Repository navigation
Conversation
|
NNlib should be registered in a few minutes |
|
The GNN julia 1 job terminated because the resource limits. Locally the TransformerConv took around 7gb to run, plus other tests might have exhausted the resources |
Signed-off-by: Parvm1102 <parvmittal31757@gmail.com>
Signed-off-by: Parvm1102 <parvmittal31757@gmail.com>
`Flux.destructure`'s `re` writes the gradient into the trainable leaves but keeps every other array (e.g. BatchNorm's running stats) at its model value. Zygote returns `nothing` for those fields so `check_equal_leaves` skipped them, but Enzyme returns a full shadow model with the correct zero gradient there and the comparison failed on state that is not a gradient at all. Rebuild the reference from a zeroed model so those leaves compare as zeros. This enables Enzyme on the GNNChain Parallel/BatchNorm test site for non-dense graphs; verified neutral for EdgeConv and GNNlib's propagate. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Parvm1102 <parvmittal31757@gmail.com>
On Julia 1.13.3 (Enzyme 0.13.216-0.13.218) Enzyme fails on these layers: GConvLSTMCell, GConvLSTM, GConvGRUCell and GConvGRU hit "LLVM error: Referring to an instruction in another function!" while linking the derivative, and DCGRUCell/DCGRU hit an IllegalTypeAnalysisException in the keyword GNNGraph constructor that d_conv uses. Test them against Zygote only, as on master, with a comment explaining why. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Parvm1102 <parvmittal31757@gmail.com>
b13abcb to
4d0bac3
Compare
|
Rebased onto the latest master and disabled the 6 tests that failed in the last CI run: GConvLSTMCell, GConvLSTM, GConvGRUCell, GConvGRU: Enzyme fails to compile the gradient with LLVM error: Referring to an instruction in another function!. This started with Julia 1.13.3 and is still there with Enzyme 0.13.218. DCGRUCell, DCGRU: Enzyme throws IllegalTypeAnalysisException in the keyword GNNGraph constructor used inside d_conv. These layers now test against Zygote only. Looking for CI behaviour here. |
|
we can re run the two failing tests. Those were failing due to a HTTP 504 error |
I am creating a draft PR for enabling enzyme tests after the recent fixes on NNlib.jl, Enzyme.jl and this repo.
It is draft because NNlibv 0.9.45 bump commit is there but hasn't released yet on the julia registry. All CI checks will be failing because of that.
Another reason for being draft PR is because of enzyme, the test suite time has increased drastically. It will now take hours to complete running on CI. Most layers add 3-15 minutes of time each. TransformerConv now takes ~45 minutes, GConvLSTM 33 min, GConvGRU 27 min, GCNConv 22 min and so on. Need to discuss how should we proceed.
Current status for Enzyme:
The Dense failures and DConv will only require simple fixes.