Fix quantized_w8a32_gru output shape to preserve batch dimensions #17291
+12
−10
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary:
The meta function for
quantized_w8a32_grureturned shape(2, hidden_dim),dropping the batch dimensions of the hidden state. After the fusion pass replaces
aten.gru.inputwith the custom op,getitem[0]on the output produced shape(hidden_dim,)instead of the original(batch, ..., hidden_dim), causing ashape mismatch with the reference output and NaN in the comparison.
Fix the meta function to return
(2, *hidden.shape)to preserve all dimensions.Also update the reference implementation to handle 3D hidden states and preserve
the original shape in the output, consistent with the shape fix applied to
quantized_w8a32_convin D84745967.Differential Revision: D92570517