feat: sketch for a Julia template and MWE

[jpbrodrick89]

@dionhaefner I know you're working on your own version of this and I definitely don't want to step on any toes, but I've been playing around with this a bit in the background and thought it better to have in the open to work out what direction we want to go in (both/and is potentially a possibility too). I'm not a Julia expert by any means and this probably needs some more cycles if we go with it, but this seems to work when packages support Enzyme (not all do). Happy to hear how you want play it.

I tried to give an example using a simple linear solver with verifiable gradients that could be useful in a Jax setting as no Jax packages expose modified Cholesky.

[codecov]

Codecov Report

❌ Patch coverage is 20.87912% with 72 lines in your changes missing coverage. Please review.
✅ Project coverage is 69.63%. Comparing base (ea2dbcc) to head (328a661).
⚠️ Report is 9 commits behind head on main.

Files with missing lines	Patch %	Lines
...esseract_core/sdk/templates/julia/tesseract_api.py	11.68%	68 Missing ⚠️
tesseract_core/sdk/engine.py	60.00%	0 Missing and 2 partials ⚠️
...sseract_core/runtime/testing/finite_differences.py	75.00%	0 Missing and 1 partial ⚠️
...seract_core/sdk/templates/julia/julia/Project.toml	75.00%	1 Missing ⚠️

❗ There is a different number of reports uploaded between BASE (ea2dbcc) and HEAD (328a661). Click for more details.

HEAD has 2 uploads less than BASE

Flag	BASE (ea2dbcc)	HEAD (328a661)
	33	31

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #552      +/-   ##
==========================================
- Coverage   77.13%   69.63%   -7.50%     
==========================================
  Files          32       35       +3     
  Lines        4356     4499     +143     
  Branches      717      745      +28     
==========================================
- Hits         3360     3133     -227     
- Misses        712     1082     +370     
  Partials      284      284              

☔ View full report in Codecov by Sentry.
📢 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.

[PasteurBot]

Benchmark Results

Benchmarks use a no-op Tesseract to measure pure framework overhead.

🚀 0 faster, ⚠️ 0 slower, ✅ 36 unchanged

✅ No significant performance changes detected.

Full results

Benchmark	Baseline	Current	Change	Status
api/apply_1,000	0.495ms	0.495ms	+0.1%	✅
api/apply_100,000	0.498ms	0.498ms	+0.1%	✅
api/apply_10,000,000	0.497ms	0.499ms	+0.4%	✅
cli/apply_1,000	1650.711ms	1657.890ms	+0.4%	✅
cli/apply_100,000	1653.257ms	1645.324ms	-0.5%	✅
cli/apply_10,000,000	1682.491ms	1684.052ms	+0.1%	✅
decoding/base64_1,000	0.029ms	0.029ms	+0.3%	✅
decoding/base64_100,000	0.970ms	0.970ms	+0.0%	✅
decoding/base64_10,000,000	106.318ms	106.556ms	+0.2%	✅
decoding/binref_1,000	0.178ms	0.177ms	-0.5%	✅
decoding/binref_100,000	0.227ms	0.233ms	+2.9%	✅
decoding/binref_10,000,000	8.648ms	8.562ms	-1.0%	✅
decoding/json_1,000	0.100ms	0.100ms	+0.2%	✅
decoding/json_100,000	9.490ms	9.601ms	+1.2%	✅
decoding/json_10,000,000	1119.859ms	1117.219ms	-0.2%	✅
encoding/base64_1,000	0.028ms	0.028ms	+0.1%	✅
encoding/base64_100,000	0.144ms	0.145ms	+0.3%	✅
encoding/base64_10,000,000	22.061ms	21.778ms	-1.3%	✅
encoding/binref_1,000	0.263ms	0.262ms	-0.1%	✅
encoding/binref_100,000	0.430ms	0.427ms	-0.8%	✅
encoding/binref_10,000,000	16.496ms	16.755ms	+1.6%	✅
encoding/json_1,000	0.144ms	0.144ms	+0.4%	✅
encoding/json_100,000	14.514ms	14.551ms	+0.3%	✅
encoding/json_10,000,000	1514.957ms	1512.533ms	-0.2%	✅
http/apply_1,000	2.912ms	3.099ms	+6.4%	✅
http/apply_100,000	8.007ms	8.276ms	+3.4%	✅
http/apply_10,000,000	663.527ms	666.111ms	+0.4%	✅
roundtrip/base64_1,000	0.059ms	0.059ms	+0.7%	✅
roundtrip/base64_100,000	1.126ms	1.123ms	-0.2%	✅
roundtrip/base64_10,000,000	128.577ms	129.030ms	+0.4%	✅
roundtrip/binref_1,000	0.454ms	0.453ms	-0.2%	✅
roundtrip/binref_100,000	0.667ms	0.664ms	-0.5%	✅
roundtrip/binref_10,000,000	25.671ms	25.378ms	-1.1%	✅
roundtrip/json_1,000	0.253ms	0.252ms	-0.5%	✅
roundtrip/json_100,000	21.975ms	21.197ms	-3.5%	✅
roundtrip/json_10,000,000	2630.813ms	2636.945ms	+0.2%	✅

• Runner: Linux 6.17.0-1008-azure x86_64

[dionhaefner]

Thanks @jpbrodrick89, this looks really comprehensive. I wonder what an appropriate path to getting this merged would look like. Maybe someone who knows Julia well can give this a spin and report back? Perhaps @llueg ?

[dionhaefner]

IMO this should be a vectoradd-like example. This way we have a nice symmetry: vectoradd vanilla, vectoradd-jax, vectoradd-julia, ...

Alternatively univariate if we want to cover more edge cases.

[dionhaefner]

Why those changes?

[jpbrodrick89]

I import expand_path_pattern to separate diff and nondiff arguments, it recurses through a tree into optional arguments but the pattern always includes the the attribute name of the optional argument. If the optional argument is itself a container or a model then it will try to recurse further e.g. by calling None.sub (for TridiagBlock | None) and then error.

[dionhaefner]

Sounds like this should be an independent fix?

[jpbrodrick89]

thanks @dionhaefner , @llueg would be great to collaborate, I have some changes locally I need to tidy up. Basically I think we aim to have three examples

1. Fortran with Enzyme
2. Julia with Enzyme
3. Julia with SciMLSensitivities (locally I got this working for jvp as well as vjp)

we can reuse some boilerplate between 1 and 2 I think. I was going to deprioritise this for the next couple of weeks but if you're keen to dig into this as (re-)onboarding @llueg I'll tidy up this week.

[llueg]

This sounds very cool! Happy to try this out - doesn't necessarily fit in my onboarding project but I can make time whenever you think it's ready for review/testing. I should add a disclaimer that I wouldn't consider myself a Julia expert 😄