[NPUW] LazyTensor refactoring #27108
Conversation
smirnov-alexey
added
category: NPU
| @@ -193,18 +211,10 @@ ov::Tensor LazyTensorImpl::eval() const { | |||
| const auto& shape = std::get<3>(unpack_meta); | |||
| const auto& type = std::get<4>(unpack_meta); | |||
There was a problem hiding this comment.
it seems like variable1, variable2, ... = std::tie(unpack_meta) will be useful here
There was a problem hiding this comment.
Done, thanks!
| LazyTensor permute(const std::vector<std::size_t>& axes); | ||
| LazyTensor convert(); |
There was a problem hiding this comment.
Here's the question, where the operations can be defined.
Things like .permute() and .tof16() can stay non-static members - they will use this as an operand to construct a new LT.
Things like .unpack(w,z,s) or .concat([...]) you may want to make static.
There was a problem hiding this comment.
There are constructors above for Concat, Unpack, etc. Do you want to change them to new_tensor = LazyTensor::concat(to_concat, axis), etc?
|
Any updates here @smirnov-alexey ? |
|
Note: there is a problem with shapes (or it's a side-effect) after the last commit. Debug WIP |
…into as/npuw_lazy_tensor_refactor
smirnov-alexey
force-pushed
the
as/npuw_lazy_tensor_refactor
branch
from
3d94b6d to
6421c75
Compare
Fixed |
| struct Tensor { | ||
| ov::Tensor tensor; | ||
| }; |
There was a problem hiding this comment.
Where this one comes from? Isn't Const enough for the purpose?
There was a problem hiding this comment.
E.g. with Unpack we introduce new input tensors
No description provided.