• Home
  • Features
  • Pricing
  • Docs
  • Announcements
  • Sign In

PrincetonUniversity / PsyNeuLink / 28132814695
84%
master: 85%

Build:
Build:
LAST BUILD BRANCH: devel
DEFAULT BRANCH: master
Ran 24 Jun 2026 11:34PM UTC
Jobs 1
Files 168
Run time 1min
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst

24 Jun 2026 10:08PM UTC coverage: 83.907% (-0.1%) from 84.029%
28132814695

push

github

web-flow
Distributed (Dask) fitting for ParameterEstimationComposition (#3561)

## Summary

Adds distributed evaluation to `ParameterEstimationComposition` fitting.
Candidate
parameterizations are evaluated in parallel across a Dask cluster
instead of
serially, controlled by a single constructor flag. Serial behavior is
unchanged
when the flag is off, and Dask is imported only when distributed fitting
is
requested.

## User-facing API

`PECOptimizationFunction` and `ParameterEstimationComposition` gain two
arguments:

- `distributed: bool = False` — evaluate candidates across a Dask
cluster.
- `distributed_options: Mapping` — configuration with the following
keys:
- `pec_factory` — required; a picklable callable that takes the observed
data and
    returns a fresh `(pec, inputs)` for a worker to score.
- `worker_cores` — LLVM threads per worker; defaults to
`$SLURM_CPUS_PER_TASK`.
- `max_concurrent_evaluations` — candidates per ask/tell round; defaults
to the
    live worker count.
  - `client` — a Dask client to use instead of an auto-resolved cluster.
  - `n_workers` — size of the auto-created single-node `LocalCluster`.

Both the optuna samplers and `differential_evolution` are supported.
LLVM
execution and a single scalar objective are required.

## Launch models

- Single node: `python study.py` forms a `LocalCluster` automatically.
- Multiple nodes: `srun -n N python -m psyneulink.dask_run study.py`. A
new
launcher module forms the scheduler, driver, and workers from the SLURM
ranks,
so the study script contains no scheduler addresses or cluster setup.
The worker
  count is `N - 2`.

A worked DDM example and an SBATCH template are under
`Scripts/Debug/pec_dask/`.

## How it works

The driver owns the optimizer. Each ask/tell round dispatches a batch of
candidates to workers; a worker rebuilds one PEC locally from
`pec_factory`,
caches it, and returns `float(pec.log_likelihood(...))`. A live PEC is
never
shipped across the wire, since Compositions are not... (continued)

11056 of 14436 branches covered (76.59%)

Branch coverage included in aggregate %.

158 of 272 new or added lines in 3 files covered. (58.09%)

1 existing line in 1 file now uncovered.

37804 of 43795 relevant lines covered (86.32%)

0.86 hits per line

Uncovered Changes

Lines Coverage ∆ File
78
81.41
-10.02% psyneulink/core/components/functions/nonstateful/fitfunctions.py
34
0.0
psyneulink/dask_run.py
2
86.92
2.5% psyneulink/core/compositions/parameterestimationcomposition.py

Coverage Regressions

Lines Coverage ∆ File
1
78.69
-1.64% psyneulink/library/components/projections/pathway/autoassociativeprojection.py
Jobs
ID Job ID Ran Files Coverage
1 28132814695.1 24 Jun 2026 11:34PM UTC 168
83.91
GitHub Action Run
Source Files on build 28132814695
  • Tree
  • List 168
  • Changed 3
  • Source Changed 0
  • Coverage Changed 3
Coverage ∆ File Lines Relevant Covered Missed Hits/Line Branch Hits Branch Misses
  • Back to Repo
  • 70e6b15e on github
  • Prev Build on devel (#28116648026)
  • Next Build on devel (#28142869122)
STATUS · Troubleshooting · Open an Issue · Sales · Support · CAREERS · ENTERPRISE · START FREE · SCHEDULE DEMO
ANNOUNCEMENTS · TWITTER · TOS & SLA · Supported CI Services · What's a CI service? · Automated Testing

© 2026 Coveralls, Inc