Lorax.Shape (lorax v0.2.1)
Summary
Functions
Infers the shape of the LoRA matrices. Only supports :dense and :conv operations
Functions
Link to this function
calc_ab(op, r, parameters)
Infers the shape of the LoRA matrices. Only supports :dense and :conv operations
Dense LoRA kernels
Suppose we have a target node W
and input tensor x
,
During injection, the A matrix will project the input tensor down to
r-dimensional space. Afterwards, the B matrix will project the result back up
to some unknown dimensionality. To figure out the output dimensionality,
we inspect W's kernel shape.
Convolution LoRA kernels
In addition to figuring out the output dimensionality, we need to retrieve
the convolution kernel_size of W
. When calling Axon.conv, it's passed as an
option, but then is no longer stored inside the Axon node. To figure out the
kernel size, we also inspect W's kernel shape.