Using 'expand' with GPU backend

[YotamKa]

Dear ITensorMPS team,

I found that the new 'expand' function (thanks for fixing that, btw) is not working over a GPU backend.

I saw an error in line 90 in 'expand.jl', i.e., projectorⱼ = idⱼ - prime(basisⱼ, rinds) * dag(basisⱼ) , and tried transforming idⱼ to a CuArray; however, the trace function is bugging.

Probably, the fix is straightforward. Thanks, Yotam

Here's a minimal code to reproduce the error:

using ITensors
using ITensorMPS 
using CUDA: cu

device = cu
numtype = Float32

N = 2
sites = [siteind("S=1/2") for _ in 1:N]

mps = device(MPS(sites, ["↑" for _ in 1:N]))

mpo_os = OpSum()
mpo_os += "Sz",1
mpo = device(MPO(mpo_os,sites))

psi_expanded = expand(
  mps,
  mpo;
  alg="global_krylov",
  krylovdim=2,
  cutoff=sqrt(eps(numtype))
)

And here is the error statement:

ERROR: Type parameter position not defined for type `DenseVector{Float32}` and position name `eltype`.
Stacktrace:
  [1] error(s::String)
    @ Base ./error.jl:35
  [2] position(type::Type, pos::Function)
    @ NDTensors.TypeParameterAccessors ~/.julia/packages/NDTensors/vDdB4/src/lib/TypeParameterAccessors/src/position.jl:8
  [3] set_type_parameter(type::Type, pos::Function, param::Type)
    @ NDTensors.TypeParameterAccessors ~/.julia/packages/NDTensors/vDdB4/src/lib/TypeParameterAccessors/src/set_parameters.jl:14
  [4] set_eltype(::Type{SimpleTraits.Not{…}}, type::Type{DenseVector{…}}, param::Type)
    @ NDTensors.TypeParameterAccessors ~/.julia/packages/NDTensors/vDdB4/src/lib/TypeParameterAccessors/src/base/abstractarray.jl:68
  [5] set_eltype
    @ ~/.julia/packages/SimpleTraits/l1ZsK/src/SimpleTraits.jl:331 [inlined]
  [6] similartype(::Type{SimpleTraits.Not{…}}, arraytype::Type{DenseVector{…}}, eltype::Type)
    @ NDTensors ~/.julia/packages/NDTensors/vDdB4/src/abstractarray/similar.jl:109
  [7] similartype
    @ ~/.julia/packages/SimpleTraits/l1ZsK/src/SimpleTraits.jl:331 [inlined]
  [8] promote_rule(::Type{NDTensors.Dense{Float32, Vector{…}}}, ::Type{NDTensors.Dense{Float32, CuArray{…}}})
    @ NDTensors ~/.julia/packages/NDTensors/vDdB4/src/dense/dense.jl:128
  [9] promote_type
    @ ./promotion.jl:313 [inlined]
 [10] promote_rule(::Type{NDTensors.DenseTensor{…}}, ::Type{NDTensors.DenseTensor{…}})
    @ NDTensors ~/.julia/packages/NDTensors/vDdB4/src/tensor/tensor.jl:266
 [11] promote_type
    @ ./promotion.jl:313 [inlined]
 [12] permutedims!!(R::NDTensors.DenseTensor{…}, T::NDTensors.DenseTensor{…}, perm::Tuple{…}, f::Function)
    @ NDTensors ~/.julia/packages/NDTensors/vDdB4/src/dense/densetensor.jl:198
 [13] _map!!(f::Function, R::NDTensors.DenseTensor{…}, T1::NDTensors.DenseTensor{…}, T2::NDTensors.DenseTensor{…})
    @ ITensors ~/.julia/packages/ITensors/4M2ep/src/itensor.jl:1961
 [14] map!(f::Function, R::ITensor, T1::ITensor, T2::ITensor)
    @ ITensors ~/.julia/packages/ITensors/4M2ep/src/itensor.jl:1966
 [15] copyto!
    @ ~/.julia/packages/ITensors/4M2ep/src/broadcast.jl:330 [inlined]
 [16] materialize!
    @ ./broadcast.jl:914 [inlined]
 [17] materialize!
    @ ./broadcast.jl:911 [inlined]
 [18] -(A::ITensor, B::ITensor)
    @ ITensors ~/.julia/packages/ITensors/4M2ep/src/itensor.jl:1883
 [19] expand(::Algorithm{:orthogonalize, @NamedTuple{}}, state::MPS, references::Vector{MPS}; cutoff::Float32)
    @ ITensorTDVP ~/.julia/packages/ITensorTDVP/YqF4o/src/expand.jl:90
 [20] expand
    @ ~/.julia/packages/ITensorTDVP/YqF4o/src/expand.jl:73 [inlined]
 [21] expand(state::MPS, reference::Vector{MPS}; alg::String, kwargs::@Kwargs{cutoff::Float32})
    @ ITensorTDVP ~/.julia/packages/ITensorTDVP/YqF4o/src/expand.jl:29
 [22] expand
    @ ~/.julia/packages/ITensorTDVP/YqF4o/src/expand.jl:28 [inlined]
 [23] expand(::Algorithm{…}, state::MPS, operator::MPO; krylovdim::Int64, cutoff::Float32, apply_kwargs::@NamedTuple{…})
    @ ITensorTDVP ~/.julia/packages/ITensorTDVP/YqF4o/src/expand.jl:158
 [24] expand(state::MPS, reference::MPO; alg::String, kwargs::@Kwargs{krylovdim::Int64, cutoff::Float32})
    @ ITensorTDVP ~/.julia/packages/ITensorTDVP/YqF4o/src/expand.jl:29

[mtfishman]

Hi @YotamKa, thanks for the report, you're on the right track that we need to convert that δ tensor to GPU, though I see it looks like even after you tried that it is hitting some other bug. We'll look into it, I'm sure it is a simple fix.

Also thanks for reporting this to ITensorMPS.jl! The situation around where MPS code is defined is a bit complicated, glad to hear someone is following along haha.

[mtfishman]

I think the issue with tr on GPU should have been fixed by https://github.com/ITensor/ITensors.jl/pull/1453, are you sure your packages are up to date?

[YotamKa]

Hi, that's a good point. I saw the GitHub discussion about that (tr() with GPU). I'll update and then report again.

[mtfishman]

Should be fixed by https://github.com/ITensor/ITensorTDVP.jl/pull/85, I don't see any issues with tr when I use the latest version of NDTensors.jl.