Revert "[SLM] Allow modules to define pre-processing of weights" #16777
Conversation
…)" This reverts commit 1cccc3b.
|
@tqchen Can you add a unit test demonstrating the behavior to be retained? Is it only for the string names of symbolic variables, as mentioned in this comment? |
|
Likely i think it would be primarily two parameters with same |
|
Happy to also test followup redos of the PR in case we miss other cases |
|
I replied with a test to the other PR before noticing the on-going thread here. #16757 (comment) @tqchen so could you help pasting it into |
|
Merged without the unit test to keep the reversion standalone. Let's add the test case separately. Thank you folks! |
|
I have a parametrized version of the test case, along with testing for cases beyond |
Follow-up to apache#16777, add unit tests demonstrating desired behavior.
Prior to this commit, the weights used by `nn.Module` instances were required to be `nn.Parameter` instances. This commit allows the weights to instead be `nn.Tensor` instances, defined in terms of other `nn.Parameter` weights. This allows a model to define both the original weights that would be present in an external checkpoint (e.g. a Pytorch or Safetensors file), and the pre-processing that should be performed on those weights. This is a re-implementation of apache#16757, which was reverted in apache#16777. The re-implementation preserves the handling of dynamic shapes specified as python strings, enabling the test cases that were added in apache#16784.
* [SLM] Add unit tests for SLM to Relax exporter Follow-up to #16777, add unit tests demonstrating desired behavior. * Updated docstrings based on review comment
…che#16777) Revert "[SLM] Allow modules to define pre-processing of weights (apache#16757)" This reverts commit 1cccc3b.
* [SLM] Add unit tests for SLM to Relax exporter Follow-up to apache#16777, add unit tests demonstrating desired behavior. * Updated docstrings based on review comment
Reverts #16757