[scttfrdmn]

Fixes #554.

Problem

OpenEvolve.__init__ derives one llm_seed and writes it into every model config. Each worker's _lazy_init_worker_components then builds an LLMEnsemble seeded from it. So every worker draws the same sequence of models, and a recycled worker (max_tasks_per_child) restarts that sequence. With N workers, the realized model mix is K draws repeated N times rather than N·K independent draws.

Change

_run_iteration_worker calls _seed_worker_ensemble(iteration) after lazy init. It reseeds the worker's sampler from md5(f"{llm_seed}:{iteration}"), derived the same way llm_seed itself is:

• draws are independent across workers;
• runs stay reproducible, and an iteration's draw doesn't depend on which worker runs it or whether that worker was recycled;
• with random_seed unset, the sampler is left alone (already independent).

The evaluator-model ensemble isn't touched: its only use is generate_all_with_context, which doesn't sample.

Tests

tests/test_worker_model_seeding.py (6 tests) simulates workers as _worker_init builds them:

• draws are identical for 1, 3 or 4 workers and any iteration order;
• workers no longer repeat one sequence;
• realized mix follows weights (0.8 ± 0.06 over 500 iterations);
• a recycled worker continues the sequence;
• an unseeded sampler is untouched;
• _run_iteration_worker seeds with its iteration.

With the reseed made a no-op (the old behaviour), the three behavioural tests fail. Full suite: python -m unittest discover tests, 918 tests, OK (2 skipped). Black-formatted.

Also checked end to end: in openevolve-bedrock, a 40-iteration spawn-mode run with 3 workers against a fake endpoint matches per-model call counts predicted from this derivation exactly. That project currently patches this in from outside, and will drop the patch once a release contains this fix.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}