replicate-cli

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Replicate CLI

The Replicate CLI is a command-line tool for interacting with Replicate's AI model platform. It enables running predictions, managing models, creating deployments, and fine-tuning models directly from the terminal.

Replicate CLI是一款用于与Replicate AI模型平台交互的命令行工具，它支持直接在终端中运行预测、管理模型、创建部署以及微调模型。

Authentication

身份验证

Before using the Replicate CLI, set the API token:

bash

export REPLICATE_API_TOKEN=<token-from-replicate.com/account>

Alternatively, authenticate interactively:

bash

replicate auth login

Verify authentication:

bash

replicate account current

使用Replicate CLI之前，需设置API令牌：

bash

export REPLICATE_API_TOKEN=<token-from-replicate.com/account>

或者，通过交互式方式进行身份验证：

bash

replicate auth login

验证身份状态：

bash

replicate account current

Core Commands

核心命令

Running Predictions

运行预测

The primary use case is running predictions against hosted models.

Basic prediction:

bash

replicate run <owner/model> input_key=value

Examples:

Image generation:

bash

replicate run stability-ai/sdxl prompt="a studio photo of a rainbow colored corgi"

Text generation with streaming:

bash

replicate run meta/llama-2-70b-chat --stream prompt="Tell me a joke"

Prediction flags:

```
--stream
```
- Stream output tokens in real-time (for text models)
```
--no-wait
```
- Submit prediction without waiting for completion
```
--web
```
- Open prediction in browser
```
--json
```
- Output result as JSON
```
--save
```
- Save outputs to local directory

--output-directory <dir>

- Specify output directory (default:

./{prediction-id}

)

主要使用场景是针对托管模型运行预测。

基础预测：

bash

replicate run <owner/model> input_key=value

示例：

图像生成：

bash

replicate run stability-ai/sdxl prompt="a studio photo of a rainbow colored corgi"

流式文本生成：

bash

replicate run meta/llama-2-70b-chat --stream prompt="Tell me a joke"

预测参数：

```
--stream
```
- 实时流式输出令牌（适用于文本模型）
```
--no-wait
```
- 提交预测后无需等待完成
```
--web
```
- 在浏览器中打开预测页面
```
--json
```
- 以JSON格式输出结果
```
--save
```
- 将输出保存到本地目录

--output-directory <dir>

- 指定输出目录（默认：

./{prediction-id}

）

Input Handling

输入处理

File uploads: Prefix local file paths with

bash

replicate run nightmareai/real-esrgan image=@photo.jpg

Output chaining: Use

{{.output}}

template syntax to chain predictions:

bash

replicate run stability-ai/sdxl prompt="a corgi" | \
replicate run nightmareai/real-esrgan image={{.output[0]}}

文件上传： 本地文件路径前添加

前缀：

bash

replicate run nightmareai/real-esrgan image=@photo.jpg

输出链式调用： 使用

{{.output}}

模板语法实现预测链式调用：

bash

replicate run stability-ai/sdxl prompt="a corgi" | \
replicate run nightmareai/real-esrgan image={{.output[0]}}

Model Operations

模型操作

View model schema (see required inputs and outputs):

bash

replicate model schema <owner/model>
replicate model schema stability-ai/sdxl --json

List models:

bash

replicate model list
replicate model list --json

Show model details:

bash

replicate model show <owner/model>

Create a new model:

bash

replicate model create <owner/name> \
  --hardware gpu-a100-large \
  --private \
  --description "Model description"

Model creation flags:

```
--hardware <sku>
```
- Hardware SKU (see
```
references/hardware.md
```
)
```
--private
```
/
```
--public
```
- Visibility setting
```
--description <text>
```
- Model description
```
--github-url <url>
```
- Link to source repository
```
--license-url <url>
```
- License information
```
--cover-image-url <url>
```
- Cover image for model page

查看模型架构（了解必填输入与输出）：

bash

replicate model schema <owner/model>
replicate model schema stability-ai/sdxl --json

列出模型：

bash

replicate model list
replicate model list --json

查看模型详情：

bash

replicate model show <owner/model>

创建新模型：

bash

replicate model create <owner/name> \
  --hardware gpu-a100-large \
  --private \
  --description "Model description"

模型创建参数：

```
--hardware <sku>
```
- 硬件SKU（详见
```
references/hardware.md
```
）
```
--private
```
/
```
--public
```
- 可见性设置
```
--description <text>
```
- 模型描述
```
--github-url <url>
```
- 源代码仓库链接
```
--license-url <url>
```
- 许可证信息
```
--cover-image-url <url>
```
- 模型页面封面图

Training (Fine-tuning)

训练（微调）

Fine-tune models using the training command:

bash

replicate train <base-model> \
  --destination <owner/new-model> \
  input_key=value

Example - Fine-tune SDXL with DreamBooth:

bash

replicate train stability-ai/sdxl \
  --destination myuser/custom-sdxl \
  --web \
  input_images=@training-images.zip \
  use_face_detection_instead=true

List trainings:

bash

replicate training list

Show training details:

bash

replicate training show <training-id>

使用训练命令微调模型：

bash

replicate train <base-model> \
  --destination <owner/new-model> \
  input_key=value

示例 - 使用DreamBooth微调SDXL：

bash

replicate train stability-ai/sdxl \
  --destination myuser/custom-sdxl \
  --web \
  input_images=@training-images.zip \
  use_face_detection_instead=true

列出训练任务：

bash

replicate training list

查看训练任务详情：

bash

replicate training show <training-id>

Deployments

部署

Deployments provide dedicated, always-on inference endpoints with predictable performance.

Create deployment:

bash

replicate deployments create <name> \
  --model <owner/model> \
  --hardware <sku> \
  --min-instances 1 \
  --max-instances 3

Example:

bash

replicate deployments create text-to-image \
  --model stability-ai/sdxl \
  --hardware gpu-a100-large \
  --min-instances 1 \
  --max-instances 5

Update deployment:

bash

replicate deployments update <name> \
  --max-instances 10 \
  --version <version-id>

List deployments:

bash

replicate deployments list

Show deployment details and schema:

bash

replicate deployments show <name>
replicate deployments schema <name>

部署功能提供专用的、始终在线的推理端点，性能可预测。

创建部署：

bash

replicate deployments create <name> \
  --model <owner/model> \
  --hardware <sku> \
  --min-instances 1 \
  --max-instances 3

示例：

bash

replicate deployments create text-to-image \
  --model stability-ai/sdxl \
  --hardware gpu-a100-large \
  --min-instances 1 \
  --max-instances 5

更新部署：

bash

replicate deployments update <name> \
  --max-instances 10 \
  --version <version-id>

列出部署：

bash

replicate deployments list

查看部署详情与架构：

bash

replicate deployments show <name>
replicate deployments schema <name>

Hardware

硬件

List available hardware options:

bash

replicate hardware list

See

references/hardware.md

for detailed hardware information and selection guidelines.

列出可用的硬件选项：

bash

replicate hardware list

详见

references/hardware.md

获取硬件详细信息与选型指南。

Scaffolding

脚手架

Create a local development environment from an existing prediction:

bash

replicate scaffold <prediction-id-or-url> --template=<node|python>

This generates a project with the prediction's model and inputs pre-configured.

从现有预测创建本地开发环境：

bash

replicate scaffold <prediction-id-or-url> --template=<node|python>

该命令会生成一个预配置好预测模型与输入参数的项目。

Command Aliases

命令别名

For convenience, these aliases are available:

Alias	Equivalent Command
`replicate run`	`replicate prediction create`
`replicate stream`	`replicate prediction create --stream`
`replicate train`	`replicate training create`

Short aliases for subcommands:

```
replicate m
```
=
```
replicate model
```
```
replicate p
```
=
```
replicate prediction
```
```
replicate t
```
=
```
replicate training
```
```
replicate d
```
=
```
replicate deployments
```
```
replicate hw
```
=
```
replicate hardware
```
```
replicate a
```
=
```
replicate account
```

为方便使用，提供以下命令别名：

别名	等效命令
`replicate run`	`replicate prediction create`
`replicate stream`	`replicate prediction create --stream`
`replicate train`	`replicate training create`

子命令的短别名：

```
replicate m
```
=
```
replicate model
```
```
replicate p
```
=
```
replicate prediction
```
```
replicate t
```
=
```
replicate training
```
```
replicate d
```
=
```
replicate deployments
```
```
replicate hw
```
=
```
replicate hardware
```
```
replicate a
```
=
```
replicate account
```

Common Workflows

常见工作流

Image Generation Pipeline

图像生成流水线

Generate an image and upscale it:

bash

replicate run stability-ai/sdxl \
  prompt="professional photo of a sunset" \
  negative_prompt="blurry, low quality" | \
replicate run nightmareai/real-esrgan \
  image={{.output[0]}} \
  --save

生成图像并进行超分辨率处理：

bash

replicate run stability-ai/sdxl \
  prompt="professional photo of a sunset" \
  negative_prompt="blurry, low quality" | \
replicate run nightmareai/real-esrgan \
  image={{.output[0]}} \
  --save

Check Model Inputs Before Running

运行前检查模型输入

Always check the model schema to understand required inputs:

bash

replicate model schema owner/model-name

运行前务必查看模型架构，了解必填输入：

bash

replicate model schema owner/model-name

Batch Processing

批量处理

Run predictions and save outputs:

bash

for prompt in "cat" "dog" "bird"; do
  replicate run stability-ai/sdxl prompt="$prompt" --save --output-directory "./outputs/$prompt"
done

运行预测并保存输出：

bash

for prompt in "cat" "dog" "bird"; do
  replicate run stability-ai/sdxl prompt="$prompt" --save --output-directory "./outputs/$prompt"
done

Monitor Long-Running Tasks

监控长时间运行的任务

Submit without waiting, then check status:

bash

undefined

提交任务后无需等待，后续再检查状态：

bash

undefined

Submit

提交任务

replicate run owner/model input=value --no-wait --json > prediction.json

Check status later

后续检查状态

replicate prediction show $(jq -r '.id' prediction.json)

undefined

replicate prediction show $(jq -r '.id' prediction.json)

undefined

Best Practices

最佳实践

Always check schema first - Run
```
replicate model schema <model>
```
to understand required and optional inputs before running predictions.
Use streaming for text models - Add
```
--stream
```
flag when running language models to see output in real-time.
Save outputs explicitly - Use
```
--save
```
and
```
--output-directory
```
to organize prediction outputs.
Use JSON output for automation - Add
```
--json
```
flag when parsing outputs programmatically.
Open in web for debugging - Add
```
--web
```
flag to view predictions in the Replicate dashboard for detailed logs.
Chain predictions efficiently - Use the
```
{{.output}}
```
syntax to pass outputs between models without intermediate saves.

务必先查看架构 - 运行
```
replicate model schema <model>
```
以了解运行预测前需要的必填与可选输入。
文本模型使用流式输出 - 运行语言模型时添加
```
--stream
```
参数，实时查看输出内容。
显式保存输出 - 使用
```
--save
```
与
```
--output-directory
```
参数来整理预测输出文件。
自动化场景使用JSON输出 - 当需要以编程方式解析输出时，添加
```
--json
```
参数。
调试时使用网页查看 - 添加
```
--web
```
参数，在Replicate控制台中查看预测的详细日志。
高效实现链式预测 - 使用
```
{{.output}}
```
语法在模型间传递输出，无需中间保存文件。

Troubleshooting

故障排查

Authentication errors:

Verify
```
REPLICATE_API_TOKEN
```
is set correctly
Run
```
replicate account current
```
to test authentication

Model not found:

Check model name format:
```
owner/model-name
```
Verify model exists at replicate.com

Input validation errors:

Run
```
replicate model schema <model>
```
to see required inputs
Check input types (string, number, file)

File upload issues:

Ensure
```
@
```
prefix is used for local files
Verify file path is correct and file exists

身份验证错误：

确认
```
REPLICATE_API_TOKEN
```
已正确设置
运行
```
replicate account current
```
测试身份验证状态

模型未找到：

检查模型名称格式：
```
owner/model-name
```
确认模型在replicate.com上存在

输入验证错误：

运行
```
replicate model schema <model>
```
查看必填输入
检查输入类型（字符串、数字、文件）

文件上传问题：

确保本地文件路径前添加了
```
@
```
前缀
确认文件路径正确且文件存在

Additional Resources

额外资源

Replicate documentation: https://replicate.com/docs
Model explorer: https://replicate.com/explore
API reference: https://replicate.com/docs/reference/http
GitHub repository: https://github.com/replicate/cli

Replicate官方文档：https://replicate.com/docs
模型探索器：https://replicate.com/explore
API参考：https://replicate.com/docs/reference/http
GitHub仓库：https://github.com/replicate/cli