tlopex commented on code in PR #20188:
URL: https://github.com/apache/tvm/pull/20188#discussion_r3858858559
##########
python/tvm/relax/frontend/onnx/onnx_frontend.py:
##########
@@ -2173,6 +2173,23 @@ def _impl_v13(cls, bb, inputs, attr, params):
axis = tuple([int(x) for x in axis.data.numpy()])
return cls._squeeze(bb, data, axis)
+ @classmethod
+ def _check_squeeze_axes_are_unit_dims(cls, data, axes):
+ """Raise if any axis to be squeezed has a statically known extent
other than 1."""
+ rank = _get_known_tensor_rank(data)
+ if rank is None:
+ return
+ ty = data.ty
+ if not (isinstance(ty, relax.TensorType) and isinstance(ty.shape,
relax.ShapeExpr)):
+ return
+ for axis in _normalize_constant_axes(list(axes), rank, "Squeeze"):
+ extent = ty.shape.values[axis]
+ if isinstance(extent, tirx.IntImm) and int(extent.value) != 1:
Review Comment:
This only rejects statically known non-unit extents, so the same ONNX
violation remains for symbolic dimensions. For an input declared as `(N, 3)`
with constant axes `[0]`, running with shape `(2, 3)` correctly raises in ONNX
Runtime, while TVM imports it and silently returns a `(3,)` view containing the
first row. Please add a runtime check that selected extents equal 1, or reject
the import when this cannot be proven, and add a symbolic-dimension regression
test. The runtime-axes path needs the same validation.
--
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]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
