Conversation
miladm
force-pushed
the
add_ne_size_node
branch
from
aa2dce5 to
e78d7c8
Compare
miladm
force-pushed
the
add_ne_size_node
branch
from
e78d7c8 to
0249d08
Compare
miladm
force-pushed
the
add_ne_size_node
branch
from
0249d08 to
3592f80
Compare
| return dim_node_0->getDynamicValue() != dim_node_1->getDynamicValue() ? 1 : 0; | ||
| } | ||
|
|
||
| std::string SizeNe::ToString() const { return "SizeNe"; } |
There was a problem hiding this comment.
can you make ToString return "aten::ne_size"? The string has to contain the "aten::ne" otherwise you won't be able to dump the ir of it.
|
if possible, can you add some python level unit test? |
|
|
||
| SizeNe::SizeNe(torch::lazy::Value a, torch::lazy::Value b) | ||
| : XlaNode(torch::lazy::OpKind{c10::Symbol::fromQualString("aten::ne")}, | ||
| {a, b}, xla::ShapeUtil::MakeShape(xla::S64, {}), 1) { |
There was a problem hiding this comment.
let' use GetShapeDimensionType, @vanbasten23 when can we update all those s64 in the sizeNode to GetShapeDimensionType?
There was a problem hiding this comment.
Ack. I'm blocked on #4258 to update all return type: I could've done the casting to unblock me but I'd really like to understand where the s32/s64 mismatch comes from before blindly do the casting. I left a comment. Can you take another look @JackCaoG ?
On another thought, I'm not technically blocked by the above PR to update all those s64 in the sizeNode to use GetShapeDimensionType , let me do that today. Thanks for the reminder.
3 participants
No description provided.