Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add tensormux/kernel-skills --skill port-cuda-kernel-to-hipgit clone --depth 1 https://github.com/tensormux/kernel-skillsWrote 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.
[](https://agentmods.dev/skills/tensormux/kernel-skills/port-cuda-kernel-to-hip)<a href="https://agentmods.dev/skills/tensormux/kernel-skills/port-cuda-kernel-to-hip"><img src="https://agentmods.dev/badge/skills/tensormux/kernel-skills/port-cuda-kernel-to-hip/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.
<a href="https://agentmods.dev/skills/tensormux/kernel-skills/port-cuda-kernel-to-hip"><img src="https://agentmods.dev/badge/skills/tensormux/kernel-skills/port-cuda-kernel-to-hip.svg" alt="Reviewed on agentmods" width="80" height="20"></a>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 | $0.03305 |
| Opus 5 | $0.00000 | $0.01653 |
| Sonnet 5 | $0.00000 | $0.00661 |
| Haiku 4.5 | $0.00000 | $0.00331 |
Grade B, and why
port-cuda-kernel-to-hip scanned grade B with 1 finding 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 12d ago.
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.
Strips warnings and disclaimersmediumAnti-refusal
Omitting safety caveats hides risk from the user and is a common jailbreak preamble.
- [ ] Has the kernel been compiled with `hipcc --offload-arch=<target>` without warnings? How it starts
The opening of the file, as written. The whole thing — 137 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Skill: Port a CUDA Kernel to HIP
Purpose
Guide the agent through translating a CUDA kernel to AMD HIP for ROCm-compatible hardware (MI250, MI300, RDNA), preserving correctness and performance intent while adapting to the HIP execution model, memory model, and AMD-specific toolchain.
Use this when
- An existing, correct CUDA kernel needs to run on AMD GPUs (MI250, MI300X, RDNA3) via the ROCm stack.
- Building a codebase that needs to support both NVIDIA and AMD hardware from a single source using HIP's compatibility layer.
- Using
hipify-perlorhipify-clangautomated conversion and needing to audit and fix what the tool missed or got wrong. - The target is MI300X and performance parity with the CUDA original is required (not just functional correctness).
Do not use this when
- The kernel uses CUDA features with no HIP equivalent and cannot be rewritten (e.g., NVLink-specific topology, NVIDIA Tensor Core PTX instructions for Hopper). In this case, write a backend-agnostic plan first.
- The kernel is trivially expressed as a higher-level library call (rocBLAS, MIOpen, rocFFT) that already targets AMD platforms. Use the library rather than porting a custom kernel.
- The goal is Triton instead of HIP — Triton's AMD backend (ROCm) is often a better path for compute-intensive kernels than manual HIP porting.
Inputs the agent should gather first
- AMD hardware target: CDNA (MI250X, MI300X) for datacenter inference/training, or RDNA (RX 7900, etc.) for consumer. Wavefront size differs: CDNA and RDNA3 are 64-wide wavefronts by default, not 32-wide warps like CUDA.
- ROCm version: ROCm 5.x vs 6.x — HIP API surface and library availability differs across versions.
- CUDA features used: identify which features require attention — warp primitives (
__shfl_*,__ballot_sync,__any_sync,__all_sync), Tensor Core intrinsics (WMMA/WGMMA), cooperative groups, dynamic parallelism, unified memory, NVTX, NCCL. - Performance requirements: is functional correctness sufficient (pass CI), or must performance match or exceed the CUDA original on comparable hardware?
- Single-source or fork: will the code maintain one source base with
#ifdef __HIP_PLATFORM_AMD__guards, or will there be a separate HIP file?
What ships with it
1 file 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.
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.
- 12d ago First seen · 137 lines · 0 tokens per session scan B f55f922f50b9
port-cuda-kernel-to-hip is a skill published in the GitHub repository tensormux/kernel-skills (75 stars, last pushed 2mo ago), licensed MIT. It costs nothing until one of its globs matches a file; then it loads 3,305 tokens. A static security scan graded it B with 1 finding (strips warnings and disclaimers). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
cuopt-developer
Modify, build, test, debug, and contribute to NVIDIA cuOpt (C++/CUDA, Python, server, CI). Use for solver internals, PRs, DCO, and code conventions.
readable-cpp
Readable C/C++/Rust/CUDA code rules inspired by The Art of Readable Code. Use when writing, reviewing, or refactoring C, C++, Rust, or CUDA code. Enforces short functions, flat control flow, clear naming, readable structure, and idiomatic patterns.
cutedsl_megamoe
Skill "cutedsl_megamoe" from flashinfer-ai/flashinfer, covering updating the cutedsl megamoe kernel src, layout, when the kernel team drops a new version of src/ and what not to update here.
swapab_cutedsl_megakernel
Skill "swapab_cutedsl_megakernel" from flashinfer-ai/flashinfer, covering updating the sm120 swap-ab cutedsl megamoe kernel src, layout, when the kernel team drops a new version of src/ and what not to update here.
cpp
Comprehensive C/C++ programming reference covering everything from C11-C23 and C++11-C++23, system programming, CUDA GPU computing, debugging tools, Rust interop, and advanced topics. Use for: C/C++ questions, C/C++ interview preparation, modern language features, RAII/memory management, templates/generics, CUDA…
hip-kernel-optimization
This skill should be used when writing or tuning HIP kernels on AMD/NVIDIA GPUs, covering memory coalescing, shared-memory tiling, bank conflict avoidance, warp primitives, occupancy, vectorization, async ops, loop unrolling, and profiling.