refactor: trigger build with latest Lux #882
Conversation
avik-pal
force-pushed
the
ap/trigger
branch
7 times, most recently
from
0c724f7
to
b25ed28
Compare
|
@ChrisRackauckas I tracked down and fixed the upstream issues. The only one remaining is https://github.com/SciML/NeuralPDE.jl/actions/runs/10485955321/job/29043159135?pr=882#step:6:1216 but that is kind of hard to handle -- it is matrix multiply with non-concrete element types. Might be best to get the |
I can look into it. Can you open an issue about it? Also, in CI, why is precompilation failing - https://github.com/SciML/NeuralPDE.jl/actions/runs/10485955321/job/29043160310?pr=882#step:6:982 |
Depwarn errors are on, and currently Enzyme doesn't support the latest GPUCompiler which doesn't have those depwarns (same as SciML/DeepEquilibriumNetworks.jl#159) |
That's the right thing to do. |
|
@wsmoses what should we do about the GPUCompiler Enzyme depwarn issue? How close is that to getting the release? |
|
Enzyme released a version with the latest gpucompiler last night |
|
@avik-pal you're unblocked here? |
|
No I need #883 to be fixed, else it tries to propagate |
|
Okay, but at least unblocked from non-NeuralPDE things 😅 . @KirillZubov do you think you can have a look at that? |
|
@ChrisRackauckas ok, I will try figure out with it , it's a little related to my current task |
src/discretize.jl
Outdated
| @@ -217,20 +217,20 @@ function generate_training_sets(domains, dx, eqs, bcs, eltypeθ, dict_indvars::D | |||
|
|
|||
| bcs_train_sets = map(bound_args) do bt | |||
| span = map(b -> get(dict_var_span, b, b), bt) | |||
| _set = adapt(eltypeθ, | |||
| _set = adapt(Array{eltypeθ}, | |||
There was a problem hiding this comment.
Do not force an Array type, since that won't work in most scenarios. Just add a post-adapt convert.(eltypeθ, x)
|
I updated this PR also to Lux v1.0 #887 |
|
@sathvikbhagavan @ChrisRackauckas Unfortunately, I couldn't cope with this task and I don't have time to do it in the near future. In addition to updating the packages(Lux to v1 ) that I've already done, It also need to review all the tests and update the solvers which is need some time on it. I fixed the problem with Adapt and Convert types(#883) in a couple of places. This template can be used for other solvers. There are also other errors, the cause of which I haven't had time to establish. |
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
avik-pal
changed the title
|
Lets merge once all the tests pass. I think docs build might fail due to linkchecks but it also takes like 17hrs. I will try to go through and see if we can reduce the time there. |
|
I am going to merge and fix the docs in a followup PR |
No description provided.