Feature/numba integration - #99

Merged

lenarddome merged 35 commits into

mainfrom

feature/numba-integration

Sep 30, 2026

Commits on Sep 24, 2026

perf(generators): share priors between copies, evaluate them in closed form

lenarddome
and
claude
committed
2f0a967
feat(generators): add SessionWrapper, for models that run a whole session per call

lenarddome
and
claude
committed
09db79c
feat: add numba-compiled session versions of the built-in applications

lenarddome
and
claude
committed
4b051fa
perf: speed up cpm.models, per-trial runs, the losses and meta-d'

lenarddome
and
claude
committed
68c2d9f

Commits on Sep 25, 2026

refactor(applications): one implementation per model, same API

lenarddome
and
claude
committed
aadbc63

Commits on Sep 28, 2026

refactor(models): one formula per building block, shared with the compiled loops

lenarddome
and
claude
committed
f7eb422
docs: add the numba install option

lenarddome
and
claude
committed
eaebd8b
docs(changelog): bring the numba entries up to date

lenarddome
and
claude
committed
e145a7c
docs(changelog): shorten the entries of the development version

lenarddome
and
claude
committed
755cd3f
chore: bump the version to 0.26.0.dev0

lenarddome
and
claude
committed
cc7160d
bench: time the hierarchical tutorials; correct the compile times in the docs

lenarddome
and
claude
committed
2d48bad
docs: re-run the hierarchical tutorials

lenarddome
and
claude
committed
9debd33
docs: add a News section, with a post on faster fitting in cpm 0.26

lenarddome
and
claude
committed
6b31a12
fix(hierarchical): seed the starting priors of later chains, within both bounds

lenarddome
and
claude
committed
8c61db8
docs(tutorials): point the hierarchical tutorials to RLRW, and re-run them

lenarddome
and
claude
committed
9d275aa
docs: add a how-to guide on speeding up your own model

lenarddome
and
claude
committed
63ba29d
docs(numba): correct the caching and debugging advice

lenarddome
and
claude
committed
fd56207

Commits on Sep 29, 2026

fix(applications): check indices and activations in the compiled loops

lenarddome
and
claude
committed
832a2ff
fix(generators): pickling and invalid priors, and SessionWrapper and reset fixes

lenarddome
and
claude
committed
03ffcf0
fix(jit): fall back to plain Python or no caching when numba cannot be used

lenarddome
and
claude
committed
ee689cd
chore: fix the wheel tag, Python requirement and classifiers; drop unused imports

lenarddome
and
claude
committed
dcb3b93
ci: run the tests on Python 3.11-3.14, with and without numba

lenarddome
and
claude
committed
18b44bb
docs(changelog): note that Wrapper.reset keeps the current priors

lenarddome
and
claude
committed
a1a3e41
chore(benchmarks): remove results that nothing reads

lenarddome
and
claude
committed
897f97f
perf(applications): check the activations with one call per trial

lenarddome
and
claude
committed
f553901
fix(models): NaN policies and short feedback behave as on main again

lenarddome
and
claude
committed
2980bfd
bench: re-measure the speed-ups, and update them in the News post

lenarddome
and
claude
committed
3f4894a
docs: complete the changelog, and describe the numba fallbacks

lenarddome
and
claude
committed
435da17
Merge origin/main into feature/numba-integration

lenarddome
and
claude
committed
313825f
ci: use actions/setup-python v7 in the tests, as in the other workflows

lenarddome
and
claude
committed
d7bf882
build: require pandas and matplotlib releases that work with NumPy 2

lenarddome
and
claude
committed
2468bae
fix(applications): reject NaN indices on every platform, not only on x86

lenarddome
and
claude
committed
82bbfd0
test: compare to 1e-12 across CPUs, and store the reference without pandas objects

lenarddome
and
claude
committed
68f623b
docs(news): results can differ in the last digit on other machines

lenarddome
and
claude
committed
f9e4367
build: require pybads 1.0.5, the first release that works with NumPy 2

lenarddome
and
claude
committed
040c87b