Add support for nightly version (1.13)

[yebai]

TODO

CI Summary — GitHub Actions

Documentation Preview

Mooncake.jl documentation for PR #856 is available at:
https://chalk-lab.github.io/Mooncake.jl/previews/PR856/

Performance

Performance Ratio:
Ratio of time to compute gradient and time to compute function.
Warning: results are very approximate! See here for more context.

┌────────────────────────────┬──────────┬──────────┬─────────────┬─────────┬─────────────┬────────┐
│                      Label │   Primal │ Mooncake │ MooncakeFwd │  Zygote │ ReverseDiff │ Enzyme │
│                     String │   String │   String │      String │  String │      String │ String │
├────────────────────────────┼──────────┼──────────┼─────────────┼─────────┼─────────────┼────────┤
│                   sum_1000 │ 130.0 ns │     1.85 │        2.32 │     1.0 │        5.01 │   1.85 │
│                  _sum_1000 │  1.07 μs │     6.55 │        1.02 │  5660.0 │        31.5 │   1.09 │
│               sum_sin_1000 │  7.45 μs │     2.42 │        1.19 │    1.73 │        9.29 │   1.81 │
│              _sum_sin_1000 │  5.17 μs │     3.24 │        2.44 │   355.0 │        13.2 │   2.76 │
│                   kron_sum │ 192.0 μs │     12.7 │        3.54 │    7.72 │       441.0 │   19.3 │
│              kron_view_sum │ 257.0 μs │     12.7 │        4.98 │    17.4 │       470.0 │   15.4 │
│      naive_map_sin_cos_exp │  2.34 μs │     2.26 │        1.44 │ missing │        6.95 │   2.04 │
│            map_sin_cos_exp │  2.29 μs │      2.8 │        1.52 │    2.62 │         6.0 │   2.59 │
│      broadcast_sin_cos_exp │  2.43 μs │     2.57 │        1.44 │    5.47 │        1.38 │   1.97 │
│                 simple_mlp │ 352.0 μs │     5.02 │        2.92 │    1.65 │        8.86 │   3.12 │
│                     gp_lml │ 399.0 μs │      5.9 │        1.82 │    2.59 │     missing │    2.6 │
│ turing_broadcast_benchmark │  2.04 ms │     4.03 │        3.35 │ missing │        36.5 │    1.8 │
│         large_single_block │ 450.0 ns │     5.83 │        1.96 │  3780.0 │        28.4 │   2.18 │
└────────────────────────────┴──────────┴──────────┴─────────────┴─────────┴─────────────┴────────┘

[github-actions]

Mooncake.jl documentation for PR #856 is available at:
https://chalk-lab.github.io/Mooncake.jl/previews/PR856/

[github-actions]

Performance Ratio:
Ratio of time to compute gradient and time to compute function.
Warning: results are very approximate! See here for more context.

┌────────────────────────────┬──────────┬──────────┬─────────────┬─────────┬─────────────┬────────┐
│                      Label │   Primal │ Mooncake │ MooncakeFwd │  Zygote │ ReverseDiff │ Enzyme │
│                     String │   String │   String │      String │  String │      String │ String │
├────────────────────────────┼──────────┼──────────┼─────────────┼─────────┼─────────────┼────────┤
│                   sum_1000 │ 266.0 ns │     1.75 │        1.89 │    0.44 │        1.52 │   3.26 │
│                  _sum_1000 │  1.21 μs │     8.06 │        1.03 │  1600.0 │        23.0 │   1.42 │
│               sum_sin_1000 │  6.56 μs │     3.24 │        1.88 │    2.15 │        10.1 │   1.95 │
│              _sum_sin_1000 │  6.67 μs │     2.85 │        1.67 │   243.0 │        10.1 │   1.77 │
│                   kron_sum │ 389.0 μs │     7.37 │        2.59 │    15.6 │       287.0 │   12.9 │
│              kron_view_sum │ 458.0 μs │     7.21 │        3.36 │    15.2 │       254.0 │   11.1 │
│      naive_map_sin_cos_exp │  2.84 μs │     3.77 │         1.3 │ missing │        5.24 │   1.77 │
│            map_sin_cos_exp │  3.01 μs │     4.38 │        2.71 │    1.85 │        4.19 │   2.07 │
│      broadcast_sin_cos_exp │  3.01 μs │     2.21 │        2.05 │    3.11 │         1.1 │    1.7 │
│                 simple_mlp │ 477.0 μs │      4.0 │        2.58 │    1.59 │        6.84 │    2.9 │
│                     gp_lml │ 306.0 μs │     7.29 │        2.31 │    4.75 │     missing │    4.2 │
│ turing_broadcast_benchmark │   1.9 ms │     4.37 │        3.04 │ missing │        27.7 │   1.91 │
│         large_single_block │ 484.0 ns │     6.32 │        2.04 │  3800.0 │        23.0 │   1.86 │
└────────────────────────────┴──────────┴──────────┴─────────────┴─────────┴─────────────┴────────┘

[Technici4n]

I don't think that nightly is worth it, but we should add pre once it's available.

[penelopeysm]

I agree, nightly is quite fragile. BTW, current nightly is 1.14.

[codecov]

Codecov Report

❌ Patch coverage is 58.33333% with 5 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
ext/MooncakeAllocCheckExt.jl	50.00%	3 Missing ⚠️
src/interpreter/ir_utils.jl	66.66%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!

[Technici4n]

Note that pre is currently julia 1.12. Do we know when the first 1.13 pre will be released?

[penelopeysm]

Yeahhh, it's kind of in an awkward spot, that's why I haven't added 1.13 tests for the Turing repos. I'm not actually sure how to specify 1.13 in CI. Maybe version: '1.13-nightly' might work (following from https://github.com/julia-actions/setup-julia?tab=readme-ov-file#julia-versions) but I've never tried it.

1.13.0-alpha1 was supposed to be released last week, but it didn't build: https://julialang.slack.com/archives/C7E3LFG82/p1763643474697259

[penelopeysm]

1.13-alpha2 is released, so might be worth checking whether pre resolves to 1.13

[MasonProtter]

Frustratingly, we can't even start a CI run for 1.13 because JET won't install. See aviatesk/JET.jl#796

We should perhaps factor the JET tests out into a separate test suite that only runs if JET is installable.

[yebai]

The test suite doesn't hard-depend on JET; if JET isn't installed, it will skip JET tests. This means one can temporarily remove JET from the test environment.

Julia 1.13 appears to be less disruptive than 1.12, since some tests (e.g., linear_algebra) pass without change. This indicates the core compiler part may be mostly fine.