注意: 本文正在更新中，请随时关注最新版本。

我最近参加了 FlashInfer AI Kernel Generation Contest（FlashInfer Contest, 2026）。这篇博客并不是一篇关于 CUDA kernel 优化技巧的教程，我本身并不是 GPU 算子开发专家；参加这次比赛的主要目的，是想借助一个高度可验证、反馈明确的任务环境，研究 如何让 Coding Agent 在持续闭环中产出高质量的 GPU kernel。完整的技术报告参见 Harness Engineering for LLM-Driven GPU Kernel Generation（Shui et al., 2026），公开仓库在 mlsys26-flashinfer-contest。

Harness Engineering

Self-Evolving Agents 中提到 Harness Engineering（OpenAI, 2026；Anthropic, 2026）的核心范式：人类负责设计约束、构建反馈机制并定义评估标准，Agent 在受控环境内迭代生成更高质量的代码。

Fig. 1. Closed-loop harness/controller workflow. Harness compiles, validates, benchmarks, profiles, and archives candidates; controller converts workload/profile/history into the next optimization hypothesis. (Image source: Shui, 2026)

Fig. 1. Closed-loop harness/controller workflow. Harness compiles, validates, benchmarks, profiles, and archives candidates; controller converts workload/profile/history into the next optimization hypothesis. (Image source: Shui, 2026)

整个 harness 由四层组成：

  • Grounding inputs：算子定义、参考实现和 workload JSON 等必要上下文输入。
  • Shape discovery：从 workload 中根据 batch size 和 sequence length 等参数进行分组，并从每组抽取代表性维度，使 agent 可以快速评估和迭代候选，而不需要每次都做全量验证。
  • Closed-loop optimization：在 baseline → profile → diagnose → generate → evaluate → archive 这条循环里生成候选，验证代码能否编译、结果是否正确，并评估性能；同时利用 Torch Profiler 和 NVIDIA Nsight Compute (NCU) 等工具分析算子瓶颈。
  • Outputs：归档相关代码和性能指标等文件，后续提供给 agent 继续迭代。

人类编写优化 skills、构建评测脚手架。这些做法类似 Codex Skills（OpenAI, 2026）和 Codex Subagents（OpenAI, 2026）的思路：通过复用上下文、工具和并行探索来提升 agent 的搜索效率，避免陷入局部最优。

实验结果

下面的数据来自 Modal 平台 B200 GPU 上的本地评测，并不是公开 leaderboard 结果；评测方法参考 FlashInfer（Ye et al., 2025）和 FlashInfer-Bench（Xing et al., 2026）的 correctness-gated kernel benchmark 设定。下表为了便于阅读使用平均延迟之比作为汇总口径，并不是官方逐算子或逐赛道 contest score：

算子定义方法平均延迟 (ms)相对 PyTorch 加速相对 FlashInfer 加速
DSA AttentionAgent-Assisted0.011175217.17×29.68×
Full-Agent0.022811106.39×14.54×
FlashInfer baseline0.3316507.32×1.00×
DSA IndexerAgent-Assisted0.006893494.13×18.05×
Full-Agent0.032659104.29×3.81×
FlashInfer baseline0.12442027.38×1.00×
GDN PrefillAgent-Assisted0.05199221,078×13.70×
Full-Agent0.6888751,591×1.03×
FlashInfer baseline0.7121661,539×1.00×
MoE FP8Agent-Assisted0.28634063.78×1.62×
FlashInfer baseline0.46387439.37×1.00×
Full-Agent1.74263010.48×0.27×
GDN DecodeAgent-Assisted0.0062017,970×1.12×
FlashInfer baseline0.0069407,121×1.00×
Full-Agent0.0083665,907×0.83×
Fig. 2. Final retained speedups over the supplied FlashInfer baseline, measured using mean latency. Agent-Assisted artifacts improve all five definitions under the reporting normalization. (Image source: Shui, 2026)

Fig. 2. Final retained speedups over the supplied FlashInfer baseline, measured using mean latency. Agent-Assisted artifacts improve all five definitions under the reporting normalization. (Image source: Shui, 2026)

在这些本地评估的结果中，Agent-Assisted 在五个算子上全部优于 FlashInfer baseline，加速比从 1.12×（GDN Decode）到 29.68×（DSA Attention）；Full-Agent 则在 MoE FP8 与 GDN Decode 上反而比 baseline 更慢。

Agent-Assisted 优化轨迹

Fig. 3. Retained speedup trajectories over the supplied FlashInfer baseline. The curves show long plateaus and discrete jumps rather than smooth monotonic progress. (Image source: Shui, 2026)

Fig. 3. Retained speedup trajectories over the supplied FlashInfer baseline. The curves show long plateaus and discrete jumps rather than smooth monotonic progress. (Image source: Shui, 2026)

从轨迹可以看到，性能提升并不是平滑发生的，而是长期平台期之后出现少数几次明显跃迁。有效的 Agent-Assisted kernel 优化不是单纯依赖提示词，而是依赖可测量的系统闭环：把算子约束、评测脚手架、性能分析反馈和历史轨迹组织成可复用流程，再让智能体在其中生成、验证和保留候选。这个过程需要人类持续设计和维护 harness。

Full-Agent 优化轨迹

Fig. 4. Full-Agent optimization trajectories from LoongFlow trace logs. Gray dots are correctness-passing candidates, solid lines are running best, and dashed lines mark the FlashInfer baseline. (Image source: Shui, 2026)

Fig. 4. Full-Agent optimization trajectories from LoongFlow trace logs. Gray dots are correctness-passing candidates, solid lines are running best, and dashed lines mark the FlashInfer baseline. (Image source: Shui, 2026)

Full-Agent 轨迹来自 LoongFlow（Wan et al., 2025）的自动搜索日志。它也能在部分算子上找到有效候选，例如 DSA Attention 达到 14.54×，但整体仍慢于 Agent-Assisted，并且 MoE FP8 和 GDN Decode 甚至低于 FlashInfer baseline。这个差距说明，完全自动化的智能体搜索仍然困难。人类提供的高质量参考实现和持续积累的轨迹记忆，往往比让智能体从零开始探索更有效率。未来系统需要把控制状态和历史记忆纳入 harness，同时保持严格的最终验证。

参考文献

[1] FlashInfer Contest. “FlashInfer AI Kernel Generation Contest.” MLSys 2026 Competition, NVIDIA Track (2026).

[2] Shui, Yue, Chenyu Ma, Hangfei Xu, Shengzhao Wen, and Yanpeng Wang. “Harness Engineering for LLM-Driven GPU Kernel Generation.” Technical Report (2026).

[3] OpenAI. “Harness Engineering: Leveraging Codex in an Agent-First World.” OpenAI Blog (2026).

[4] Anthropic. “Harness Design for Long-Running Application Development.” Anthropic Engineering Blog (2026).

[5] OpenAI. “Codex Skills.” OpenAI Developers (2026).

[6] OpenAI. “Codex Subagents.” OpenAI Developers (2026).

[7] Ye, Zihao, et al. “FlashInfer: Efficient and Customizable Attention Engine for LLM Inference Serving.” Proceedings of Machine Learning and Systems (2025).

[8] Xing, Shanli, et al. “FlashInfer-Bench: Building the Virtuous Cycle for AI-driven LLM Systems.” arXiv preprint arXiv:2601.00227 (2026).

[9] Wan, Chunhui, et al. “LoongFlow: Directed Evolutionary Search via a Cognitive Plan-Execute-Summarize Paradigm.” arXiv preprint arXiv:2512.24077 (2025).

引用

引用：转载或引用本文内容时，请注明原作者和来源。

Cited as:

Yue Shui. (May 2026). 代码智能体生成 GPU Kernel.
https://syhya.github.io/zh/posts/2026-05-18-flashinfer-contest

Or

@article{syhya2026-mlsys26-flashinfer-contest,
  title   = "代码智能体生成 GPU Kernel",
  author  = "Yue Shui",
  journal = "syhya.github.io",
  year    = "2026",
  month   = "May",
  url     = "https://syhya.github.io/zh/posts/2026-05-18-flashinfer-contest"
}