Skip to content

setup_detail.md

Latest commit

 

History

History
141 lines (101 loc) · 4.74 KB

setup_detail.md

File metadata and controls

141 lines (101 loc) · 4.74 KB

AlphaDiana Setup Details

For detailed manual deployment, ROCK port detection, Redis/Ray/admin/proxy startup order, OpenClaw deployment, and FAQ, refer to the root README.md.

Quick entry:

bash scripts/quickstart.sh

If you need to run manually, pay attention to the following:

  • Run source scripts/rock_env.sh from the repository root directory
  • Run unset TMPDIR before running python scripts/find_rock_ports.py --write-env scripts/.rock_ports.env
  • After newgrp docker, you must re-run conda activate, source scripts/rock_env.sh, and source scripts/.rock_ports.env
  • Run ray stop before starting Ray
  • RAY_TMPDIR defaults to /tmp/ray

Further manual steps from README.md can be migrated here over time.

ROCK Proxy Timeout Configuration

The ROCK proxy post_proxy timeout has been changed from a hardcoded 120s to reading from ProxyServiceConfig.timeout (default 180s). To set a larger value (600s recommended), configure it in the ROCK YAML config:

# rock-config.yml (specified via the ROCK_CONFIG environment variable)
proxy_service:
  timeout: 600

If your ROCK version still has timeout=120 hardcoded, you need to modify it manually:

  1. Open rock/sandbox/service/sandbox_proxy_service.py in the ROCK installation directory
  2. Find timeout=120 (around line 454)
  3. Change it to timeout=self.proxy_config.timeout

FAQs

Common issues at a glance:

Issue Cause Fix
permission denied on Docker socket Current shell lacks docker group Run newgrp docker or re-login
deploy.py sandbox timeout Redis not running or .venv symlink broken Check docker ps and ls -la ref/ROCK/.venv
Ray port 8265 already in use Shared server port conflict Use find_rock_ports.py to detect free ports
rock.admin.main Redis ConnectionError Redis container not started docker start "$ROCK_REDIS_CONTAINER"
Sandbox container exits immediately ref/ROCK/.venv missing or invalid ln -sfn "$(python -c 'import sys; print(sys.prefix)')" ref/ROCK/.venv

Port Cleanup

On shared servers, you may need to clean up previously allocated ports:

bash scripts/cleanup_rock_ports.sh

This only terminates processes owned by the current user on ROCK-related ports.

CLI Reference

alphadiana run <config.yaml>                # Run evaluation
alphadiana validate <config.yaml>           # Validate config without running
alphadiana report <results_dir>             # Generate report from result files
alphadiana batch <c1.yaml> <c2.yaml> ...    # Run multiple configs (supports --parallel)
alphadiana list-benchmarks                  # List registered benchmarks

Override config values from CLI:

alphadiana run config.yaml \
  --override agent.config.temperature=0.5 \
  --override max_concurrent=4

Re-run failed tasks:

alphadiana run config.yaml --redo-all

Configuration

Experiments are defined by a single YAML file. See configs/schema.yaml for the full schema.

run_id: "openclaw-qwen3-8b-aime2024-001"   # auto-generated if omitted

agent:
  name: openclaw                  # openclaw | direct_llm
  version: "2026.3.7"
  config:
    # Supports environment variables: ${SANDBOX_ID}, ${ROCK_PORT}
    api_base: "http://127.0.0.1:9001/apis/envs/sandbox/v1/sandboxes/${SANDBOX_ID}/proxy/v1"
    model: openclaw
    gateway_token: "OPENCLAW"
    max_tokens: 65536             # recommend 65536+ for thinking models
    max_attempts: 5               # retry attempts (openclaw only)
    request_timeout: 1800         # seconds (openclaw only)

benchmark:
  name: aime
  config:
    dataset: "HuggingFaceH4/aime_2024"
    split: "train"

sandbox: null                     # null | rock | local

scorer:
  name: math_verify
  config:
    tolerance: 1e-6

max_concurrent: 1                 # parallel task count
num_samples: 32                   # samples per task (for pass@k)
output_dir: "./results"
metadata:                         # free-form tags (optional)
  author: "team-xyz"

Example configs: configs/examples/

API Key Handling in Dashboard

The Dashboard API Key input supports two modes:

  1. Direct paste: Paste the API key directly
  2. Environment variable reference: Enter $VAR_NAME (e.g., $OPENROUTER_API_KEY) to read from .env or system environment

The Dashboard auto-matches API key variables based on the API base URL domain:

API Base URL Auto-matched Variable
https://openrouter.ai/api/v1/ $OPENROUTER_API_KEY
https://api.openai.com/v1/ $OPENAI_API_KEY
https://api.siliconflow.cn/v1/ $SILICONFLOW_API_KEY
https://ark.cn-beijing.volces.com/api/... $ARK_API_KEY
https://api.deepseek.com/v1/ $DEEPSEEK_API_KEY