transformers recently updated usage of SiLU to the SiLUActivation class. When upgrading transformers this will cause backwards compatibility issues with existing recipes.
This PR proposes a fix by adding class alias support in quantization modifier name matching. Any reference to SiLU in a recipe will now also check for SiLUActivation
transformers recently updated usage of
SiLU
to theSiLUActivation
class. When upgrading transformers this will cause backwards compatibility issues with existing recipes.This PR proposes a fix by adding class alias support in quantization modifier name matching. Any reference to
SiLU
in a recipe will now also check forSiLUActivation
test_plan: @dbogunowicz to revert test patches in https://github.com/neuralmagic/sparseml/pull/2204 to test this PR