What is next_cutedsl_megamoe?

A maintenance guide for importing or adapting a kernel-team release of the Rubin CuTeDSL MegaMoE code used by FlashInfer. A kernel is low-level code tuned to run computations on a processor or accelerator.

Use it when bringing in a new upstream drop, updating exports, checking compiler and target support, adapting dependencies, or tuning and measuring the kernel.

It records which upstream source is used and separates unchanged vendor code from FlashInfer-specific adaptations. This makes updates and tuning work easier to trace.

next_cutedsl_megamoe — Skill by flashinfer-ai/flashinfer

About the project

FlashInfer is a library and kernel generator that supplies GPU operations used to run large language model inference, including attention, matrix multiplication, and mixture-of-experts computations. It helps engineers build and optimize LLM serving systems across supported GPU hardware and backend implementations. Its catalogue add-ons provide skills and instructions for working with FlashInfer.

flashinfer-ai/flashinfer · 6,626 stars · on GitHub · flashinfer.ai

Install

Getting it into your agent

It runs from inside its repository, so the clone comes first — what it calls does not travel with the file alone.

Clone the repo

git clone --depth 1 https://github.com/flashinfer-ai/flashinfer

agentmods

npx agentmods add skills/flashinfer-ai/flashinfer/next_cutedsl_megamoe

Made for: Claude Code, Codex.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

README.md

[![agentmods](https://agentmods.dev/badge/skills/flashinfer-ai/flashinfer/next_cutedsl_megamoe/github.svg)](https://agentmods.dev/skills/flashinfer-ai/flashinfer/next_cutedsl_megamoe)

Your own site

<a href="https://agentmods.dev/skills/flashinfer-ai/flashinfer/next_cutedsl_megamoe"><img src="https://agentmods.dev/badge/skills/flashinfer-ai/flashinfer/next_cutedsl_megamoe/github.svg" alt="Measured on agentmods" height="20"></a>

Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.

Your own site · 80×15

<a href="https://agentmods.dev/skills/flashinfer-ai/flashinfer/next_cutedsl_megamoe"><img src="https://agentmods.dev/badge/skills/flashinfer-ai/flashinfer/next_cutedsl_megamoe.svg" alt="Reviewed on agentmods" width="80" height="20"></a>

Per session 0 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.

When invoked 3,557 The whole file, excluding the scripts and references it only reads on demand.

Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe.

Origin original No closer match found in the catalogue.

Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

Model	Per session	Once invoked
Fable 5.1	$0.00000	$0.03557
Opus 5.5	$0.00000	$0.01423
Sonnet 5.5	$0.00000	$0.00711
Haiku 4.5	$0.00000	$0.00356

Measured yesterday against content hash 27a7c055a756, method: parsed. Prices are Anthropic first-party input rates as of 2026-10-07, from the pricing page.

Security

Grade A, and why

next_cutedsl_megamoe scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.

The scan reads SKILL.md. This mod also ships 51 executable files (__init__.py, shim/__init__.py, shim/_paths.py, …), listed below but not scanned — reading those needs a real analyzer, not pattern matching.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

flashinfer/moe_ep/kernel_src/sm107/next_cutedsl_megamoe/SKILL.md · 292 lines

How it starts

The opening of the file, as written. The whole thing — 292 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Updating the Rubin CuTeDSL MegaMoE kernel src

Use this guide when importing a new kernel-team drop or adapting FlashInfer to changes in that drop. VENDOR.md records the pinned source and export exceptions; TUNING.md covers tuning and measurements.

Layout

kernel_src/sm107/next_cutedsl_megamoe/
├── src/sources/            ← immutable output of upstream next/export_src.py
│   ├── api.py, quant_def.py
│   ├── helpers/, communication/
│   └── kernel_src/         ← Rubin inference/mega plus shared dependencies
├── __init__.py             ← public package API used by FlashInfer backends
├── shim/                   ← FlashInfer-owned adaptation
│   ├── _paths.py           ← exposes src/ as the top-level sources package
│   ├── dependencies.py     ← native SM107 compiler/target checks
│   ├── block_scaled.py     ← config, descriptors, workspace, compile/launch
│   ├── comm.py             ← symmetric allocation, peer mapping, lifetime
│   ├── kernel_helpers.py   ← weight conversion, quantization, Torch reference
│   ├── knob_cache.py       ← heuristic profiles, cache keys and policy
│   ├── autotune.py         ← collective candidate validation and timing
│   └── correctness.py      ← sampled numerical oracle
├── VENDOR.md               ← provenance, supported subset, recorded exceptions
├── SKILL.md                ← this update procedure
└── TUNING.md               ← tuning and benchmark conventions

Keep src/ byte-for-byte equal to the pinned upstream export, except for the explicit local patches inventoried in VENDOR.md. Reapply and requalify those patches after an export until upstream absorbs them. Record the exporter revision, selected kernels and generated transformations in VENDOR.md. Fix device/exporter code upstream and re-export; put FlashInfer adaptation in shim/ and the backend wrappers. Preserve the vendored-path exclusions in formatting/lint tools.

Import layering

Backends → package __init__.py → shim/ → raw sources modules.

Read the full file on GitHub · 292 lines

Files

What ships with it

53 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

yesterday Changed · +83 lines 27a7c055a756
11d ago Changed · +2 lines e5ee42cc6298
13d ago Changed · +1 lines 0bd99db84407
22d ago First seen · 206 lines · 0 tokens per session scan A ce495f03afea

Subscribe to this mod's changes

next_cutedsl_megamoe is a skill published in the GitHub repository flashinfer-ai/flashinfer (6,626 stars, last pushed today), licensed Apache-2.0. It costs nothing until one of its globs matches a file; then it loads 3,557 tokens. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-24.

Data and AI GPU PyTorch cuda jit large-large-models llm-inference

gpu-server-setup

Prepare a Linux server with NVIDIA GPUs for neural-network and LLM workloads, or diagnose one that misbehaves. Use when the user asks to set up or prepare a GPU server, install NVIDIA drivers or CUDA on Debian/Ubuntu, wire Docker to GPUs (NVIDIA Container Toolkit), deploy vLLM / Infinity / OpenWebUI / Ollama /…

EvilFreelancer/gpu-server-setup · 167 tokens

sglang-diffusion-performance

Use when choosing the fastest SGLang Diffusion flags for a model, GPU, and VRAM budget.

sgl-project/sglang · 29 tokens

sglang-diffusion-cache-dit

Workflow for upgrading/integrating cache-dit in SGLang diffusion (multimodalgen): DBCache, DMD calibrator, TaylorSeer, SVDQuant DQ; porting upstream PRs and resolving conflicts against the per-request knob system; adding new cache knobs; building the sglang generate CLI test matrix; precision validation (PSNR / log…

sgl-project/sglang · 161 tokens

tensorboard

Visualize training metrics, debug models with histograms, compare experiments, visualize model graphs, and profile performance with TensorBoard - Google's ML visualization toolkit.

davila7/claude-code-templates · 32 tokens

ray-train

Distributed training orchestration across clusters. Scales PyTorch/TensorFlow/HuggingFace from laptop to 1000s of nodes. Built-in hyperparameter tuning with Ray Tune, fault tolerance, elastic scaling. Use when training massive models across multiple machines or running distributed hyperparameter sweeps.

davila7/claude-code-templates · 63 tokens

skypilot-multi-cloud-orchestration

Multi-cloud orchestration for ML workloads with automatic cost optimization. Use when you need to run training or batch jobs across multiple clouds, leverage spot instances with auto-recovery, or optimize GPU costs across providers.

davila7/claude-code-templates · 51 tokens