colab-remote

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Colab Remote — SSH-Operated GPU Training

Colab Remote — 通过 SSH 操作的 GPU 训练

Operate Google Colab Pro/Pro+ instances as headless GPU backends from the local terminal.

从本地终端将 Google Colab Pro/Pro+ 实例作为无头 GPU 后端进行操作。

Architecture

架构

Local Mac (Claude Code)
  ├── agent-browser → Chrome → colab.research.google.com
  │   └── Opens notebook, runs colab-ssh setup cell
  ├── SSH tunnel → Colab runtime (via ngrok or cloudflared)
  │   └── Run training, monitor GPU, transfer files
  └── /autoany EGRI loop (local)
      └── Proposes mutations → SSH executes on Colab → evaluates results

Local Mac (Claude Code)
  ├── agent-browser → Chrome → colab.research.google.com
  │   └── Opens notebook, runs colab-ssh setup cell
  ├── SSH tunnel → Colab runtime (via ngrok or cloudflared)
  │   └── Run training, monitor GPU, transfer files
  └── /autoany EGRI loop (local)
      └── Proposes mutations → SSH executes on Colab → evaluates results

Phase 1: Launch Colab Session (Browser Automation)

阶段 1：启动 Colab 会话（浏览器自动化）

Use

/agent-browser

to open Colab and set up SSH access.

使用

/agent-browser

打开 Colab 并设置 SSH 访问。

Step 1: Open Colab and create notebook

步骤 1：打开 Colab 并创建 notebook

bash

agent-browser open "https://colab.research.google.com/#create=true"
agent-browser wait --load networkidle
agent-browser snapshot -i

If login is required, prompt the user to authenticate manually, then re-snapshot.

bash

agent-browser open "https://colab.research.google.com/#create=true"
agent-browser wait --load networkidle
agent-browser snapshot -i

如果需要登录，提示用户手动完成身份验证，然后重新执行 snapshot。

Step 2: Select GPU runtime

步骤 2：选择 GPU 运行时

Navigate Runtime > Change runtime type, select GPU (T4/V100/A100 depending on plan), and save.

前往 Runtime > Change runtime type，选择 GPU（根据套餐选择 T4/V100/A100），然后保存。

Step 3: Install colab-ssh and get connection details

步骤 3：安装 colab-ssh 并获取连接信息

Type the SSH setup code into a cell. Two methods supported:

Method A: ngrok (recommended)

python

!pip install colab-ssh --upgrade
from colab_ssh import launch_ssh
launch_ssh("YOUR_NGROK_TOKEN")

User must provide ngrok authtoken from https://ngrok.com.

Method B: cloudflared (no account needed)

python

!pip install colab-ssh --upgrade
from colab_ssh import launch_ssh_cloudflared
launch_ssh_cloudflared(password="your-password-here")

在 cell 中输入 SSH 设置代码。支持以下两种方式：

方式 A：ngrok（推荐）

python

!pip install colab-ssh --upgrade
from colab_ssh import launch_ssh
launch_ssh("YOUR_NGROK_TOKEN")

用户必须提供来自 https://ngrok.com 的 ngrok authtoken。

方式 B：cloudflared（无需账户）

python

!pip install colab-ssh --upgrade
from colab_ssh import launch_ssh_cloudflared
launch_ssh_cloudflared(password="your-password-here")

Step 4: Extract and save connection details

步骤 4：提取并保存连接信息

After the cell runs, snapshot output to extract hostname/port. Save for reuse:

bash

mkdir -p ~/.colab-remote
cat > ~/.colab-remote/session.env << 'EOF'
COLAB_HOST=0.tcp.ngrok.io
COLAB_PORT=12345
COLAB_USER=root
COLAB_METHOD=ngrok
EOF

Load in subsequent commands:

source ~/.colab-remote/session.env

cell 运行后，对输出执行 snapshot，以提取 hostname/port。保存以供后续使用：

bash

mkdir -p ~/.colab-remote
cat > ~/.colab-remote/session.env << 'EOF'
COLAB_HOST=0.tcp.ngrok.io
COLAB_PORT=12345
COLAB_USER=root
COLAB_METHOD=ngrok
EOF

在后续命令中加载：

source ~/.colab-remote/session.env

Phase 2: SSH Operations

阶段 2：SSH 操作

Connect

连接

bash

undefined

bash

undefined

ngrok

ssh -o StrictHostKeyChecking=no -p $COLAB_PORT root@$COLAB_HOST

cloudflared

ssh -o StrictHostKeyChecking=no -o ProxyCommand="cloudflared access ssh --hostname %h" root@$COLAB_HOST

undefined

ssh -o StrictHostKeyChecking=no -o ProxyCommand="cloudflared access ssh --hostname %h" root@$COLAB_HOST

undefined

Verify GPU

验证 GPU

bash

ssh -p $COLAB_PORT root@$COLAB_HOST "nvidia-smi"

bash

ssh -p $COLAB_PORT root@$COLAB_HOST "nvidia-smi"

Transfer files

传输文件

bash

undefined

bash

undefined

Upload

scp -P $COLAB_PORT -r ./data root@$COLAB_HOST:/content/data

Download

scp -P $COLAB_PORT -r root@$COLAB_HOST:/content/checkpoints ./checkpoints

undefined

scp -P $COLAB_PORT -r root@$COLAB_HOST:/content/checkpoints ./checkpoints

undefined

Run training

运行训练

bash

undefined

bash

undefined

Foreground

ssh -p $COLAB_PORT root@$COLAB_HOST "cd /content && python train.py --epochs 10"

Background (survives SSH disconnect)

ssh -p $COLAB_PORT root@$COLAB_HOST "cd /content && nohup python train.py > train.log 2>&1 &"

Monitor

ssh -p $COLAB_PORT root@$COLAB_HOST "tail -f /content/train.log"

undefined

ssh -p $COLAB_PORT root@$COLAB_HOST "tail -f /content/train.log"

undefined

Monitor GPU

监控 GPU

bash

ssh -p $COLAB_PORT root@$COLAB_HOST "nvidia-smi --query-gpu=utilization.gpu,utilization.memory,memory.used,memory.total,temperature.gpu --format=csv"

bash

ssh -p $COLAB_PORT root@$COLAB_HOST "nvidia-smi --query-gpu=utilization.gpu,utilization.memory,memory.used,memory.total,temperature.gpu --format=csv"

Install dependencies

安装依赖

bash

ssh -p $COLAB_PORT root@$COLAB_HOST "pip install torch transformers peft bitsandbytes accelerate datasets"

bash

ssh -p $COLAB_PORT root@$COLAB_HOST "pip install torch transformers peft bitsandbytes accelerate datasets"

Phase 3: EGRI Integration (/autoany)

阶段 3：EGRI 集成（/autoany）

Wire Colab as the execution backend for an EGRI optimization loop. See

references/egri-colab.md

for the full problem-spec template and harness patterns.

将 Colab 接入 EGRI 优化循环，作为执行后端。完整的 problem-spec 模板和 harness 模式请参阅

references/egri-colab.md

。

Execution loop (summary)

执行循环（摘要）

for each trial:
  1. Upload mutated artifact → scp to Colab
  2. Execute on Colab GPU → ssh python train.py
  3. Evaluate results → ssh python evaluate.py
  4. Download metrics → scp results.json
  5. Score locally (immutable evaluator)
  6. Promote or discard based on policy

for each trial:
  1. Upload mutated artifact → scp to Colab
  2. Execute on Colab GPU → ssh python train.py
  3. Evaluate results → ssh python evaluate.py
  4. Download metrics → scp results.json
  5. Score locally (immutable evaluator)
  6. Promote or discard based on policy

Phase 4: Session Lifecycle

阶段 4：会话生命周期

Tier	Max runtime	Idle timeout	GPU
Free	12h	90min	T4, limited
Pro	24h	90min	T4, V100, priority
Pro+	24h	90min	T4, V100, A100

套餐	最长运行时长	空闲超时	GPU
Free	12h	90min	T4，有限
Pro	24h	90min	T4, V100，优先
Pro+	24h	90min	T4, V100, A100

Keep-alive

保活

bash

ssh -p $COLAB_PORT root@$COLAB_HOST "while true; do sleep 300; echo keepalive; done &"

bash

ssh -p $COLAB_PORT root@$COLAB_HOST "while true; do sleep 300; echo keepalive; done &"

Reconnect after timeout

超时后重新连接

Check:

ssh -p $COLAB_PORT root@$COLAB_HOST "echo ok" 2>/dev/null && echo "UP" || echo "DOWN"

If dead, re-launch via Phase 1 (browser automation)
Resume from last checkpoint

检查：

ssh -p $COLAB_PORT root@$COLAB_HOST "echo ok" 2>/dev/null && echo "UP" || echo "DOWN"

如果连接已断开，则通过阶段 1（浏览器自动化）重新启动
从最后一个 checkpoint 恢复

Google Drive persistence

Google Drive 持久化

Mount Drive to persist across sessions:

bash

ssh -p $COLAB_PORT root@$COLAB_HOST "python -c 'from google.colab import drive; drive.mount(\"/content/drive\")'"

挂载 Drive，以跨会话持久化数据：

bash

ssh -p $COLAB_PORT root@$COLAB_HOST "python -c 'from google.colab import drive; drive.mount(\"/content/drive\")'"

Checkpoints survive in /content/drive/MyDrive/

undefined


Checkpoint 将保存在 /content/drive/MyDrive/ 中。

Quick Reference

快速参考

Task	Command
Check GPU	`ssh -p $COLAB_PORT root@$COLAB_HOST "nvidia-smi"`
Upload	`scp -P $COLAB_PORT ./file root@$COLAB_HOST:/content/`
Download	`scp -P $COLAB_PORT root@$COLAB_HOST:/content/file ./`
Run script	`ssh -p $COLAB_PORT root@$COLAB_HOST "python /content/script.py"`
Background job	`ssh -p $COLAB_PORT root@$COLAB_HOST "nohup python train.py > log 2>&1 &"`
Tail log	`ssh -p $COLAB_PORT root@$COLAB_HOST "tail -20 /content/log"`
Disk space	`ssh -p $COLAB_PORT root@$COLAB_HOST "df -h /content"`
Kill job	`ssh -p $COLAB_PORT root@$COLAB_HOST "pkill -f train.py"`
Session alive?	`ssh -p $COLAB_PORT root@$COLAB_HOST "echo ok" 2>/dev/null`

任务	命令
检查 GPU	`ssh -p $COLAB_PORT root@$COLAB_HOST "nvidia-smi"`
上传	`scp -P $COLAB_PORT ./file root@$COLAB_HOST:/content/`
下载	`scp -P $COLAB_PORT root@$COLAB_HOST:/content/file ./`
运行脚本	`ssh -p $COLAB_PORT root@$COLAB_HOST "python /content/script.py"`
后台任务	`ssh -p $COLAB_PORT root@$COLAB_HOST "nohup python train.py > log 2>&1 &"`
查看日志末尾	`ssh -p $COLAB_PORT root@$COLAB_HOST "tail -20 /content/log"`
磁盘空间	`ssh -p $COLAB_PORT root@$COLAB_HOST "df -h /content"`
终止任务	`ssh -p $COLAB_PORT root@$COLAB_HOST "pkill -f train.py"`
会话是否存活？	`ssh -p $COLAB_PORT root@$COLAB_HOST "echo ok" 2>/dev/null`

Prerequisites

前置条件

ngrok account (free): https://ngrok.com — or
```
cloudflared
```
:
```
brew install cloudflared
```
Colab Pro/Pro+ for GPU priority and longer runtimes
agent-browser installed and working
Google account signed into Chrome

ngrok 账户（免费）：https://ngrok.com — 或使用
```
cloudflared
```
：
```
brew install cloudflared
```
Colab Pro/Pro+，用于 GPU 优先调度和更长的运行时长
已安装且可正常运行的 agent-browser
Chrome 中已登录 Google 账户