skills.rest
igamenovoer
igamenovoerCOMMUNITY
Explore all skills by this author →
GitHub

krnopt-low-precision-kernel-formats

Compare low-precision CUDA kernel formats and contracts for engineers.

Updated May 27, 2026
Search Tags:#comparison#cuda#fp8#low-precision#kernel-format#dequant
One-click install
npx skills add https://github.com/igamenovoer/domain-skills --skill krnopt-low-precision-kernel-formats
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: krnopt-low-precision-kernel-formats
Source: https://github.com/igamenovoer/domain-skills/tree/main/domain/cuda/krnopt-low-precision-kernel-formats
Command: npx skills add https://github.com/igamenovoer/domain-skills --skill krnopt-low-precision-kernel-formats

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This skill helps engineers and researchers understand, compare, and select appropriate low-precision CUDA kernel formats and contracts, ensuring correct interpretation of value formats, scales, and layouts.

Core Features & Use Cases

  • Identify and distinguish FP8 variants, block-scaled FP8, NVFP4, MXFP4/6/8, and other low-precision contracts.
  • Determine where dequant, requant, or scale handling belongs in a kernel path and how it affects the ABI.
  • Use cases include evaluating format compatibility for profiling, optimization routing, and hardware-aware decisions in CUDA kernels.

Quick Start

Follow the guidelines to explain a given low-precision kernel format contract and propose the next appropriate skill to route to.

Frequently Asked Questions about krnopt-low-precision-kernel-formats

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I compare low-precision CUDA kernel formats like FP8 and NVFP4?▼

To compare low-precision CUDA kernel formats, evaluate the value dtype, scale dtype, granularity, layout contracts, and dequant placement to distinguish FP8 variants, NVFP4, and MXFP formats for hardware-aware optimization.

What is the difference between block scaling and standard FP8 in CUDA kernels?▼

Block scaling in CUDA kernels groups multiple values to share a single scale factor, whereas standard FP8 applies per-tensor scaling, affecting granularity, scale layout, and dequant location within the kernel path.

Where does dequantization belong in a low-precision CUDA kernel path?▼

Dequantization placement in a low-precision CUDA kernel path depends on the specific format contract, determining whether scale handling occurs during memory loads, compute stages, or output writes to maintain ABI compatibility.

When should I use MXFP4 or MXFP6 formats instead of FP8 for CUDA optimization?▼

Use MXFP4 or MXFP6 formats instead of FP8 when your CUDA optimization requires lower precision compute with block-scaled granularity, trading numerical range for improved memory bandwidth and throughput.

Does block-scaled FP8 affect the ABI of my CUDA kernel?▼

Block-scaled FP8 affects the CUDA kernel ABI by introducing specific layout contracts for scale tensors, requiring precise dequant and requant location handling to ensure correct value interpretation.

What are the architecture-specific constraints for NVFP4 format in CUDA?▼

NVFP4 format in CUDA imposes architecture-specific constraints on scale dtype and granularity, requiring engineers to verify hardware support and dequant placement to ensure correct low-precision kernel execution.

Related Skills

Explore More Software Engineering Skills→
igamenovoerigamenovoer
7

pixi-make-cu-build-env

Configure CUDA build environments for Pixi projects with CMake and nvcc.

Community
Advanced
Ced3-hanCed3-han

awq-quantization

Automate 4-bit activation-aware quantization for large language models.

Community
Advanced
alibabaalibaba
71

gpu-kernel-bottleneck-analysis

Analyzes GPU kernel performance using Roofline models and profiling data.

Official
Advanced
hung-phanhung-phan
1

triton-lang

Create high-performance GPU kernels for deep learning using Python.

Community
Advanced