Can't connect to Gemini API?

[Garfield2013]

1. I have this in the default_config.yaml file + in the examples/signal_processing/config.yaml file:

  api_base: "https://generativelanguage.googleapis.com/v1beta/openai/"
  api_key: "MY_GOOGLE_AI_STUDIO_API_KEY"

2. I have executed this command in Ubuntu terminal:

export OPENAI_API_KEY=MY_GOOGLE_AI_STUDIO_API_KEY

But I still get the following error (Invalid JSON payload received. Unknown name "seed"):

python openevolve-run.py examples/signal_processing/initial_program.py examples/signal_processing/evaluator.py --config examples/signal_processing/config.yaml
2025-06-24 20:52:22,238 - INFO - Logging to examples/signal_processing/openevolve_output/logs/openevolve_20250624_205222.log
2025-06-24 20:52:22,243 - INFO - Set random seed to 42 for reproducibility
2025-06-24 20:52:22,282 - INFO - Initialized OpenAI LLM with model: gemini-2.5-flash-lite-preview-06-17
2025-06-24 20:52:22,304 - INFO - Initialized OpenAI LLM with model: gemini-2.5-flash
2025-06-24 20:52:22,304 - INFO - Initialized LLM ensemble with models: gemini-2.5-flash-lite-preview-06-17 (weight: 0.80), gemini-2.5-flash (weight: 0.20)
2025-06-24 20:52:22,326 - INFO - Initialized OpenAI LLM with model: gemini-2.5-flash-lite-preview-06-17
2025-06-24 20:52:22,349 - INFO - Initialized OpenAI LLM with model: gemini-2.5-flash
2025-06-24 20:52:22,349 - INFO - Initialized LLM ensemble with models: gemini-2.5-flash-lite-preview-06-17 (weight: 0.80), gemini-2.5-flash (weight: 0.20)
2025-06-24 20:52:22,349 - INFO - Initialized prompt sampler
2025-06-24 20:52:22,349 - INFO - Initialized prompt sampler
2025-06-24 20:52:22,349 - INFO - Set custom templates: system=evaluator_system_message, user=None
2025-06-24 20:52:22,350 - INFO - Initialized program database with 0 programs
2025-06-24 20:52:22,786 - INFO - Successfully loaded evaluation function from examples/signal_processing/evaluator.py
2025-06-24 20:52:22,786 - INFO - Initialized evaluator with examples/signal_processing/evaluator.py
2025-06-24 20:52:22,786 - INFO - Initialized OpenEvolve with examples/signal_processing/initial_program.py and examples/signal_processing/evaluator.py
2025-06-24 20:52:22,786 - INFO - Adding initial program to database
2025-06-24 20:52:22,875 - INFO - Evaluated program a91dbe79-2a47-425b-a60c-12224e6cca6e in 0.09s: runs_successfully=1.0000, composite_score=0.4200, output_length=91.0000, overall_score=0.3899, slope_changes=66.0000, lag_error=0.8697, avg_error=1.2418, false_reversals=55.2000, correlation=0.3768, noise_reduction=0.0000, smoothness_score=0.2326, responsiveness_score=0.5349, accuracy_score=0.3768, efficiency_score=1.0000, execution_time=0.0064, success_rate=1.0000
2025-06-24 20:52:22,875 - INFO - Starting evolution from iteration 0 for 100 iterations (total: 100)
2025-06-24 20:52:22,875 - INFO - Using island-based evolution with 4 islands
2025-06-24 20:52:22,875 - INFO - Island Status:
2025-06-24 20:52:22,875 - INFO -  * Island 0: 1 programs, best=13.7281, avg=13.7281, diversity=0.00, gen=0
2025-06-24 20:52:22,875 - INFO -    Island 1: 0 programs, best=0.0000, avg=0.0000, diversity=0.00, gen=0
2025-06-24 20:52:22,875 - INFO -    Island 2: 0 programs, best=0.0000, avg=0.0000, diversity=0.00, gen=0
2025-06-24 20:52:22,875 - INFO -    Island 3: 0 programs, best=0.0000, avg=0.0000, diversity=0.00, gen=0
2025-06-24 20:52:23,462 - INFO - HTTP Request: POST https://generativelanguage.googleapis.com/v1beta/openai/chat/completions "HTTP/1.1 400 Bad Request"
**2025-06-24 20:52:23,463 - WARNING - Error on attempt 1/4: Error code: 400 - [{'error': {'code': 400, 'message': 'Invalid JSON payload received. Unknown name "seed": Cannot find field.', 'status': 'INVALID_ARGUMENT', 'details': [{'@type': 'type.googleapis.com/google.rpc.BadRequest', 'fieldViolations': [{'description': 'Invalid JSON payload received. Unknown name "seed": Cannot find field.'}]}]}}]. Retrying...**

[Garfield2013]

FIXED:

default_config.yaml:

llm:
  #Models for evolution
  models:
    #List of available models with their weights
    - name: "google/gemini-2.5-flash-lite-preview-06-17"
      weight: 0.8
    - name: "google/gemini-2.5-flash-lite-preview-06-17"
      weight: 0.2

  #Models for LLM feedback
  evaluator_models:
    #List of available models with their weights
    - name: "google/gemini-2.5-flash-lite-preview-06-17"
      weight: 0.8
    - name: "google/gemini-2.5-flash-lite-preview-06-17"
      weight: 0.2

  #API configuration
  api_base: "https://openrouter.ai/api/v1/"  # Base URL for API (change for non-OpenAI models)
  api_key: "MY_OPEN_ROUTER_API_KEY"     # API key (defaults to OPENAI_API_KEY env variable)

examples/signal_processing/config.yaml file:

#LLM configuration

llm:
  primary_model: "google/gemini-2.5-flash-lite-preview-06-17"
  primary_model_weight: 0.8
  secondary_model: "google/gemini-2.5-flash-lite-preview-06-17"
  secondary_model_weight: 0.2
  api_base: "https://openrouter.ai/api/v1/"
  api_key: "MY_OPEN_ROUTER_API_KEY"
  temperature: 0.6
  top_p: 0.95
  max_tokens: 32000

[codelion]

I think the gemini's openai compatible endpoint doesn't support the seed param that we added recently for reproducibility. We should fix it to not fail in that case and may be run with a warning.

[pjf67546]

I think the gemini's openai compatible endpoint doesn't support the seed param that we added recently for reproducibility. We should fix it to not fail in that case and may be run with a warning.

I used the Mistral API for the circle packing example in version 0.0.9 and encountered the seed problem too. Version 0.0.8 works fine, and I set the seed to null to fix the issue. Is there a better way to handle this problem? Not just for Gemini, like the change in version 0.0.10, but also because some models support the seed while others do not, even within the same API endpoint. BTW, Gemma3 does not support either the seed or system messages.

Finally, thank you for the great work in reproducing DeepMind's work, as well as for the frequent updates to improve this repo. I really appreciate it!

[codelion]

I think if the LLM API doesn't support a seed param it is best to set it to null. Since we do not control the API provider we will anyways not be able to do much w.r.t reproducibility. For Gemma3, if you are doing inference yourself you can always set some seed in either transformers or vllm or any other inference server. We can default to seed: null in the config in openevolve and only ask users to use it if they want to enable reproducibility.

[pjf67546]

I think if the LLM API doesn't support a seed param it is best to set it to null. Since we do not control the API provider we will anyways not be able to do much w.r.t reproducibility. For Gemma3, if you are doing inference yourself you can always set some seed in either transformers or vllm or any other inference server. We can default to seed: null in the config in openevolve and only ask users to use it if they want to enable reproducibility.

Perhaps the seed can be set globally, and if the LLM is needed, a switch can be turned on for that design? Because in the packing example I mentioned earlier, the code generated by the LLM actually had "randomness," specifically in terms of initialization and global optimization.

The introduction of AlphaEvolve also mentions:
"AlphaEvolve can also propose new approaches to complex mathematical problems. Provided with a minimal code skeleton for a computer program, AlphaEvolve designed many components of a novel gradient-based optimization procedure that discovered multiple new algorithms for matrix multiplication, a fundamental problem in computer science."

The gradient optimization algorithms mentioned here are likely to have randomness as well. So even if the LLM doesn't support a seed, setting a seed may still help stabilize the evaluator.