chore: prepare release

[janfb]

• draft changelog
• bump version

TODO

• GPU tests passing
• slow tests passing
• lock and env files updated.

[codecov]

Codecov Report

❌ Patch coverage is 50.00000% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 87.73%. Comparing base (c8fac87) to head (951d9b1).
⚠️ Report is 5 commits behind head on main.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
sbi/inference/potentials/vector_field_adaptor.py	40.00%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1718      +/-   ##
==========================================
- Coverage   88.01%   87.73%   -0.28%     
==========================================
  Files         140      140              
  Lines       12791    13258     +467     
==========================================
+ Hits        11258    11632     +374     
- Misses       1533     1626      +93     

Flag	Coverage Δ
fast	82.75% <50.00%> (?)

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

Files with missing lines	Coverage Δ
sbi/__version__.py	100.00% <100.00%> (ø)
sbi/utils/torchutils.py	68.11% <ø> (ø)
sbi/inference/potentials/vector_field_adaptor.py	88.72% <40.00%> (-0.62%)	⬇️

... and 1 file with indirect coverage changes

[manuelgloeckler]

Hey @janfb ,

given due to the arviz release problem i.e. #1816. We should do the release soonish.

I think the only PR that we planned to include is #1752 (I can also have a closer look into that). The only larger bug fix #1803 will be merged soon.

Let me know if I can help on something i.e. running slow tests or so.

[janfb]

good point @manuelgloeckler !

Yes, running slow tests and gpu tests on your end would be great!

We should also include the

• Re-enable time-dependent z-scoring for Flow Matching #1752 (will finish this soon) in the release
• an update on the FMPE / NPSE tutorials / how-to-guide (working on it)
• update the Changelog with all the new PRs and new contributors (will do it)

[manuelgloeckler]

Hey, all right. Thanks!

I will do the GPU test + slow test. And post here.

[manuelgloeckler]

For the slow tests locally on a Mac, they pass but I think we should remove some XFail:

• LC2ST Module Refactoring #1727 is still open so I guess this is fine (not sure how far you came there)
• test_fmpe_shifted_data_c2st[None-False] - No z-scoring fails should be fixed by Add back z-scoreing for flow matching #1623 , no?

SKIPPED [2] tests/lc2st_test.py:252: flaky due to evaluation error, will be fixed in #1727
XFAIL tests/inference_with_NaN_simulator_test.py::test_inference_with_nan_simulator[NLE_A-0.05]
XFAIL tests/inference_with_NaN_simulator_test.py::test_inference_with_nan_simulator[NRE_B-0.05]
XFAIL tests/embedding_net_test.py::test_npe_with_with_iid_embedding_varying_num_trials - Padding with NaNs causes error in new NaN check on x_o, see #1701, #1717
XFAIL tests/ensemble_test.py::test_c2st_posterior_ensemble_on_linearGaussian[NPE_C-5]
XFAIL tests/linearGaussian_vector_field_test.py::test_fmpe_shifted_data_c2st[None-False] - No z-scoring fails

[manuelgloeckler]

There are two failures on GPU tests:

FAILED tests/inference_on_device_test.py::test_multiround_mdn_training_on_device[cpu-NPE_A] - ValueError: Posterior precision matrix is not positive definite. This is a known issue with SNPE-A when the proposal and density estimator don't align well. Try di...
FAILED tests/inference_on_device_test.py::test_multiround_mdn_training_on_device[gpu-NPE_A] - ValueError: Posterior precision matrix is not positive definite. This is a known issue with SNPE-A when the proposal and density estimator don't align well. Try di...

Which are however not related to the device nut is probably from the SNPE-A refactoring that changed the numerics.

I am not entirely sure why or where, but the device tests specifically inference_on_device_test.py are pretty slow (or got stuck for me). In this run only around 80% finished.

[janfb]

For the slow tests locally on a Mac, they pass but I think we should remove some XFail:

• LC2ST Module Refactoring #1727 is still open so I guess this is fine (not sure how far you came there)
• test_fmpe_shifted_data_c2st[None-False] - No z-scoring fails should be fixed by Add back z-scoreing for flow matching #1623 , no?

SKIPPED [2] tests/lc2st_test.py:252: flaky due to evaluation error, will be fixed in #1727
XFAIL tests/inference_with_NaN_simulator_test.py::test_inference_with_nan_simulator[NLE_A-0.05]
XFAIL tests/inference_with_NaN_simulator_test.py::test_inference_with_nan_simulator[NRE_B-0.05]
XFAIL tests/embedding_net_test.py::test_npe_with_with_iid_embedding_varying_num_trials - Padding with NaNs causes error in new NaN check on x_o, see #1701, #1717
XFAIL tests/ensemble_test.py::test_c2st_posterior_ensemble_on_linearGaussian[NPE_C-5]
XFAIL tests/linearGaussian_vector_field_test.py::test_fmpe_shifted_data_c2st[None-False] - No z-scoring fails

• LC2ST: Is waiting for a review, but let's do this next release.
• no z-scoring: this is on purpose, it should fail without z-scoring because data is shifted.

[janfb]

tests/inference_on_device_test.py::test_multiround_mdn_training_on_device

this is because num_sims and max_num_epochs were too small so that NPE-A couldn't train and the analytical correction is off. I pushed a fix.

[janfb]

Regarding the slow GPU tests: can you run it with a time out or with --durations=20 or so to check which tests are so slow? @manuelgloeckler

Update: there is a pytest-timeout plugin; I am running this now on my CUDA machine to find the hanging tests.

With a time out of 300sec I get:

XFAIL tests/inference_on_device_test.py::test_boxuniform_device_handling[cpu-gpu]
FAILED tests/inference_on_device_test.py::test_to_method_on_posteriors[direct-cpu] - Failed: Timeout (>300.0s) from pytest-timeout.
FAILED tests/inference_on_device_test.py::test_vector_field_methods_device_handling[NPSE-2-cpu-cpu] - Failed: Timeout (>300.0s) from pytest-timeout.
FAILED tests/inference_on_device_test.py::test_vector_field_methods_device_handling[NPSE-2-cpu-gpu] - Failed: Timeout (>300.0s) from pytest-timeout.
FAILED tests/mnle_test.py::test_mnle_on_device[cpu] - TypeError: MCMCPosterior.sample() got an unexpected keyword argument 'mcmc_method'
FAILED tests/mnle_test.py::test_mnle_on_device[gpu] - TypeError: MCMCPosterior.sample() got an unexpected keyword argument 'mcmc_method'

[janfb]

@manuelgloeckler the issues should be fixed now. can you test again to double check?

[janfb]

@manuelgloeckler the issues should be fixed now. can you test again to double check?

Done, all GPU tests are passing and are fast.

[janfb]

CI will fail until #1827 is merged into main and this branch is rebased.

Slow and GPU tests are passing ✅

What is missing is:

• fully updated Changelog
• FMPE / NPSE tutorial

[manuelgloeckler]

Thanks @janfb ! I updated the Changelog and should be up to date now (atleast to my knowledge).

So I think the only part missing is the notebooks?

[janfb]

Thanks @janfb ! I updated the Changelog and should be up to date now (atleast to my knowledge).

So I think the only part missing is the notebooks?

Yes, the VE notebooks and the notebook tests. I plan to do the VE notebooks today.

For the notebook tests, I will now run CD on this branch.

Can you review #1758 ? @manuelgloeckler

[manuelgloeckler]

Alright #1758 is merged, now. Let me know if I should have a look on the notebook.

[janfb]

Alright #1758 is merged, now. Let me know if I should have a look on the notebook.

Great! Just pushed it, #1830.

[manuelgloeckler]

Great, made a review. Looks pretty much good to go!