.md`
4. Run `scripts/count-workflows.sh` to refresh the header **workflows** badge (`badges/workflows.json`)

---

## License

Per-workflow files should state license terms if derived from upstream projects. The repository default license applies when `LICENSE` is present at the repo root.

---

If this collection helps you operationalize kernel-optimization agents, consider starring the repo and nominating the next method to port.

Awesome-Kernel-Workflows

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Awesome-Kernel-Workflows

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Download README