Adding decomposition MPSPrep #6896

KetpuntoG · 2025-01-28T22:27:33Z

Before submitting

Please complete the following checklist when submitting a PR:

All new features must include a unit test.
If you've fixed a bug or added code that should be tested, add a test to the
test directory!
All new functions and code must be clearly commented and documented.
If you do make documentation changes, make sure that the docs build and
render correctly by running make docs.
Ensure that the test suite passes, by running make test.
Add a new entry to the doc/releases/changelog-dev.md file, summarizing the
change, and including a link back to the PR.
The PennyLane source code conforms to
PEP8 standards.
We check all of our code against Pylint.
To lint modified files, simply pip install pylint, and then
run pylint pennylane/path/to/file.py.

When all the above are checked, delete everything above the dashed
line and fill in the pull request template.

Context:

[SC-83733]
[SC-81348]
MPSPrep only works in "lightning.tensor". With this decomposition, it will work in any device

Description of the Change:

Benefits:

Possible Drawbacks:

Related GitHub Issues:

github-actions · 2025-01-28T22:27:47Z

Hello. You may have forgotten to update the changelog!
Please edit doc/releases/changelog-dev.md with:

A one-to-two sentence description of the change. You may include a small working example for new features.
A link back to this PR.
Your name (or GitHub username) in the contributors section.

codecov · 2025-01-28T22:43:39Z

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 99.59%. Comparing base (04bfe4d) to head (4a1b932).
Report is 10 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #6896      +/-   ##
==========================================
- Coverage   99.59%   99.59%   -0.01%     
==========================================
  Files         478      479       +1     
  Lines       45306    45362      +56     
==========================================
+ Hits        45124    45179      +55     
- Misses        182      183       +1

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

…I/pennylane into MPSPrep-decomposition

pennylane/templates/state_preparations/state_prep_mps.py

…I/pennylane into MPSPrep-decomposition

albi3ro

We worked hard to remove global random number generation from pennylane, and so I'd rather not have it go back in, especially with interfering with the user's choice of seed.

global random generation has been removed. No longer blocking the PR.

mlxd

Hey @KetpuntoG
just a few fly-by questions from me

mlxd · 2025-01-29T22:25:50Z

pennylane/templates/state_preparations/state_prep_mps.py

+        unitary[:, :k] = columns
+
+    # Complete the remaining columns using Gram-Schmidt
+    rng = np.random.default_rng(42)


Does this need to be seeded? Won't this ensure that each and every call to this will use the same correlated RNG outputs for the same input?
If None it will pull entropy from the operating system's default source, which is likely a better source of randomness. Or, allowing the provision of a seed as a function input would be the best option, as this ensure control is at the caller's level. What do you think?

I had to do it or qml.assert_valid complained that the decomposition did not match the operator in the queue (two different Unitaries were generated)

Assert valid in the tests? But we should be able to specify the seed to ensure the tests generate the same output? I'd expect numerical differences in the outcomes with generators that are seeded differently.

Actually these values of the matrices are not used for anything, we generate them just to be able to put the matrix in QubitUnitary. Since it is also a private function, I think it is ok to set it in this case :) User does not need to know about it

pennylane/templates/state_preparations/state_prep_mps.py

obliviateandsurrender

Thanks @KetpuntoG! Just leaving the first set of comments and will take a deeper look once these are addressed.

obliviateandsurrender · 2025-01-30T19:43:54Z

pennylane/templates/state_preparations/state_prep_mps.py

+    ortogonal_projection_matrix = qml.math.eye(d) - qml.math.dot(unitary, qml.math.conj(unitary.T))
+    new_columns = qml.math.dot(ortogonal_projection_matrix, new_columns)


Could we directly be able to use QR decomposition here? 🤔 I was trying something like this -

n, k = 8, 3 # number of rows, fixed given number of columns vectors = random_orthonormal_vectors(n, k) # this function gives me random orthonormal vectors randext = np.random.randn(n, n - k) Q, _ = np.linalg.qr(np.hstack([vectors, randext])) assert np.allclose(Q @ Q.T, np.eye(len(unitary))) and np.allclose(Q @ Q.T, Q.T @ Q)

It looks good but sometimes it changes the sign of columns, see the following example

For fixing this sign flip, maybe this should work:

Q, R = np.linalg.qr(np.hstack([vectors, randext])) Q *= np.sign(np.diag(R))

Awesome 🤩

pennylane/templates/state_preparations/state_prep_mps.py

obliviateandsurrender · 2025-01-30T19:52:56Z

pennylane/templates/state_preparations/state_prep_mps.py

+
+            for column in Ai:
+
+                vector = qml.math.zeros(2**n_wires, like=mps[0])


I thought argument like should be given a string - "jax", "numpy", etc. Might be nice to check once.

It is not very well documented that argument 🤔
If this help, I see this also in other parts in PL:

doing the latter is twice faster though. We can store the interface outside the loop and use that within the loop. It would also help with the line 265 check.

After in person discussion, we keep it as it is right now (jit complains about the change)

After testing something like this should work -

interface, dtype = qml.math.get_interface(mps[0]), mps[0].dtype vector = qml.math.zeros(2**n_wires, like=interface, dtype=dtype)

pennylane/templates/state_preparations/state_prep_mps.py

obliviateandsurrender · 2025-01-31T21:20:25Z

pennylane/templates/state_preparations/state_prep_mps.py

+        work_wires (Sequence[int]): list of extra qubits needed in the decomposition. The bond dimension of the mps
+                        is defined as ``2^len(work_wires)``. Default is ``None``


Should we define this as the maximum permissible bond dimension of the provided MPS?

obliviateandsurrender · 2025-01-31T21:21:44Z

pennylane/templates/state_preparations/state_prep_mps.py

+        self.hyperparameters["input_wires"] = qml.wires.Wires(wires)
+
+        if work_wires:
+            self.hyperparameters["work_wires"] = qml.wires.Wires(work_wires)


Do we require a check that sufficient work wires have been provided?

obliviateandsurrender · 2025-01-31T21:23:43Z

pennylane/templates/state_preparations/state_prep_mps.py

+            if i == 0:
+                Ai = Ai.reshape((1, *Ai.shape))
+            elif i == len(mps) - 1:
+                Ai = Ai.reshape((*Ai.shape, 1))


Could we do this rank-2 to rank-3 promotion outside the loop itself to prevent checking these conditions for every iteration?

Conditional not needed, agree.

obliviateandsurrender · 2025-01-31T21:29:32Z

pennylane/templates/state_preparations/state_prep_mps.py

+
+            for column in Ai:
+
+                vector = qml.math.zeros(2**n_wires, like=mps[0])


doing the latter is twice faster though. We can store the interface outside the loop and use that within the loop. It would also help with the line 265 check.

pennylane/templates/state_preparations/state_prep_mps.py

Co-authored-by: Utkarsh <[email protected]>

austingmhuang

@KetpuntoG Nice work, looks good.

austingmhuang · 2025-02-04T15:18:41Z

pennylane/templates/state_preparations/state_prep_mps.py



    Args:
        mps (List[Array]):  list of arrays of rank-3 and rank-2 tensors representing an MPS state as a
            product of site matrices. See the usage details section for more information.

        wires (Sequence[int]): wires that the template acts on
+        work_wires (Sequence[int]): list of extra qubits needed in the decomposition. The bond dimension of the mps
+                        is defined as ``2^len(work_wires)``. Default is ``None``


Suggested change

is defined as ``2^len(work_wires)``. Default is ``None``

is defined as ``2^len(work_wires)``. Default is ``None``.

austingmhuang · 2025-02-04T15:18:51Z

pennylane/templates/state_preparations/state_prep_mps.py



    Args:
        mps (List[Array]):  list of arrays of rank-3 and rank-2 tensors representing an MPS state as a
            product of site matrices. See the usage details section for more information.

        wires (Sequence[int]): wires that the template acts on
+        work_wires (Sequence[int]): list of extra qubits needed in the decomposition. The bond dimension of the mps


austingmhuang · 2025-02-04T19:05:37Z

pennylane/templates/state_preparations/state_prep_mps.py

+        This operator is natively supported on the ``lightning.tensor`` device, designed to run MPS structures efficiently. For other devices, implementing this operation uses a gate-based decomposition which requires auxiliary qubits (via ``work_wires``) to prepare the state vector represented by the MPS in a quantum circuit.
+


Just curious, is there a reference to this implementation using a gate-based decomposition?

Ah I see it is in the compute_decomposition function. not sure where it's preferred.

austingmhuang · 2025-02-04T19:23:38Z

pennylane/templates/state_preparations/state_prep_mps.py

+        hyperparameters = (
+            ("wires", self.hyperparameters["input_wires"]),
+            ("work_wires", self.hyperparameters["work_wires"]),
+        )


Can we make this a dictionary so we don't have to do hyperparams_dict = dict(metadata)?

austingmhuang · 2025-02-04T19:27:04Z

pennylane/templates/state_preparations/state_prep_mps.py

+        filtered_hyperparameters = {
+            key: value for key, value in self.hyperparameters.items() if key != "input_wires"
+        }
+        return self.compute_decomposition(
+            self.parameters, wires=self.hyperparameters["input_wires"], **filtered_hyperparameters
+        )
+


Can we store self.hyperparameters["input_wires"] in a variable, then simply del self.hyperparameters["input_wires"], and do **self.hyperparameters?

austingmhuang · 2025-02-04T19:28:53Z

pennylane/templates/state_preparations/state_prep_mps.py

+
+        Args:
+            mps (List[Array]):  list of arrays of rank-3 and rank-2 tensors representing an MPS state as a
+                product of site matrices. See the usage details section for more information.


Suggested change

product of site matrices. See the usage details section for more information.

product of site matrices.

austingmhuang · 2025-02-04T19:32:16Z

pennylane/templates/state_preparations/state_prep_mps.py

+        """
+
+        if work_wires is None:
+            raise AssertionError(


Suggested change

raise AssertionError(

raise ValueError(

IMO, AssertionError is more for things like:

try: assert work_wires is None except AssertionError: ....

Also, what if not enough work_wires are provided? Be cautious here.

austingmhuang · 2025-02-04T19:35:18Z

pennylane/templates/state_preparations/state_prep_mps.py

+            if i == 0:
+                Ai = Ai.reshape((1, *Ai.shape))
+            elif i == len(mps) - 1:
+                Ai = Ai.reshape((*Ai.shape, 1))


Conditional not needed, agree.

decomposition MPS

3527063

Merge branch 'master' into MPSPrep-decomposition

67f04b6

KetpuntoG added 4 commits January 29, 2025 09:24

extra tests

f1ec861

Merge branch 'MPSPrep-decomposition' of https://github.com/PennyLaneA…

57b0c31

…I/pennylane into MPSPrep-decomposition

change log

0d1ab8d

Merge branch 'master' into MPSPrep-decomposition

9a55803

KetpuntoG mentioned this pull request Jan 29, 2025

Test MPS PennyLaneAI/pennylane-lightning#1058

Open

albi3ro reviewed Jan 29, 2025

View reviewed changes

pennylane/templates/state_preparations/state_prep_mps.py Outdated Show resolved Hide resolved

KetpuntoG added 2 commits January 29, 2025 09:45

black

ac71564

Merge branch 'MPSPrep-decomposition' of https://github.com/PennyLaneA…

3259d71

…I/pennylane into MPSPrep-decomposition

albi3ro previously requested changes Jan 29, 2025

View reviewed changes

KetpuntoG added 2 commits January 29, 2025 10:07

local random

c440144

Merge branch 'master' into MPSPrep-decomposition

ac0a31e

KetpuntoG requested a review from obliviateandsurrender January 29, 2025 15:42

mlxd reviewed Jan 29, 2025

View reviewed changes

Update state_prep_mps.py

625758f

obliviateandsurrender requested changes Jan 30, 2025

View reviewed changes

This comment was marked as duplicate.

Sign in to view

KetpuntoG added 2 commits January 31, 2025 09:35

comments Utkarsh

298cec9

QR decomposition

4a1b932

obliviateandsurrender requested changes Jan 31, 2025

View reviewed changes

KetpuntoG requested a review from austingmhuang February 3, 2025 15:20

KetpuntoG and others added 2 commits February 3, 2025 17:19

Update pennylane/templates/state_preparations/state_prep_mps.py

39af831

Co-authored-by: Utkarsh <[email protected]>

Update pennylane/templates/state_preparations/state_prep_mps.py

8ca8615

Co-authored-by: Utkarsh <[email protected]>

austingmhuang requested changes Feb 4, 2025

View reviewed changes

adding test that now fails

10bbbaf

		ortogonal_projection_matrix = qml.math.eye(d) - qml.math.dot(unitary, qml.math.conj(unitary.T))
		new_columns = qml.math.dot(ortogonal_projection_matrix, new_columns)


		for column in Ai:

		vector = qml.math.zeros(2**n_wires, like=mps[0])

		work_wires (Sequence[int]): list of extra qubits needed in the decomposition. The bond dimension of the mps
		is defined as ``2^len(work_wires)``. Default is ``None``

		This operator is natively supported on the ``lightning.tensor`` device, designed to run MPS structures efficiently. For other devices, implementing this operation uses a gate-based decomposition which requires auxiliary qubits (via ``work_wires``) to prepare the state vector represented by the MPS in a quantum circuit.

	product of site matrices. See the usage details section for more information.
	product of site matrices.

Adding decomposition MPSPrep #6896

Are you sure you want to change the base?

Adding decomposition MPSPrep #6896

Conversation

KetpuntoG commented Jan 28, 2025 • edited Loading

Before submitting

github-actions bot commented Jan 28, 2025

codecov bot commented Jan 28, 2025 • edited Loading

Codecov Report

albi3ro left a comment

Choose a reason for hiding this comment

mlxd left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

obliviateandsurrender left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

obliviateandsurrender Jan 31, 2025 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

This comment was marked as duplicate.

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

austingmhuang left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

KetpuntoG commented Jan 28, 2025 •

edited

Loading

codecov bot commented Jan 28, 2025 •

edited

Loading

obliviateandsurrender Jan 31, 2025 •

edited

Loading