Memory Usage

[cfengno1]

Hi Huanchen,

I have a few quesitons about the memory usage in block2.

1. I notice that there is a parameter ``min_mpo_mem" for dynamically loading MPO. Can I do the same for MPS?

2. According to the document, the parameter ``stack_mem" stands for the memory used for storing renormalized operators (in bytes). Does it mean that all environment tensors are stored in memory instead of on disk?

3. I notice three additional parameters that save memory: fused_contraction_multiplication, fused_contraction_rotation, lowmem_numerical_transform. Can you help explain what is fused contraction and numerical transform used here?

Thank you so much!

Best,
Feng

[hczhai]

1. MPS is by default dynamically loaded into memory.
2. Only renormalized operators that are relevant to the current iteration is loaded into memory.
3. As shown in the Fig. 6 of the paper J. Chem. Phys. 154, 224116 (2021), "contraction" is Fig. 6(a), "rotation" is Fig. 6 (b), and "multiplication" is Fig. 6(c). "Numerical transform" is explained in https://block2.readthedocs.io/en/latest/theory/spatial.html#middle-site-transformation. When you are not doing DMRG for quantum chemistry Hamiltonian using the conventional implementation, there is no "numerical transform" step.

[cfengno1]

Hi Huanchen,

THank you for the quick reply!

Regarding the contraction, rotation and multiplication, my queston is what does "fused" mean? How does it save memory?

Best,
Feng

[hczhai]

Fused X and Y means opY(opX(A, B), C) is implemented as opXY(A, B, C) without storing the intermediate opX(A, B) in memory. Since these options may conflict with other more effective optimizations, it is very likely for a specific usage case there is no saving of the peak memory during the whole calculation. Normally the default option is recommended.