How to use Float64 with GPU backend

ITensor Julia Questions
1

I need Float64 precision for high-accuracy computations on GPU, but I haven’t been able to find relevant documentation or discussions on how to switch from Float32 to Float64 in ITensor. For reference, here’s a minimal example:

using ITensors, ITensorMPS
using Random: Random
using CUDA
using cuTENSOR

device_id = 0
CUDA.device!(device_id)


function main(; dtype=Float64, device=identity)
    Random.seed!(1234)

    L = 2

    H = OpSum()
    H += 1.0, "Sz", 1, "Sz", 2
    
    sites = siteinds("S=1/2", L)
    H_MPO = device(MPO(dtype, H, sites))
    
    psi_init = device(random_mps(dtype, sites; linkdims=10))
    outputlevel = 0

    nsweeps = 50
    maxdim = [10]
    cutoff = [1E-10]


    E, psi = dmrg(H_MPO, psi_init; nsweeps, maxdim, cutoff, outputlevel=outputlevel)
    
    println("Energy data type: ", typeof(E))
    flush(stdout)

    GC.gc()
    CUDA.reclaim()
end

main(;dtype=Float64, device=cu)

The printed energy data type is Float32, despite explicitly setting the data type of psi_init and H_MPO to Float64. This suggests that the DMRG algorithm is actually operating in Float32. Could someone clarify how to properly configure ITensor/ITensorMPS to use Float64 throughout the computation?

2

cu(x) defined in CUDA.jl automatically converts the element type to single precision (Float32). In order to preserve the element type, you can use adapt(CuArray, x) using the adapt function from Adapt.jl. See Memory management · CUDA.jl (that’s not anything specific to the design of ITensors.jl or ITensorMPS.jl, we are just following the design and conventions of the Julia GPU libraries).

EDIT: A trick you can use is calling your main function as main(; dtype=Float64, device=adapt(CuArray)), adapt(CuArray)(x) is equivalent to adapt(CuArray, x).

3

This works perfectly! Thank you for your reply.