Skip to main content

forward-kernel-validation

Validate optimized Recirculation forward kernels against an unfused, unbatched oracle when changing batching, fusion, attention, CUDA graphs, or accelerator inference code.

来源信息

仓库
ModelCloud/Recirculation
最近来源活动
2026年8月22日 12:54
检测到的 SKILL.md 语言
英语
星标
9
分支
4

安装方式

默认使用会先检查来源的 Prompt;你也可以切换为直接命令,或下载本地副本。

检查来源文件

决定是否安装前,请先阅读 SKILL.md,以及 SkillsMP 当前展示的配套文件。

正在显示 SKILL.md

SKILL.md
来源说明 · 只读预览
name
forward-kernel-validation
description
Validate optimized Recirculation forward kernels against an unfused, unbatched oracle when changing batching, fusion, attention, CUDA graphs, or accelerator inference code.
# Forward kernel validation Compare optimized forward outputs with the unfused, unbatched implementation using identical weights, dtype, tokens, positions, masks, cache state, and Recirculation configuration. Use mean absolute forward error as the release gate: - Fused or batched kernels: mean absolute error must be at most `4e-3`. - Kernels that are both non-fused and non-batched: mean absolute error must be at most `2e-3`. A path qualifies for the `4e-3` limit if it uses fusion, batching, or both. Record max absolute error, relative L2, and normalized max for diagnostics, but do not use those metrics to accept or reject the kernel. A stricter explicitly requested gate may still be used. Benchmark accuracy, task scores, and greedy-token agreement are outcome diagnostics, not substitutes for the forward gate. Do not reject an otherwise passing optimization solely because a benchmark score or greedy output changes; report those changes separately. Save the selected gate, kernel classification, oracle configuration, and all error metrics with benchmark artifacts.
在 GitHub 查看