Try to reduce overspecialization of annotations#3037
Conversation
Benchmark Results
Benchmark PlotsA plot of the benchmark results has been uploaded as an artifact at https://github.com/EnzymeAD/Enzyme.jl/actions/runs/24255958491/artifacts/6376552547. |
|
FWIW this took my compile trace from 22k lines to a mere 4k. |
|
@vchuravy can you give this a review? |
| TT = Base.unwrap_unionall(TT0) | ||
| ft = TT.parameters[1] | ||
| tt = map(_annotate ∘ Base.Fix2(Base.rewrap_unionall, TT0), TT.parameters[2:end]) | ||
| tt = [] |
There was a problem hiding this comment.
Looks like this won't work with either Type or TypeVar type hint to the vector 😞
There was a problem hiding this comment.
really, aren't all the results of _annotate, types though?
There was a problem hiding this comment.
Some are TypeVar and some are Type apparently
|
|
Yes need a quick patch fix |
|
Any updates here? |
|
Ok so this actually had been patched already. The problem is that I also had to put that hard upper bound on Enzyme because Comrade is broken on the newer versions (see #2873). So some versions are getting mixed up. I'd ignore the Comrade failure for now because it is already broken. |
Seems to have helped #3036 a little bit. Even the
Fix2version was not inheriting the@nospecialize...