Memory increases with growing number of time steps in cudaq.dynamics

[chnixi]

It seems that the total memory requirement on GPUs increases almost linearly with growing number of time steps in time evolution of cudaq.dynamics (cu12-0.14.0), despite the fact that only a small number of observables are being monitored. Is there a way for a fixed problem which works well for a small number of time steps to scale up to a larger number of time steps, without running out of memory?

[sacpis]

Will look into it. Would you be able to post a sample example here?

[chnixi]

Many thanks for getting back so quickly! Below is an example script, the number of time steps is controlled by 'n_steps' which is set to 50 currently: ``` import numpy as np import cudaq from cudaq import boson, Schedule, ScipyZvodeIntegrator, spin, RungeKuttaIntegrator import cupy as cp L=16 cudaq.set_target("dynamics") #initial condition: every site is singly occupied init=np.zeros(np.power(3,L)) index=0 for i in range(L): index=index+pow(3,i) init[index]=1 hamiltonian = boson.create(0)*boson.annihilate(L-1) for i in range(L-1): hamiltonian = hamiltonian + boson.number(i)*boson.number(i) # Dimensions of sub-systems dimensions={} for i in range(L): dimensions[i] = 3 # Initial state of the system (ground state). rho0 = cudaq.State.from_data(cp.array(init , dtype=cp.complex128)) t_final=25 n_steps=50 # Schedule of time steps. steps = np.linspace(0, t_final, n_steps) schedule = Schedule(steps, ["t"]) # Run the simulation. evolution_result = cudaq.evolve(hamiltonian, dimensions, schedule, rho0, observables=[boson.number(0), boson.number(L-1)], collapse_operators=[], store_intermediate_results=True) ```

…

________________________________ From: Sachin Pisal ***@***.***> Sent: Wednesday, April 8, 2026 2:37 PM To: NVIDIA/cuda-quantum ***@***.***> Cc: Xiao, Chengnian ***@***.***>; Author ***@***.***> Subject: Re: [NVIDIA/cuda-quantum] Memory increases with growing number of time steps in cudaq.dynamics (Discussion #4279) Will look into it. Would you be able to post a sample example here? — Reply to this email directly, view it on GitHub<#4279 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AJJHPHUCRXF5ROXRDNTUJFD4U2L63AVCNFSM6AAAAACXRN7QMGVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTMNBZGMYDKOI>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[sacpis]

The store_intermediate_results=True is deprecated alias for IntermediateResultSave.ALL. It saves the entire state vector to the GPU at every time step. So that's why the memory growth is like O(n_steps * state_size).

Please use store_intermediate_results=cudaq.IntermediateResultSave.EXPECTATION_VALUE instead as it saves only the expectation values of 2 observables (boson.number(0) and boson.number(L-1)) at each step and not the state. It is like O(n_steps * 2) now.

[chnixi]

That seems to work. Many thanks for your help!