Any idea on why it doesn’t work with CUDA julia> using Flux, LinearAlgebra julia> x = randn(10, 10); julia> xgpu = gpu(x); julia> gradient(x -> sum(svd(x).S), xgpu) ERROR: MethodError: no method matching svd_rev(::CUDA.CUSOLVER.CuSVD{Float32,Float32,CUDA.CuArray{Float32,2}}, ::ChainRulesCore.Zer…

Gradient of sum of singular values of a matrix with CUDA.jl

maleadt February 16, 2021, 11:52am 2

ChainRules calls svd_rev, which it defines only on Base.SVD while CUDA.jl has its own SVD type.

Topic		Replies	Views
Incorporating SVD factorization into GPUs with CUDA Machine Learning question , cuda , svd	7	864	May 23, 2023
Using Flux to optimize a function of the Singular Values Optimization (Mathematical) differentiation , linearalgebra , flux	6	1574	June 11, 2020
Gradient of ldiv trace on CUDA GPU cuda , linearalgebra , flux , zygote	3	572	March 23, 2022
GPU gradient issues with matrix tranpose GPU flux	0	348	March 14, 2021
CUDA.jl and svd getting automatic casting to Array{Float32,2} General Usage cuda	1	760	May 18, 2021