-> *Every LLM server I tried made me choose between convenience and control. I wanted to pin everyday models in memory, auto-swap heavier ones on demand, set context limits - and manage it all from a menu bar.*
+> *我试过的每个 LLM 服务器都让我在便利性和可控性之间做选择。我想把常用模型固定在内存中，按需自动切换较重的模型，设置上下文限制，并从菜单栏管理这一切。*
>
-> *oMLX persists KV cache across a hot in-memory tier and cold SSD tier - even when context changes mid-conversation, all past context stays cached and reusable across requests, making local LLMs practical for real coding work with tools like Claude Code. That's why I built it.*
+> *oMLX 将 KV 缓存持久化在热内存层和冷 SSD 层之间。即使对话中途上下文发生变化，所有历史上下文仍然保留在缓存中，可跨请求复用，让本地 LLM 在配合 Claude Code 等工具做实际编码时真正变得可用。这就是我做 oMLX 的原因。*
-## Install
+## 安装
-### macOS App
+### macOS 应用
-Download the `.dmg` from [Releases](https://github.com/jundot/omlx/releases), drag to Applications, done. The app includes in-app auto-update, so future upgrades are just one click. The macOS app also installs a lightweight `~/.omlx/bin/omlx` CLI shim so terminal commands and Apple Shortcuts can control the app-managed server.
+从 [Releases](https://github.com/jundot/omlx/releases) 下载 `.dmg`，拖到 Applications 即可。应用支持自动更新，后续升级只需一键完成。macOS 应用也会安装轻量的 `~/.omlx/bin/omlx` CLI shim，因此可以从终端命令或 Apple Shortcuts 控制由应用管理的服务器。
### Homebrew
@@ -62,217 +68,191 @@ Download the `.dmg` from [Releases](https://github.com/jundot/omlx/releases), dr
brew tap jundot/omlx https://github.com/jundot/omlx
brew install omlx
-# Upgrade to the latest version
+# 升级到最新版本
brew update && brew upgrade omlx
-# Run as a background service (auto-restarts on crash)
-omlx start
+# 作为后台服务运行（崩溃时自动重启）
+brew services start omlx
-# Optional: MCP (Model Context Protocol) support
+# 可选：MCP（Model Context Protocol）支持
/opt/homebrew/opt/omlx/libexec/bin/pip install mcp
```
-Optional GLM-5.2 / MiniMax M3 native custom kernels currently require a HEAD build:
+可选的 GLM-5.2 / MiniMax M3 原生自定义内核目前需要 HEAD 构建：
```bash
brew install omlx --HEAD --with-custom-kernel
```
-### From Source
+### 从源码安装
```bash
git clone https://github.com/jundot/omlx.git
cd omlx
-pip install -e . # Core only
-pip install -e ".[mcp]" # With MCP (Model Context Protocol) support
+pip install -e . # 仅核心
+pip install -e ".[mcp]" # 含 MCP（Model Context Protocol）支持
-# GLM-5.2 / MiniMax M3 / Qwen3.5 native custom kernels (strongly recommended
-# if you serve those families -- see note below)
+# 可选：GLM-5.2 / MiniMax M3 原生自定义内核
OMLX_WITH_CUSTOM_KERNEL=1 pip install -e .
```
-Requires macOS 15.0+ (Sequoia), Python 3.11–3.13, and Apple Silicon (M1/M2/M3/M4).
+需要 macOS 15.0+ (Sequoia), Python 3.10+ 和 Apple Silicon（M1/M2/M3/M4）。
-> **Note on native custom kernels:** a plain `pip install -e .` does NOT build
-> them, and the affected model families then silently fall back to much slower
-> generic paths -- for GLM-5.2 the fused DSA prefill is roughly 30x faster with
-> the kernels (measured 845 vs ~29 tok/s on an M3 Ultra), and the fallback also
-> uses more memory (#2137). Building them requires the Metal toolchain, which
-> Command Line Tools alone do not provide (`xcrun: error: unable to find utility
-> "metal"`): install full Xcode, or use the official DMG which ships the kernels
-> precompiled. Homebrew can build them with `brew install omlx --HEAD
-> --with-custom-kernel`, but that build also needs full Xcode. To verify your
-> install:
->
-> ```bash
-> python -c "from omlx.custom_kernels import native_kernel_status; print(native_kernel_status())"
-> ```
+## 快速开始
-## Quickstart
+### macOS 应用
-### macOS App
-
-Launch oMLX from your Applications folder. The Welcome screen guides you through three steps - model directory, server start, and first model download. That's it. To connect OpenClaw, OpenCode, Codex, Hermes Agent, or Copilot, see [Integrations](#integrations).
+从 Applications 文件夹启动 oMLX。欢迎界面会引导你完成三个步骤 — 模型目录设置、服务器启动、首个模型下载。就是这样。要连接 OpenClaw、OpenCode、Codex、Hermes Agent 或 Copilot，请参阅[集成](#集成)。
-
-
+
+
### CLI
```bash
-# Managed background server (macOS app or Homebrew install)
-omlx start
-omlx stop
-omlx restart
-
-# Foreground server attached to this terminal
omlx serve --model-dir ~/models
```
-The server discovers LLMs, VLMs, embedding models, and rerankers from subdirectories automatically. Any OpenAI-compatible client can connect to `http://localhost:8000/v1`. A built-in chat UI is also available at `http://localhost:8000/admin/chat`.
+服务器会自动从子目录中发现 LLM、VLM、嵌入模型和重排序模型。任何 OpenAI 兼容客户端都可以连接到 `http://localhost:8000/v1`。内置聊天 UI 也可在 `http://localhost:8000/admin/chat` 使用。
-### Homebrew Service
+### Homebrew 服务
-If you installed via Homebrew, you can run oMLX as a managed background service:
+如果通过 Homebrew 安装，可以将 oMLX 作为托管后台服务运行：
```bash
-omlx start # Start via brew services
-omlx stop # Stop
-omlx restart # Restart
-
-brew services start omlx # Start (auto-restarts on crash)
-brew services stop omlx # Stop
-brew services restart omlx # Restart
-brew services info omlx # Check status
+brew services start omlx # 启动（崩溃时自动重启）
+brew services stop omlx # 停止
+brew services restart omlx # 重启
+brew services info omlx # 查看状态
```
-The service runs `omlx serve` with zero-config defaults (`~/.omlx/models`, port 8000). `omlx start`, `omlx stop`, and `omlx restart` are the portable lifecycle commands; Homebrew installs delegate them to `brew services`. To customize, either set environment variables (`OMLX_MODEL_DIR`, `OMLX_PORT`, etc.) or run `omlx serve --model-dir /your/path` once to persist settings to `~/.omlx/settings.json`.
+服务使用默认配置运行 `omlx serve`（`~/.omlx/models`，端口 8000）。要自定义，可以设置环境变量（`OMLX_MODEL_DIR`、`OMLX_PORT` 等），或运行一次 `omlx serve --model-dir /your/path` 将设置保存到 `~/.omlx/settings.json`。
-Logs are written to two locations:
-- **Service log**: `$(brew --prefix)/var/log/omlx.log` (stdout/stderr)
-- **Server log**: `~/.omlx/logs/server.log` (structured application log)
+日志写入两个位置：
+- **服务日志**: `$(brew --prefix)/var/log/omlx.log`（stdout/stderr）
+- **服务器日志**: `~/.omlx/logs/server.log`（结构化应用日志）
-## Features
+## 功能
-Supports text LLMs, vision-language models (VLM), OCR models, embeddings, and rerankers on Apple Silicon.
+在 Apple Silicon 上支持文本 LLM、视觉语言模型（VLM）、OCR 模型、嵌入模型和重排序模型。
-### Admin Dashboard
+### 管理后台
-Web UI at `/admin` for real-time monitoring, model management, chat, benchmark, and per-model settings. Supports English, Korean, Japanese, Chinese, French, Russian, Spanish, and Brazilian Portuguese. All CDN dependencies are vendored for fully offline operation.
+在 `/admin` 提供实时监控、模型管理、聊天、基准测试和模型级设置的 Web UI。支持英语、韩语、日语、中文和俄语。所有 CDN 依赖已内置，完全支持离线运行。
-
+
-### Vision-Language Models
+### 视觉语言模型
-Run VLMs with the same continuous batching and tiered KV cache stack as text LLMs. Supports multi-image chat, base64/URL/file image inputs, and tool calling with vision context. OCR models (DeepSeek-OCR, DOTS-OCR, GLM-OCR) are auto-detected with optimized prompts.
+使用与文本 LLM 相同的连续批处理和分层 KV 缓存堆栈运行 VLM。支持多图聊天、base64/URL/文件图像输入，以及带视觉上下文的工具调用。OCR 模型（DeepSeek-OCR、DOTS-OCR、GLM-OCR）会被自动识别，并使用优化的提示词。
-### Tiered KV Cache (Hot + Cold)
+### 分层 KV 缓存（热缓存 + 冷缓存）
-Block-based KV cache management inspired by vLLM, with prefix sharing and Copy-on-Write. The cache operates across two tiers:
+借鉴 vLLM 的基于块的 KV 缓存管理，支持前缀共享和写时复制（Copy-on-Write）。缓存分为两个层级：
-- **Hot tier (RAM)**: Frequently accessed blocks stay in memory for fast access.
-- **Cold tier (SSD)**: When the hot cache fills up, blocks are offloaded to SSD in safetensors format. On the next request with a matching prefix, they're restored from disk instead of recomputed from scratch - even after a server restart.
+- **热缓存（RAM）**: 频繁访问的块保留在内存中，实现快速读取。
+- **冷缓存（SSD）**: 热缓存满时，块会以 safetensors 格式转存到 SSD。下次请求命中相同前缀时，直接从磁盘恢复，无需重新计算 — 即使服务器重启也不会丢失。
-
+
-### Continuous Batching
+### 连续批处理
-Handles concurrent requests through mlx-lm's BatchGenerator. Max concurrent requests is configurable via CLI or admin panel.
+通过 mlx-lm 的 BatchGenerator 处理并发请求。最大并发请求数可通过 CLI 或管理面板配置。
-### Claude Code Optimization
+### Claude Code 优化
-Context scaling support for running smaller context models with Claude Code. Scales reported token counts so that auto-compact triggers at the right timing, and SSE keep-alive prevents read timeouts during long prefill.
+支持在 Claude Code 中使用较小上下文模型的上下文缩放。通过缩放上报的 Token 数量，让自动压缩在合适的时机触发，同时提供 SSE keep-alive 防止长时间预填充导致的读取超时。
-### Multi-Model Serving
+### 多模型服务
-Load LLMs, VLMs, embedding models, and rerankers within the same server. Models are managed through a combination of automatic and manual controls:
+在同一服务器中加载 LLM、VLM、嵌入模型和重排序模型。通过自动和手动控制的组合管理模型：
-- **LRU eviction**: Least-recently-used models are evicted automatically when memory runs low.
-- **Manual load/unload**: Interactive status badges in the admin panel let you load or unload models on demand.
-- **Model pinning**: Pin frequently used models to keep them always loaded.
-- **Per-model TTL**: Set an idle timeout per model to auto-unload after a period of inactivity.
-- **Process memory enforcement**: Total memory limit (default: system RAM - 8GB) prevents system-wide OOM.
+- **LRU 驱逐**: 内存不足时，最近最少使用的模型会被自动卸载。
+- **手动加载/卸载**: 在管理后台通过状态标识按需加载或卸载模型。
+- **模型固定**: 固定常用模型使其始终保持加载状态。
+- **模型级 TTL**: 为每个模型设置空闲超时，在一段时间不活动后自动卸载。
+- **进程内存限制**: 总内存限制（默认：系统 RAM - 8GB）防止系统级 OOM。
-### Per-Model Settings
+### 模型级设置
-Configure sampling parameters, chat template kwargs, TTL, model alias, model type override, and more per model directly from the admin panel. Changes apply immediately without server restart.
+在管理后台直接配置每个模型的采样参数、聊天模板参数、TTL、模型别名、模型类型覆盖等。修改即时生效，无需重启服务器。
-- **Model alias**: set a custom API-visible name. `/v1/models` returns the alias, and requests accept both the alias and directory name.
-- **Model type override**: manually set a model as LLM or VLM regardless of auto-detection.
-- **Profiles**: save named bundles of per-model settings and switch between them from the admin panel. A profile can optionally be exposed as its own model: `/v1/models` then also lists `:` (e.g. `qwen3-8b:thinking`), which serves on the same engine as the base model with the profile's settings overlaid per request — no extra memory, no reload. When the base model has an alias, the exposed ID is advertised as `:`; the directory-name form keeps working, just like for the base model.
+- **模型别名**: 设置自定义 API 显示名称。`/v1/models` 返回别名，请求时别名和目录名均可使用。
+- **模型类型覆盖**: 无论自动检测结果如何，手动设置为 LLM 或 VLM。
+- **配置文件**: 保存每个模型的命名设置组合，并在管理后台中切换。配置文件可以选择性地作为独立模型公开：`/v1/models` 随后也会列出 `<模型>:<配置文件>`（例如 `qwen3-8b:thinking`），它在与基础模型相同的引擎上运行，并按请求叠加该配置文件的设置 — 无需额外内存，无需重新加载。当基础模型有别名时，公开的 ID 以 `<别名>:<配置文件>` 形式呈现；目录名形式仍然有效，与基础模型一样。
-
+
-### Built-in Chat
+### 内置聊天
-Chat directly with any loaded model from the admin panel. Supports conversation history, model switching, dark mode, reasoning model output, and image upload for VLM/OCR models.
+从管理后台直接与已加载的模型聊天。支持对话历史、模型切换、深色模式、推理模型输出，以及 VLM/OCR 模型的图片上传。
-
+
-### Model Downloader
+### 模型下载器
-Search and download MLX models from HuggingFace directly in the admin dashboard. Browse model cards, check file sizes, and download with one click.
+在管理后台中直接搜索和下载 HuggingFace 上的 MLX 模型。浏览模型卡片、查看文件大小，一键下载。
-
+
-### Integrations
+### 集成
-Set up OpenClaw, OpenCode, Codex, Hermes Agent, Copilot, and Pi directly from the admin dashboard with a single click. No manual config editing required.
+在管理后台中一键设置 OpenClaw、OpenCode、Codex、Hermes Agent、Copilot 和 Pi。无需手动编辑配置文件。
-
+
-### Performance Benchmark
+### 性能基准测试
-One-click benchmarking from the admin panel. Measures prefill (PP) and text generation (TG) tokens per second, with partial prefix cache hit testing for realistic performance numbers.
+从管理后台一键运行基准测试。测量预填充（PP）和 Token 生成（TG）的每秒 Token 数，包含部分前缀缓存命中测试以获得真实的性能数据。
-
+
-### macOS Menubar App
+### macOS 菜单栏应用
-Native Swift / SwiftUI menubar app (not Electron). Start, stop, and monitor the server without opening a terminal. Includes persistent serving stats (survives restarts), auto-restart on crash, and Sparkle-driven auto-update.
+原生 Swift / SwiftUI 菜单栏应用（非 Electron）。无需打开终端即可启动、停止和监控服务器。包含持久化服务统计（重启后保留）、崩溃自动重启和基于 Sparkle 的自动更新。
-
+
-### API Compatibility
+### API 兼容性
-Drop-in replacement for OpenAI and Anthropic APIs. Supports streaming usage stats (`stream_options.include_usage`), Anthropic adaptive thinking, and vision inputs (base64, URL).
+OpenAI 和 Anthropic API 的直接替代品。支持流式使用统计（`stream_options.include_usage`）、Anthropic adaptive thinking 和视觉输入（base64、URL）。
-| Endpoint | Description |
-|----------|-------------|
-| `POST /v1/chat/completions` | Chat completions (streaming) |
-| `POST /v1/completions` | Text completions (streaming) |
+| 端点 | 说明 |
+|----------|------|
+| `POST /v1/chat/completions` | 聊天补全（流式） |
+| `POST /v1/completions` | 文本补全（流式） |
| `POST /v1/messages` | Anthropic Messages API |
-| `POST /v1/embeddings` | Text embeddings |
-| `POST /v1/rerank` | Document reranking |
-| `GET /v1/models` | List available models |
+| `POST /v1/embeddings` | 文本嵌入 |
+| `POST /v1/rerank` | 文档重排序 |
+| `GET /v1/models` | 列出可用模型 |
-### Tool Calling & Structured Output
+### 工具调用与结构化输出
-Supports all function calling formats available in mlx-lm, JSON schema validation, and MCP tool integration. Tool calling requires the model's chat template to support the `tools` parameter. The following model families are auto-detected via mlx-lm's built-in tool parsers:
+支持 mlx-lm 中所有可用的函数调用格式、JSON Schema 验证和 MCP 工具集成。工具调用需要模型的聊天模板支持 `tools` 参数。以下模型系列通过 mlx-lm 的内置工具解析器自动检测：
-| Model Family | Format |
+| 模型系列 | 格式 |
|---|---|
-| Llama, Qwen, DeepSeek, etc. | JSON `` |
-| Qwen3.5 Series | XML `` |
+| Llama、Qwen、DeepSeek 等 | JSON `` |
+| Qwen3.5 系列 | XML `` |
| Gemma | `` |
| GLM (4.7, 5) | `/` XML |
| MiniMax | Namespaced `` |
@@ -280,11 +260,11 @@ Supports all function calling formats available in mlx-lm, JSON schema validatio
| Kimi K2 | `<\|tool_calls_section_begin\|>` |
| Longcat | `` |
-Models not listed above may still work if their chat template accepts `tools` and their output uses a recognized `` XML format. For tool-enabled streaming, assistant text is emitted incrementally while known tool-call control markup is suppressed from visible content; structured tool calls are emitted after parsing the completed turn.
+上表未列出的模型，只要聊天模板支持 `tools` 参数且输出采用可识别的 `` XML 格式，也有可能正常工作。针对支持工具调用的流式请求，系统会增量发射助手文本，同时隐藏已知的工具调用控制标记；结构化工具调用将在完成整个回合解析后发射。
-## Models
+## 模型
-Point `--model-dir` at a directory containing MLX-format model subdirectories. Two-level organization folders (e.g., `mlx-community/model-name/`) are also supported.
+将 `--model-dir` 指向包含 MLX 格式模型子目录的目录。支持两级目录结构（如 `mlx-community/model-name/`）。
```
~/models/
@@ -295,83 +275,75 @@ Point `--model-dir` at a directory containing MLX-format model subdirectories. T
└── bge-m3/
```
-Models are auto-detected by type. You can also download models directly from the admin dashboard.
+模型会按类型自动识别。也可以直接在管理后台下载模型。
-| Type | Models |
-|------|--------|
-| LLM | Any model supported by [mlx-lm](https://github.com/ml-explore/mlx-lm) |
-| VLM | Qwen3.5 Series, GLM-4V, Pixtral, and other [mlx-vlm](https://github.com/Blaizzy/mlx-vlm) models |
-| OCR | DeepSeek-OCR, DOTS-OCR, GLM-OCR |
-| Embedding | BERT, BGE-M3, ModernBERT |
-| Reranker | ModernBERT, XLM-RoBERTa |
+| 类型 | 模型 |
+|------|------|
+| LLM | [mlx-lm](https://github.com/ml-explore/mlx-lm) 支持的所有模型 |
+| VLM | Qwen3.5 系列、GLM-4V、Pixtral 及其他 [mlx-vlm](https://github.com/Blaizzy/mlx-vlm) 模型 |
+| OCR | DeepSeek-OCR、DOTS-OCR、GLM-OCR |
+| 嵌入 | BERT、BGE-M3、ModernBERT |
+| 重排序 | ModernBERT、XLM-RoBERTa |
-## CLI Configuration
+## CLI 配置
```bash
-# Managed background server (macOS app or Homebrew install)
-omlx start
-omlx stop
-omlx restart
+# 已加载模型的内存限制
+omlx serve --model-dir ~/models --max-model-memory 32GB
-# Start with default settings (memory guard tier = balanced, manage via admin UI)
-omlx serve --model-dir ~/models
+# 进程级内存限制（默认：auto = RAM - 8GB）
+omlx serve --model-dir ~/models --max-process-memory 80%
-# Choose a memory guard tier at startup
-omlx serve --model-dir ~/models --memory-guard safe
-
-# Set a custom memory guard ceiling in GB
-omlx serve --model-dir ~/models --memory-guard-gb 48
-
-# Enable SSD cache for KV blocks
+# 启用 KV 块的 SSD 缓存
omlx serve --model-dir ~/models --paged-ssd-cache-dir ~/.omlx/cache
-# Set in-memory hot cache size
+# 设置内存热缓存大小
omlx serve --model-dir ~/models --hot-cache-max-size 20%
-# Adjust max concurrent requests (default: 8)
+# 调整最大并发请求数（默认: 8）
omlx serve --model-dir ~/models --max-concurrent-requests 16
-# With MCP tools
+# 使用 MCP 工具
omlx serve --model-dir ~/models --mcp-config mcp.json
-# HuggingFace mirror endpoint (for restricted regions)
+# HuggingFace 镜像端点（适用于受限地区）
omlx serve --model-dir ~/models --hf-endpoint https://hf-mirror.com
-# API key authentication
+# API 密钥认证
omlx serve --model-dir ~/models --api-key your-secret-key
-# Localhost-only: skip verification via admin panel global settings
+# 仅限 Localhost：在管理后台全局设置中跳过验证
```
-All settings can also be configured from the web admin panel at `/admin`. Settings are persisted to `~/.omlx/settings.json`, and CLI flags take precedence.
+以上所有设置也可以在 `/admin` 的 Web 管理后台中配置。设置保存在 `~/.omlx/settings.json`，CLI 参数优先级更高。
-Architecture
+架构
```
FastAPI Server (OpenAI / Anthropic API)
│
- ├── EnginePool (multi-model, LRU eviction, TTL, manual load/unload)
- │ ├── BatchedEngine (LLMs, continuous batching)
- │ ├── VLMEngine (vision-language models)
+ ├── EnginePool (多模型、LRU 驱逐、TTL、手动加载/卸载)
+ │ ├── BatchedEngine (LLM，连续批处理)
+ │ ├── VLMEngine (视觉语言模型)
│ ├── EmbeddingEngine
│ └── RerankerEngine
│
- ├── ProcessMemoryEnforcer (total memory limit, TTL checks)
+ ├── ProcessMemoryEnforcer (总内存限制、TTL 检查)
│
- ├── Scheduler (FCFS, configurable concurrency)
+ ├── Scheduler (FCFS，可配置并发数)
│ └── mlx-lm BatchGenerator
│
└── Cache Stack
- ├── PagedCacheManager (GPU, block-based, CoW, prefix sharing)
- ├── Hot Cache (in-memory tier, write-back)
- └── PagedSSDCacheManager (SSD cold tier, safetensors format)
+ ├── PagedCacheManager (GPU，基于块，CoW，前缀共享)
+ ├── Hot Cache (内存缓存，write-back)
+ └── PagedSSDCacheManager (SSD 冷缓存，safetensors 格式)
```
-## Development
+## 开发
-### CLI Server
+### CLI 服务器
```bash
git clone https://github.com/jundot/omlx.git
@@ -380,44 +352,43 @@ pip install -e ".[dev]"
pytest -m "not slow"
```
-### macOS App
+### macOS 应用
-The native SwiftUI app lives at `apps/omlx-mac/`. Requires Xcode 26.5+ and Python 3.11+. venvstacks is declared as a dev dependency so `pip install -e ".[dev]"` (or `uv sync --dev`) brings the pinned version in. The build script also falls back to `uvx venvstacks` or `pipx run venvstacks` if you prefer a host-global tool runner.
+原生 SwiftUI 应用位于 `apps/omlx-mac/`，需要 Xcode 26.5+ 和 Python 3.11+。venvstacks 已声明为 dev 依赖，因此 `pip install -e ".[dev]"`（或 `uv sync --dev`）会引入固定版本。若偏好主机全局工具运行器，也可使用 `uvx venvstacks` 或 `pipx run venvstacks`。
```bash
-# Stage a runnable oMLX.app (xcodebuild + venvstacks Python layers + ad-hoc sign)
+# 暂存可运行的 oMLX.app（xcodebuild + venvstacks Python 层 + ad-hoc 签名）
apps/omlx-mac/Scripts/build.sh release
-# Result lands at apps/omlx-mac/build/Stage/oMLX.app
+# 结果在 apps/omlx-mac/build/Stage/oMLX.app
open apps/omlx-mac/build/Stage/oMLX.app
-# Force a fresh venvstacks rebuild (otherwise it's cached by fingerprint)
+# 强制重建 venvstacks（默认按指纹缓存）
apps/omlx-mac/Scripts/build.sh release --rebuild-donor
-# Stage with optional GLM-5.2 / MiniMax M3 native custom kernels
+# 暂存包含可选 GLM-5.2 / MiniMax M3 原生自定义内核的应用
apps/omlx-mac/Scripts/build.sh release --with-custom-kernel
```
-First cold build takes 10–20 minutes (venvstacks Python layer assembly). Subsequent builds reuse the cached `packaging/_export/` and finish in about 4 minutes. See [packaging/README.md](packaging/README.md) for the layer configuration and [apps/omlx-mac/](apps/omlx-mac/) for the Swift sources.
+首次 cold 构建需要 10–20 分钟（venvstacks Python 层组装）。后续构建复用 `packaging/_export/` 缓存，约 4 分钟完成。层配置请参阅 [packaging/README.md](packaging/README.md)，Swift 源码请参阅 [apps/omlx-mac/](apps/omlx-mac/)。
-## Contributing
+## 贡献
-Contributions are welcome! See [Contributing Guide](docs/CONTRIBUTING.md) for details.
+欢迎贡献！详情请参阅[贡献指南](docs/CONTRIBUTING.md)。
-- Bug fixes and improvements
-- Performance optimizations
-- Documentation improvements
+- Bug 修复和改进
+- 性能优化
+- 文档改进
-## License
+## 许可证
[Apache 2.0](LICENSE)
-## Acknowledgments
+## 致谢
-- [MLX](https://github.com/ml-explore/mlx) and [mlx-lm](https://github.com/ml-explore/mlx-lm) by Apple
-- [mlx-vlm](https://github.com/Blaizzy/mlx-vlm) - Vision-language model inference on Apple Silicon
-- [vllm-mlx](https://github.com/waybarrios/vllm-mlx) - oMLX started from vllm-mlx v0.1.0 and evolved significantly with multi-model serving, tiered KV caching, VLM with full paged cache support, an admin panel, and a macOS menu bar app
-- [venvstacks](https://venvstacks.lmstudio.ai) - Portable Python environment layering for the macOS app bundle
-- [mlx-embeddings](https://github.com/Blaizzy/mlx-embeddings) - Embedding model support for Apple Silicon
-- [dflash-mlx](https://github.com/bstnxbt/dflash-mlx) - Block diffusion speculative decoding on Apple Silicon
-- [MTPLX](https://github.com/youssofal/mtplx) - Lightning MTP's verify-shape Metal kernels are powered by MTPLX by Youssof Altoukhi, which also inspired the depth-k pipeline
+- [MLX](https://github.com/ml-explore/mlx) 和 [mlx-lm](https://github.com/ml-explore/mlx-lm) by Apple
+- [mlx-vlm](https://github.com/Blaizzy/mlx-vlm) - Apple Silicon 上的视觉语言模型推理
+- [vllm-mlx](https://github.com/waybarrios/vllm-mlx) - oMLX 从 vllm-mlx v0.1.0 起步，经过大幅演进，增加了多模型服务、分层 KV 缓存、完整分页缓存支持的 VLM、管理后台和 macOS 菜单栏应用
+- [venvstacks](https://venvstacks.lmstudio.ai) - macOS 应用包的便携 Python 环境分层
+- [mlx-embeddings](https://github.com/Blaizzy/mlx-embeddings) - Apple Silicon 嵌入模型支持
+- [dflash-mlx](https://github.com/bstnxbt/dflash-mlx) - Apple Silicon 上的块扩散推测解码 (Block diffusion speculative decoding)
@@ -15,7 +21,7 @@
+ 
- 
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+