masahi commented on a change in pull request #8105:
URL: https://github.com/apache/tvm/pull/8105#discussion_r641011555
##########
File path: include/tvm/relay/attrs/transform.h
##########
@@ -146,11 +146,18 @@ struct GatherAttrs : public tvm::AttrsNode<GatherAttrs> {
struct GatherNDAttrs : public tvm::AttrsNode<GatherNDAttrs> {
Integer batch_dims;
+ Integer num_indices_per_tuple;
TVM_DECLARE_ATTRS(GatherAttrs, "relay.attrs.GatherNDAttrs") {
TVM_ATTR_FIELD(batch_dims).set_default(Integer(0)).describe("The number of
batch dimensions.");
+ TVM_ATTR_FIELD(num_indices_per_tuple)
Review comment:
Yeah I spent sometime thinking about the best name. I like `index_rank`
but the confusion could arise, because it is not clear if this is the rank of
indices tensor (the second argument to this op) or the rank of indexing tuple
(which is what I want here). In the ONNX doc
https://github.com/onnx/onnx/blob/master/docs/Operators.md#GatherND, the former
is `q` and the latter one is `indices_shape[-1]`. I'm looking for a concise and
clear word to describe the latter rank. What do you think?
--
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.
For queries about this service, please contact Infrastructure at:
[email protected]
