|
Hi, first of all thanks for this great library. |
Answered by
yangcal
Sep 24, 2025
Replies: 4 comments 11 replies
|
Hello, Thanks for sharing the user experience. Please note that the Meanwhile, cuquantum-python now also offers a direct MPS simulator using RE: speeding up, generally the majority of the cost in simulating MPS amounts to the SVD step, which then heavily depends on the bond dimension. If your circuit state is sparse, you may more aggressively truncate the bonds to make it run faster, or you may try a few SVD options that we provide, e.g, Feel free to share more feedbacks if you try these APIs |
3 replies
|
@yangcal a follow-up question on contract and decompose when you get a chance. If I have a batch of circuits with exactly the same qubit count, number of layers, and entanglement. Is there a way to use the Here is what I naively tried (for a batch size of 1) import cupy as cp
from cuquantum.tensornet import contract, tensor
mps_tensors = [cp.array([1,0], dtype=cp.float32).reshape(1,1,2,1)]*2
gate = cp.array([[1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0], [0, 0, 0, -1]], dtype=cp.float32).reshape(1,2,2,2,2)
t = contract('...ipj,...jqk,...pqPQ->...iPQk', *mps_tensors, gate)
u, s, v = tensor.decompose('...iPQk->...iPx,...xQk', t, method=tensor.SVDMethod())but it seems that the subscripts I use for the decomposition do not work |
|
Even if we're to support this, it will be effectively looping over the batched indices and perform multiple SVDs separately. Note that it's not possible to enforce value based truncation in this setting as each batched SVD will have different singular value distributions. |
|
Understood, thank you! |
Answer selected by
dupuisni
|
Thanks @yangcal I did not know about the circuit = get_qiskit_circuit()
config = MPSConfig(max_extent=32, abs_cutoff=1.e-12)
state = NetworkState.from_circuit(circuit, dtype='complex128', backend='cupy', config=config)
state.compute_output_state()but it fails with: I am guessing I am doing something wrong here. What I really care to get are the MPS tensors which I would like to use in some other downstream application. |
3 replies
|
I don’t see anything going obviously wrong here. The comupute_output_state method should just return the final mps tensors. Are you able to run one of the examples? Or do you mind sharing the full script that you’re using? |
|
Maybe there is something wrong with my install, I get the same exception for all the examples that use import numpy as np
from cuquantum.tensornet.experimental import NetworkState, MPSConfig
import qiskit
def entangled_random_circuit(RNG, num_qubits, reps):
def entanglement(layer_index):
if layer_index % 2 == 0:
return [(2 * i, 2 * i + 1) for i in range(num_qubits // 2)]
return [(2 * i + 1, 2 * i + 2) for i in range((num_qubits - 1) // 2)]
circuit = qiskit.circuit.library.n_local(
num_qubits, "rx", "cy", entanglement, reps=reps
)
random_params = np.around(RNG.random(len(circuit.parameters)), decimals=3)
params = dict(zip(circuit.parameters, random_params))
return circuit.assign_parameters(params)
def main():
RNG = np.random.default_rng(seed=37)
circuit = entangled_random_circuit(RNG, num_qubits=10, reps=10)
config = MPSConfig(max_extent=32, abs_cutoff=1.e-12)
state = NetworkState.from_circuit(circuit, dtype='complex128', backend='cupy', config=config)
print(state.compute_output_state())
if __name__ == "__main__":
main() |
|
I tried running your snippet on an H100 node and it works all just fine. Do you mind sharing more information on how cuquantum was installed, what dependencies you have in your environment and cuda version / driver version and hardware info (e.g, via Meanwhile you can also do |
|
I just started a new environment installing with pip not sure about this: I am running on A100, here is the output of |
1 reply
This is very helpful. I assume that before you install pytorch, you already have some CUDA toolkit installed somewhere and exposed it to On a side note, your driver version is 550.90.07 and the maximal CUDA Version that it supports is 12.4. If pytorch brings cuda libs>12.4, this could be problematic. |
|
Thanks @yangcal, you are right, CUDA version mismatch was the issue, I got it work with: One odd thing I noticed is that I need to do and the second time I get the rest of the dependencies: I am installing with uv in a virtual environment |
4 replies
|
|
|
I miss all those: which I get when I re-run |
|
I'm not able to reproduce what you see, I started a new conda virtual env and the following appears: If I actually launch Do you possibly have various pips in your env? |
|
Sorry I was not clear, starting from a clean environment, this: works perfectly for me, but this, again starting from a clean environment: needs to be ran twice to get all the dependencies installed. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Hello,
Thanks for sharing the user experience. Please note that the
mps_algorithm.ipynbis meant to demonstrate how to leverage our primitivecontract_decomposepython API to construct your own MPS or other approximate tensor network simulator. The corresponding C API iscutensornetGateSplit. Themps_example.cuonly demonstrated the handling of two-qubit gate as single qubit gate handling amounts to pairwise contraction, which should be delegated tocutensorlibrary outside cuquantum.Meanwhile, cuquantum-python now also offers a direct MPS simulator using
cuquantum.tensornet.experimental.NetworkStateAPI, see here for documentation and here for examples.RE: speeding up, generally the ma…