Add PReLU Support for a single-value tensor #14933
Assignees
Labels
Comments
KalaivaniMCW
added a commit
that referenced
this issue
Nov 20, 2024
5 tasks
mouliraj-mcw
added a commit
that referenced
this issue
Nov 21, 2024
### Ticket Link to Github Issue #8544 ### Problem description Currently PReLU is aliased to LeakyRelu which is not correct. It needs to be properly implemented in the eltwise operation. ### What's changed - Update the prelu implementation In composite structure. - It also support sfpu(Added in this PR) for certain cases which will handled soon in #14933 . ### Additional Information In Torch PReLU, the second input tensor can only have two valid shapes: either a tensor with a single value (1) or a tensor with a size equal to the number of input channels (default is 1). Currently, This implementation only supports cases where it matches the number of channels. Support for a single-value tensor requires additional handling at the low-level kernel. ### Checklist - [ ] [All Post commit CI ](https://github.com/tenstorrent/tt-metal/actions/runs/11796920085)
KalaivaniMCW
added a commit
that referenced
this issue
Nov 21, 2024
KalaivaniMCW
added a commit
that referenced
this issue
Nov 21, 2024
KalaivaniMCW
added a commit
that referenced
this issue
Nov 21, 2024
5 tasks
KalaivaniMCW
added a commit
that referenced
this issue
Nov 23, 2024
### Ticket Link to Github Issue #14933 ### Problem description Need Support PRelu to accept single element weight array as input ### What's changed Added support for the same ### Checklist - [x] Post commit CI passes https://github.com/tenstorrent/tt-metal/actions/runs/11960853362 https://github.com/tenstorrent/tt-metal/actions/runs/11987294776 - [ ] Blackhole Post commit (if applicable) - [ ] Model regression CI testing passes (if applicable) - [ ] Device performance regression CI testing passes (if applicable) - [x] New/Existing tests provide coverage for changes
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
In Torch PReLU, the second input tensor i.e
weightcan be passed as:Currently, our implementation only supports cases where it matches the number of channels or weight passed as scalar.
We need support for the case where weight tensor is 1D list/array with length=1
The text was updated successfully, but these errors were encountered: