Hyperstitioncc/Hyperkernel

选用指南

AI 根据 README 整理 · 2026/11/6 11:17:04

使用方式

通过命令行使用，顶层命令分支为 hyperserve <llm | drive | dna | world>，每个命令有 --help 选项。

查看 README 依据

The user must provide the model checkpoints as directories of `.safetensors` files; none is included in the repository. They are DeepSeek-V4.1-Flash with its `tokenizer.json`, openpilot v0.11.2's driving model, Evo 2 7B (`arcinstitute/evo2_7b`) and Matrix-Game 3.0 (`Skywork/Matrix-Game-3.0`). Building needs Rust (see `rust-toolchain.toml`) and an NVIDIA driver R580 or newer: ```sh cargo build --release ``` ## CLI format The top level CLI branches on the type of host server you want to run: ```sh hyperserve <llm | drive | dna | world> --model <name> --weights DIR [options] ```

Each command picks its kernels from `kernels/b300/` and lists its options with `--help`. ## LLMs ```sh hyperserve llm --model deepseek-v4.1-flash --weights DIR [--preset c1] [--host 127.0.0.1] [--port 8000] [--api-key KEY] ``` Runs an OpenAI-compatible server on 4 GPUs. `--weights` is the checkpoint directory of `deepseek-ai/DeepSeek-V4.1-Flash` at revision `314113f8a7dadb388215396ba5fbee2c610a04d6`, as published, which also holds the tokenizer. Requests are OpenAI chat or completion JSON; with `--api-key` (or `HYPERSERVE_API_KEY`) they must send `Authorization: Bearer KEY`.

部署要求

需要 NVIDIA B300 GPU，LLM 需 4 块，驱动和 DNA 各需 1 块，世界模型需 2 块；LLM 还需约 203 GB 主机内存。需要 Rust 和 NVIDIA 驱动 R580 或更新版本。

查看 README 依据

# Hyperserve This repository contains the public implementations of the Megakernels generated by the Hyperkernel. This repository is in Alpha, and will be continuously updated over the next few weeks with more models and API simplifications, alongside performance improvements. ## Requirements The current megakernels are compiled for NVIDIA B300s: 4 for LLMs, 1 for driving and DNA, 2 for world models. Each command runs on the first GPUs `CUDA_VISIBLE_DEVICES` lists. LLMs also need about 203 GB of host RAM for the Engram tables, mapped once for all 4 GPU processes.

The user must provide the model checkpoints as directories of `.safetensors` files; none is included in the repository. They are DeepSeek-V4.1-Flash with its `tokenizer.json`, openpilot v0.11.2's driving model, Evo 2 7B (`arcinstitute/evo2_7b`) and Matrix-Game 3.0 (`Skywork/Matrix-Game-3.0`). Building needs Rust (see `rust-toolchain.toml`) and an NVIDIA driver R580 or newer: ```sh cargo build --release ``` ## CLI format The top level CLI branches on the type of host server you want to run: ```sh hyperserve <llm | drive | dna | world> --model <name> --weights DIR [options] ```

已知限制

仓库处于 Alpha 阶段，将不断更新。模型检查点需用户自行提供，仓库不包含任何检查点。

查看 README 依据

# Hyperserve This repository contains the public implementations of the Megakernels generated by the Hyperkernel. This repository is in Alpha, and will be continuously updated over the next few weeks with more models and API simplifications, alongside performance improvements. ## Requirements The current megakernels are compiled for NVIDIA B300s: 4 for LLMs, 1 for driving and DNA, 2 for world models. Each command runs on the first GPUs `CUDA_VISIBLE_DEVICES` lists. LLMs also need about 203 GB of host RAM for the Engram tables, mapped once for all 4 GPU processes.

The user must provide the model checkpoints as directories of `.safetensors` files; none is included in the repository. They are DeepSeek-V4.1-Flash with its `tokenizer.json`, openpilot v0.11.2's driving model, Evo 2 7B (`arcinstitute/evo2_7b`) and Matrix-Game 3.0 (`Skywork/Matrix-Game-3.0`). Building needs Rust (see `rust-toolchain.toml`) and an NVIDIA driver R580 or newer: ```sh cargo build --release ``` ## CLI format The top level CLI branches on the type of host server you want to run: ```sh hyperserve <llm | drive | dna | world> --model <name> --weights DIR [options] ```