support Azure LLM

[cometta]

does this support Azure endpoint, can show example?

[codelion]

While OpenEvolve doesn't have direct Azure OpenAI client integration, you can use OptiLLM as a proxy to access Azure (and any other provider). OptiLLM provides OpenAI-compatible endpoints that OpenEvolve can use, plus bonus optimization features.

Step 1: Install and Start OptiLLM

# Install OptiLLM
pip install optillm

# Or clone from source
git clone https://github.com/algorithmicsuperintelligence/optillm.git
cd optillm
pip install -r requirements.txt

Step 2: Configure Azure Backend

Set your Azure credentials as environment variables:

export AZURE_OPENAI_API_KEY="your-azure-api-key"
export AZURE_API_VERSION="2024-02-15-preview"
export AZURE_API_ENDPOINT="https://your-resource.openai.azure.com/"
export OPTILLM_API_KEY="optillm"

Step 3: Start OptiLLM Proxy

# Start OptiLLM on port 8000
python optillm.py --port 8000

# Or if installed via pip:
optillm --port 8000

Step 4: Configure OpenEvolve

Create your config.yaml pointing to OptiLLM:

llm:
  # Point to OptiLLM proxy (not directly to Azure)
  api_base: "http://localhost:8000/v1"

  # Use your Azure deployment names
  models:
    - name: "gpt-5"  # Replace with your Azure deployment name
      weight: 0.8
    - name: "gpt-5-nano"  # Replace with your Azure deployment name
      weight: 0.2

  temperature: 0.7
  max_tokens: 4096
  timeout: 300
  retries: 3

Step 5: Run OpenEvolve

export OPENAI_API_KEY="optillm"

python openevolve-run.py \
  path/to/initial_program.py \
  path/to/evaluator.py \
  --config config.yaml \
  --iterations 100

Bonus: OptiLLM Optimization Techniques

OptiLLM enhances code evolution with test-time compute optimizations. Prefix your model names with techniques:

Recommended for Code Evolution:

llm:
  api_base: "http://localhost:8000/v1"

  models:
    # Chain-of-thought reasoning for better logic
    - name: "cot_reflection-gpt-5"
      weight: 0.6

    # Mixture of Agents for quality
    - name: "moa-gpt-5"
      weight: 0.3

    # URL fetching for documentation-heavy tasks
    - name: "readurls-gpt-5"
      weight: 0.1

  timeout: 600  # Higher timeout for MoA
  max_tokens: 12000  # MoA generates multiple completions

Available Techniques:

• moa - Mixture of Agents (generates multiple responses, synthesizes best)
• cot_reflection - Chain-of-thought with self-reflection
• bon - Best of N sampling
• readurls - Auto-fetch web content from URLs in prompts
• executecode - Run code and include output
• rstar - Advanced reasoning for complex problems
• z3_solver - Theorem proving for logical reasoning

Chain techniques with &:

• name: "cot_reflection&readurls-gpt-5" # Both techniques

The same OptiLLM setup works for:

OpenAI

export OPENAI_API_KEY="sk-..."
# OptiLLM auto-detects and uses OpenAI

OpenRouter

export OPENAI_API_KEY="your-openrouter-key"
export OPENAI_API_BASE="https://openrouter.ai/api/v1"

Local Models (Ollama)

export OPENAI_API_BASE="http://localhost:11434/v1"
export OPENAI_API_KEY="ollama"

Any OpenAI-Compatible API

export OPENAI_API_BASE="https://your-api-endpoint/v1"
export OPENAI_API_KEY="your-key"

Complete Working Example

See the web scraper example in the OpenEvolve repo:

• Location: examples/web_scraper_optillm/
• Files: config.yaml, README.md
• Results: Improved accuracy from 72.2% to 83.3% using OptiLLM with local MLX models

[jsfreischuetz]

For OpenAI reasoning and GPT-5 models this is still failing with API errors, likely for the same reason as #300 .

"HTTP/1.1 400 Bad Request"
2025-11-17 16:08:41,459 - WARNING - Error on attempt 1/4: Error code: 400 - {'error': {'message': "Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.", 'type': 'invalid_request_error', 'param': 'max_tokens', 'code': 'unsupported_parameter'}}. Retrying...