cconvey commented on a change in pull request #8952:
URL: https://github.com/apache/tvm/pull/8952#discussion_r723746756
##########
File path: tests/python/frontend/onnx/test_forward.py
##########
@@ -4941,7 +4941,6 @@ def verify_eyelike(indata):
"test_mvn",
# This test fails llvm with a lowering error:
"test_nllloss_NCd1d2d3_none_no_weight_negative_ii_expanded",
- "test_qlinearmatmul_2D",
"test_qlinearmatmul_3D",
Review comment:
Not sure. It looks like in `test_qlinearmatmul_3D`, both input tensors
are 3D
([link](https://github.com/onnx/onnx/blob/master/onnx/backend/test/case/node/qlinearmatmul.py#L47-L62)).
But IIUC from the comments, `qnn.dense` only allows the second (i.e.,
weight) tensor to be 2D
([link](https://github.com/apache/tvm/blob/main/src/relay/qnn/op/dense.cc#L210)).
So if that comment is accurate, I would guess(?) that either (a) the
importer needs to generate extra ops to make up the difference, or (b) we'd
need to extend `qnn.dense` to handle 3D weight tensors.
Unfortunately I need to finish my work on this feature ASAP, so I can't dig
into it any deeper than this.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
