用于训练和推理开放权重模型的端到端基础设施

用于训练和推理开放权重模型的端到端基础设施
End-to-end infrastructure for training and inferencing open weight models

**Applied Compute Agent Cloud (AC2)** 是一个端到端的底层基础设施平台，专为开放权重模型的开发、推理及持续学习而设计。它充当了全面的研究人员驾驶舱，旨在简化整个模型生命周期。 **快速入门** 最快的上手方式是使用编码助手（例如 Cursor、Claude Code）。通过为助手提供配置链接（`https://platform.appliedcompute.com/agents.md`），它将自动处理 CLI 安装、身份验证和项目设置。用户也可以选择参考手动快速入门指南。 **核心组件** * **ac2.sdk：** 用于管理数据集、工作负载、密钥和部署。 * **ac2.runtime：** 用于编排智能体、环境、任务和评分器的框架。 * **ac2.tracing：** 基于 OpenTelemetry 的可观测性工具，通过简单的装饰器即可捕获函数级别的性能和逻辑。 **关键工作流** * **评估：** 运行基准测试、对追踪结果评分并比较智能体行为。 * **训练：** 支持 GRPO、监督微调（SFT）以及策略内自蒸馏（on-policy self-distillation）。 * **推理：** 部署模型，并通过可自定义的策略和端点管理流量路由。如需详细导航，用户可访问 `/llms.txt` 获取完整文档索引。

抱歉。

Applied Compute Agent Cloud (AC2) is a researcher cockpit for end-to-end model development, inference and continual learning.

Onboard with your coding agent

The fastest way to get started is to let your coding agent set AC2 up for you. Paste this prompt into Claude Code, Cursor, Codex, or any coding agent:

Fetch https://platform.appliedcompute.com/agents.md (e.g. curl -fsSL https://platform.appliedcompute.com/agents.md) and follow its instructions to set up AC2 for me.

The agent installs the CLI, logs you in, wires up the AC2 MCP server, and walks you through your first project. Prefer to run the steps yourself? Follow the Quickstart.

Packages

ac2.sdk — platform client for datasets, workloads, sessions, traces, and secrets.
ac2.runtime — agent framework for agents, environments, orchestrators, tasks, graders, and episodes.
ac2.tracing — OpenTelemetry-based observability SDK. Add @traced to any function to capture timing, inputs, outputs, and nested spans. Exports to ClickHouse and Langfuse.

Quickstart

Install the SDK and train a model.

Runtime

Define agents, environments, orchestrators, and eval tasks.

SDK

Manage datasets, deployments, runs, traces, and secrets from Python.

MCP server

Analyze eval and train traces from your editor.

Choose a workflow

Evaluate an agent

Run evals, grade traces, and compare behavior.

Train a model

Choose GRPO, supervised fine-tuning, or on-policy self-distillation.

Route inference

Deploy a model and route traffic through Policies and Endpoints.

FAQ

Find answers for failed runs, datasets, package installation, and cleanup.