Repository navigation

Feature/numba integration - #99

Merged

lenarddome merged 35 commits into

mainfrom

feature/numba-integration

Sep 30, 2026

Merged

Feature/numba integration#99
lenarddome merged 35 commits into
mainfrom
feature/numba-integration

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

Commits

Commits on Sep 24, 2026

Commits on Sep 25, 2026

Commits on Sep 28, 2026

Commits on Sep 29, 2026