- 運作狀態
- 上線中
- 30 天上線率 100%
- 最近程式碼更新
- 10月19日
- 活躍儲存庫
- 最新版本
- 無
- 貢獻者
- 2 人
- 分支 10
- nomorevibe 收錄
- 9月10日
- 儲存庫建立於 2026年4月
語言組成
- Python 71.7%
- Shell 19.5%
- TypeScript 6%
- CSS 等 2.8%
更新0 筆更新
全部尚未偵測到更新。GitHub 發布版本後會顯示在這裡。
資料 分類的熱門專案
依最近已完成的 8 小時觀測回合間的 GitHub 星號增量排序
尚未偵測到更新。GitHub 發布版本後會顯示在這裡。
依最近已完成的 8 小時觀測回合間的 GitHub 星號增量排序
個人 GitHub 帳號下公開儲存庫的擁有者可以登入並驗證所有權，自行管理介紹，無需管理員核准。
登入並產生驗證檔案，在 30 分鐘內將其提交至此儲存庫的預設分支。其他分支或分叉儲存庫中的檔案無效。然後在下方檢查檔案，即可無需管理員核准直接管理專案。請保留檔案以維持管理權限。
如需登入，將前往登入頁面。登入後請回到此處，再次點擊所選的驗證按鈕。
在專案資料夾中執行 /nomorevibe verify，即可驗證所有權並自行管理介紹。
一分鐘完成驗證kernelbench.com — GPU kernel engineering benchmarks for autonomous LLM coding agents. v3 archive + v-hard latest.
翻譯準備中
取自儲存庫 README
正在載入完整 README。 以下為已儲存的摘錄。
Frontier coding agents write GPU kernels. Each session is one autonomous
agent, graded against a roofline (or ms/speedup) ceiling, then reward-hack
audited before anything is published. Live: kernelbench.com.
This monorepo is the website and the evals. GPU sessions launch to
Lambda / Brev (or another remote worker). Operator rules: AGENTS.md, which
points at a specialized AGENTS.md in each directory. Methodology and
history: each bench's SPEC.md and DEVLOG.md.檢視完整 README ↗Frontier coding agents write GPU kernels. Each session is one autonomous agent, graded against a roofline (or ms/speedup) ceiling, then reward-hack audited before anything is published. Live: kernelbench.com.
This monorepo is the website and the evals. GPU sessions launch to
Lambda / Brev (or another remote worker). Operator rules: AGENTS.md, which
points at a specialized AGENTS.md in each directory. Methodology and
history: each bench's SPEC.md and DEVLOG.md.
| bench | path | what | site |
|---|---|---|---|
| hard | benchmarks/hard/ | per-op kernels (CUDA or Triton), roofline-graded | /hard |
| mega | benchmarks/mega/ | full fused megakernels | /mega |
| cuda | benchmarks/cuda/ | CUDA-only writing deck (Triton/DSL fail) | /cuda |
| mini | benchmarks/mini/ | small-model (<200B) deck, capped + 5-repeat (WIP) | homepage scroll category on / when debuted (not /mini) |
| multi | benchmarks/multi/ | 4×H100 NVLink multi-GPU (WIP, frontier roster) | unpublished |
Hard / mega / cuda share harness machinery and run unlimited wall-clock
(BUDGET_SECONDS=0). Mini is capped; multi is sequential on a 4-GPU node.
Next.js 16 + Tailwind. Package manager is bun (bun.lock):
bun install
bun dev # http://localhost:3000
bun run build
Site data is baked at build time from benchmarks/*/results/
(app/_lib/data.ts). Publish/deploy: kb publish then kb deploy — see
app/AGENTS.md.
app/ public/ the website (app/AGENTS.md); app/_lib/data.ts bakes benchmark data at build time
benchmarks/<bench>/ AGENTS.md, SPEC.md, DEVLOG.md, problems, src, scripts, results/, outputs/runs/ (gitignored archives)
scripts/ shared runner (lib/run_harness.sh), workers, publish helpers
kbtool/ the `kb` CLI (uv package); `uv tool install -e ./kbtool` puts `kb` on PATH
media/ tracked chart generators (media/AGENTS.md); PNGs gitignored
runs/ gitignored HF staging from `kb push-runs` (not a third archive)
This list is closed. A new thing goes in the directory that owns it. kbtool/tests fails on a new top-level dir.
Bench virtualenvs exist only on the GPU boxes that run the benchmark; the
Mac checkout carries kbtool/.venv alone.