PrincetonUniversity / PsyNeuLink / 26937910727

85%

master: 85%

push

github

Add Dask-distributed PEC MLE backend (ask/tell prototype)

Extends PEC maximum-likelihood fitting to run across multiple cluster
nodes via Dask, using an ask/tell driver model (not Optuna shared
storage):

  * Driver owns a single CMA-ES study: it asks a batch of candidate
    parameters, submits them to Dask workers as log-likelihood
    evaluations, gathers the scalars, and tells them back.
  * Workers are pure evaluators: each builds a fresh serial PEC from a
    factory (cached per worker) and returns pec.log_likelihood(...).
    A constructed PEC is never serialized -- only params, the broadcast
    inputs/data, and scalar results cross the wire.
  * Common random numbers (same_seed_for_all_parameter_combinations) for
    a smooth, reproducible likelihood surface across workers.
  * One evaluation x WORKER_CORES LLVM threads per worker, to avoid
    oversubscription.

pec_dask_mle/ package: config, factory (the PEC recipe), worker, data,
driver. Entrypoints: run_dask_mle_local.py (single-node LocalCluster)
and run_dask_mle_jobqueue.py (multi-node dask-jobqueue; ib0 interface,
walltime tuned for Della's short QOS).

Validated: single-node recovery ~7%; multi-node across 4 nodes with all
workers connected over InfiniBand.

11014 of 14376 branches covered (76.61%)

Branch coverage included in aggregate %.

37683 of 43564 relevant lines covered (86.5%)

0.87 hits per line