[ColmTalbot]

Builds on the enable-x64 PR: the first commit (2b9a91a) is that PR, and the single-precision tests here use set_backend("jax", enable_x64=False). Review from 028f525 on; this targets main so CI runs.

In single precision the linear densities underflow below ~1e-38 (~1e-308 in double), and the Monte Carlo weights with them, so ln L becomes -inf far from the bulk. This adds a log-probability path that never leaves log space, and uses it for the likelihood and selection-function sums.

Log-space models (028f525, b485a47)

• utils: ln_powerlaw, ln_powerlaw_from_logs (logs of the abscissa and bounds) and ln_truncnorm. models.mass: ln_two_component_single and ln_two_component_primary_mass_ratio. They return the log density, -inf outside the support.
• ln_powerlaw normalises from log(low) and log(high), so it doesn't overflow for large |alpha|; xlogy keeps x^0 = 1 at x = 0. The mixture uses logaddexp.
• ln_two_component_primary_mass_ratio uses dataset["ln_mass_1"] and dataset["ln_mass_ratio"] when present (mainly helps non-jax backends, as XLA hoists the logs out of a vmap anyway).
• Timing on an A100, 1000 population draws over 266k injections: 14 ms in float64 and 5.6 ms in float32, against 28 ms for log(two_component_primary_mass_ratio) in float64. b485a47 cuts the per-element transcendentals from 7 to 4 (2.57 → 2.17 ms for 500 draws in float32). Float32 matches float64 to 2e-3 in log density.

hyperpe.Model.ln_prob (c89b461)

• gwpopulation.hyperpe.Model subclasses bilby.hyper.model.Model and adds ln_prob, the sum of the component log densities; prob is unchanged.
• For each component the log density is, in order: its ln_prob attribute; ln_<name> from the component's module, if it takes the same arguments; else the log of the component, so existing models keep working. Log values are cached separately under bilby's rules.
• CosmoModel subclasses this Model; its ln_prob subtracts the log Jacobian.

Likelihood and ResamplingVT in log space (3a6c0fa)

• HyperparameterLikelihood and ResamplingVT weight samples by exp(ln_prob - ln_prior), using the model's ln_prob when it has one and log(prob) otherwise, and form the means, variances and P_det with logsumexp.
• Posteriors and injections can give their sampling prior as ln_prior, used in preference to prior.
• With unequal numbers of samples per event, each event is summed separately (utils.segment_logsumexp) instead of by differences of one cumulative sum, which lost precision when events' sums differ greatly (float32 error in ln L on 46 GWTC-4.0 events: 1e-2 → 3e-4). With jax the events are gathered into a padded array, as fast as before on GPU.
• ResamplingVT.total_generated is stored as a float; its square overflowed a 32-bit integer without x64.
• Plain functions passed to HyperparameterLikelihood and the VTs are wrapped in hyperpe.Model, so a matching ln_<name> is picked up.

Changes for existing users

• The likelihood and ResamplingVT now always use the log-space sums, in double precision too. Results agree with the linear formulas to rounding (tested), but this is the default path for everyone.
• truncnorm is now exp(ln_truncnorm), and logsubexp uses log1p(-exp(.)), so values can differ in the last few bits.
• Not converted yet: posterior_predictive_resample and the numpyro helpers still use linear weights.

For CosmoModel in single precision, distances and volumes also need the wcosmo single-precision PR (ColmTalbot/wcosmo, branch float32).

Testing

• Tests compare each log model with the log of the linear one, the log-input path with the plain one, and float32 with float64 (beta = -1, alpha = 0 and a zero lower bound included). Likelihood/VT tests compare per-event ln BF, variances and P_det with the linear formulas, ln_prior with prior, and cover weights of exp(-1000) and events whose weights differ by e^500; logsumexp and segment_logsumexp are checked against scipy on each backend.
• Full suite: 278 passed, 2 skipped, on CPU (numpy and jax) and with jax on an A100 (float64 and the float32 tests). Pre-commit hooks pass.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 94.67456% with 9 lines in your changes missing coverage. Please review.
✅ Project coverage is 95.12%. Comparing base (42c5824) to head (cf78132).

Files with missing lines	Patch %	Lines
gwpopulation/utils.py	91.07%	5 Missing ⚠️
gwpopulation/hyperpe.py	97.40%	2 Missing ⚠️
gwpopulation/models/mass.py	93.75%	1 Missing ⚠️
gwpopulation/vt.py	93.75%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #144      +/-   ##
==========================================
- Coverage   95.25%   95.12%   -0.14%     
==========================================
  Files          14       14              
  Lines         948     1085     +137     
==========================================
+ Hits          903     1032     +129     
- Misses         45       53       +8     

Flag	Coverage Δ
python3.10	95.12% <94.67%> (-0.14%)	⬇️
python3.11	95.12% <94.67%> (-0.14%)	⬇️
python3.12	95.12% <94.67%> (-0.14%)	⬇️
python3.13	95.12% <94.67%> (-0.14%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.