I was trying to reproduce the statement in https://developer.nvidia.com/blog/nvidia-announces-cuquantum-beta-availability-record-quantum-benchmark-and-quantum-container/, in particular
Quantum Fourier Transform – accelerated from 29 mins down to 19 secs
I'm running the following minimal reproducer on Python 3.8, cuQuantum 22.11, NVIDIA A100 40 GB (on a GCP instance)
import time
import cirq
import qsimcirq
import cupy
from cuquantum import contract
from cuquantum import CircuitToEinsum
simulator = qsimcirq.QSimSimulator()
# See https://quantumai.google/cirq/experiments/textbook_algorithms
def make_qft(qubits):
"""Generator for the QFT on a list of qubits."""
qreg = list(qubits)
while len(qreg) > 0:
q_head = qreg.pop(0)
yield cirq.H(q_head)
for i, qubit in enumerate(qreg):
yield (cirq.CZ ** (1 / 2 ** (i + 1)))(qubit, q_head)
def simulate_and_measure(nqubits):
qubits = cirq.LineQubit.range(nqubits)
qft = cirq.Circuit(make_qft(qubits))
myconverter = CircuitToEinsum(qft, backend=cupy)
tic = time.time()
simulator.simulate(qft)
elapsed_qsim = time.time() - tic
out = {"qsim": elapsed_qsim}
# CUDA expectation
pauli_string = {qubits[0]: 'Z'}
expression, operands = myconverter.expectation(pauli_string, lightcone=True)
tic = time.time()
contract(expression, *operands)
elapsed = time.time() - tic
out["cu_expectation"] = elapsed
# CUDA Batched amplitudes
# Fix everything but last qubit
fixed_states = "0" * (nqubits - 1)
fixed_index = tuple(map(int, fixed_states))
num_fixed = len(fixed_states)
fixed = dict(zip(myconverter.qubits[:num_fixed], fixed_states))
expression, operands = myconverter.batched_amplitudes(fixed)
tic = time.time()
contract(expression, *operands)
elapsed = time.time() - tic
out["cu_batched"] = elapsed
return out
for i in [10, 15, 20, 25, 30]:
print(i, simulate_and_measure(i))
Output (the numbers are elapsed in seconds; 10, 15, ... are number of qubits for QFT):
10 {'qsim': 0.9677999019622803, 'cu_expectation': 0.29337143898010254, 'cu_batched': 0.07590365409851074}
15 {'qsim': 0.023270368576049805, 'cu_expectation': 0.019628524780273438, 'cu_batched': 0.3687710762023926}
20 {'qsim': 0.03504538536071777, 'cu_expectation': 0.023822784423828125, 'cu_batched': 0.9347813129425049}
25 {'qsim': 0.14235782623291016, 'cu_expectation': 0.02486586570739746, 'cu_batched': 2.39030122756958}
30 {'qsim': 3.4044816493988037, 'cu_expectation': 0.028923749923706055, 'cu_batched': 4.6819908618927}
35 {'cu_expectation': 1.0615959167480469, 'cu_batched': 10.964831829071045}
40 {'cu_expectation': 0.03381609916687012, 'cu_batched': 82.43729209899902}
I wasn't able to go to 35 qubits for qsim, because I got CUDA OOM for qsim. The much reduced memory usage alone is sufficient to prefer cuQuantum for this use case.
But, I was hoping that batched_amplitudes is going to be faster than a full statevector simulation, because some qubits are fixed. But it doesn't seem to be the case. I have also tried reduced_density_matrix (not shown, so that the code snippet is short). The only one that is consistently fast is expectation. I wonder if I did it wrongly?
![[Question/Issue] How to install cuquantum on WSL2 Ubuntu-20.04?](https://avatars.githubusercontent.com/u/70229100?v=4)
Following the well designed documentation I was able to successfully get this to run. I am using terminal to write my Quantum experiments.I want to know how to run a jupyter notebook inside docker container. After installing jupyter in docker and running it gives a link that doesn't open in a browser of my linux system.I then installed firefox inside cuQuantum docker but it doesn't launch the browser. Gives a display not found error.How can I run my experiments in jupyter ?
89 Jan 05, 2023
19 Dec 19, 2022
2.3k Jan 08, 2023
149 Dec 15, 2022
115 Nov 21, 2022
3 Jun 13, 2022
4 Aug 07, 2022
38 Dec 26, 2022
10 Oct 18, 2022
556 Jan 04, 2023
223 Jan 04, 2023
63 Sep 25, 2022
219 Dec 27, 2022
38 Oct 21, 2022
6 Jul 24, 2022
153 Nov 14, 2022
7 Jan 03, 2023
2.8k Dec 31, 2022
1 Nov 12, 2021
4.2k Jan 01, 2023