OpenBMB/ForgeKernel

选用指南

AI 根据 README 整理 · 2026/10/30 20:52:12

适用场景

适用于在NVIDIA Hopper或Blackwell GPU上优化现有内核，或生成新内核，可处理source、blackbox、generate和leaderboard模式。

查看 README 依据

| `[reference]` | Mode | What the teacher sees | |---|---|---| | `run` + `source` | `source` | the reference's source and its hardware behavior | | `run` only, or `blackbox = true` | `blackbox` | hardware behavior only | | (none) | generate | a `/goal` session writes a reference first | | `[leaderboard]` instead | leaderboard | the board's problem, judge and clock | A task writes only what differs from `global_config/`, the bottom layer every task is merged onto, as in ForgeTrain: `models.toml` sets `[agent]` (`backend`, `provider`, `model`, `effort`,

使用方式

通过命令行使用，安装后运行`forge new`创建任务，`forge run`执行优化循环，`forge status`查看进度。

查看 README 依据

cp local.env.example local.env # FK_DELIVERY=local -> the GPU on this machine (FK_GPU picks the device) # FK_DELIVERY=ssh -> REMOTE_HOST (an ssh alias that logs in without a # prompt) and REMOTE_WORKDIR (a directory on it) ``` **3. Run the example task**: fuse a PyTorch RMSNorm (nine kernel launches a call) into one CuTe DSL kernel on a Blackwell GPU (set `arch = "sm90"` in `task.toml` for Hopper). ```bash forge new tasks/rmsnorm --dry-run # check the task, print the plan

forge new tasks/rmsnorm # -> .artifacts/rmsnorm_<id> tmux new -s forge 'forge run .artifacts/rmsnorm_<id>' # a run takes hours forge status .artifacts/rmsnorm_<id> # progress forge run .artifacts/rmsnorm_<id> # continue after an interruption ``` `forge run` exits 0 when the run reaches its target -- the loop stopped on it, or the final confirmation of the best kernel meets it -- and 2 otherwise. **Expected output.** The run writes the correctness judge, calibrates it against the reference,

部署要求

需要Linux或macOS，Python ≥ 3.11，git，以及Claude Code CLI（已登录或提供ANTHROPIC_API_KEY）；GPU机器需NVIDIA Hopper或Blackwell GPU，并安装CUDA版PyTorch。

查看 README 依据

install it in a virtual environment, write local.env for my GPU (this machine, or the ssh host I name) and run `forge new tasks/rmsnorm --dry-run`. Show me the plan it prints. ``` ## Quick Start **Requirements.** The machine you run `forge` on: Linux or macOS, Python ≥ 3.11, git, and the [Claude Code](https://docs.claude.com/en/docs/claude-code/overview) CLI (logged in, or an `ANTHROPIC_API_KEY`); tmux if a task has no reference. The GPU machine (the same one, or an ssh host): an NVIDIA Hopper (`sm90`) or Blackwell (`sm100`) GPU, `python3` with a CUDA build of

已知限制

参考实现有助于优化，但仅凭PyTorch实现可能停滞；自进化在噪声底限处停止，不保证无进一步提升；覆盖范围限于两类NVIDIA GPU上的四个算子家族；运行需数小时GPU时间和前沿模型使用。

查看 README 依据

- **A reference helps most.** Given only a PyTorch implementation and a device-time number, the paper's ablation stalls on all three operators tested; ForgeKernel surpasses an implementation that exists more reliably than it writes the first one. - **A plateau is an estimate, not a proof.** Self-evolution stops at the same-job noise floor: the loop found nothing more, which does not mean nothing more exists. - **Coverage.** Four operator families on two NVIDIA GPU generations; sparsity patterns and memory hierarchies unlike these are untested.

- **Cost.** A run is hours of GPU time and frontier-model usage, and the agents run with permissions bypassed inside a policy the driver writes; run it on a machine meant for it. ## License Apache License 2.0. See [LICENSE](LICENSE). The leaderboard evaluator adapts code from FlashInfer-Bench and SOL-ExecBench (both Apache-2.0); see [NOTICE](NOTICE). ## Acknowledgments ForgeKernel builds on [CUTLASS and CuTe DSL](https://github.com/NVIDIA/cutlass), [Nsight Compute](https://developer.nvidia.com/nsight-compute), [Claude Code](https://docs.claude.com/en/docs/claude-code/overview),