GPU - Unified Memory error with NDTensors.cu()

ITensor Julia Questions
1

Hi, ITensor Teams！

I am trying to use GPU to improve a code using multiples GPU at the same time trying to follow Multiple GPUs · CUDA.jl (juliagpu.org). In general is important to use the unified memory, and usually to do that it is enough to put unified = true as an extra parameter. However, in my code I am using NDTensors.cu() instead of cu() because when I use cu() sometimes I got errors as the one showed here: TEBD with GPU - error with eigen - ITensor Julia Questions - ITensor Discourse

My problem is that when I put unified = true inside NDTensors.cu() I got the following error:

using ITensors

sites = siteinds("S=1/2",50)

#cu()
A = cu(randomMPS(sites); unified = true) 
#This works without problem

#NDTensors.cu()
A = NDTensors.cu(randomMPS(sites); unified=true) 
#Here I got: MethodError: no method matching cu(::MPS; unified::Bool)

So, I am not sure how to use unified memory and NDTensors.cu() at the same time.

2

Thanks for the report, we’ll have to look into that and think about how we want to support that.

3

In the meantime, you could try using adapt(CuArray{Float64,<:Any,UnifiedMemory}, x) where x is your MPS. adapt is defined in the package Adapt.jl so you’ll have to install and load that package.

Also note that the latest syntax is random_mps, not randomMPS, we’ll be deprecating randomMPS.

1 Like
4

Note that you may have to use adapt(CuArray{Float64,1,UnifiedMemory}, x) instead because of a bug/limitation in CUDA.jl.

5

I raised an issue for the Metal.jl backend here: Generalize `adapt` to allow specifying the storage mode but not the element type and/or number of dimensions · Issue #363 · JuliaGPU/Metal.jl · GitHub but the same issue should be raised in CUDA.jl and other GPU backends.

1 Like
6

@Joacop16 I think you could also use NDTensors.cu(x; storagemode=UnifiedMemory), I forgot we had added that.

2 Likes
7

Hi @joacop16, we did change the implementation of our NDTensors.cu function and using the keyword unified should throw an error because that is not an acceptable keyword.

julia> NDTensors.cu(randomMPS(sites); unified=true)
ERROR: MethodError: no method matching cu(::MPS; unified::Bool)

Closest candidates are:
  cu(::Any; storagemode) got unsupported keyword argument "unified"
   @ NDTensorsCUDAExt ~/.julia/dev/ITensors/NDTensors/ext/NDTensorsCUDAExt/adapt.jl:10

Stacktrace:
 [1] kwerr(::@NamedTuple{unified::Bool}, ::Function, ::MPS)
   @ Base ./error.jl:165
 [2] top-level scope

@ REPL[24]:1 What version of ITensors/NDTensors are you using?
I did run your example with the code Matt suggests and it does properly construct an MPS on GPU with UnifiedMemory

julia> sites = siteinds("S=1/2",2)
2-element Vector{Index{Int64}}:
 (dim=2|id=214|"S=1/2,Site,n=1")
 (dim=2|id=199|"S=1/2,Site,n=2")

julia> NDTensors.cu(randomMPS(sites); storagemode=CUDA.UnifiedMemory)[1]
ITensor ord=2 (dim=2|id=214|"S=1/2,Site,n=1") (dim=1|id=406|"Link,l=1")
NDTensors.Dense{Float64, CuArray{Float64, 1, CUDA.UnifiedMemory}}
8

This topic was automatically closed 10 days after the last reply. New replies are no longer allowed.