Bump Numpy #7033
Comments
|
Thanks for opening this issue @MIWdlB! Our team is looking into this now, but there might be a pathway for removing our NumPy pinning for certain workflows. However, this becomes a bit more complicated if you are using JAX with PennyLane. Currently, are you using Catalyst and/or JAX for anything? |
|
Hi @MIWdlB , I just wanted to give you an update on this. We're evaluating options at the moment. We have PR #7039 open to try to unpin NumPy. Since this may not be a trivial change we may need to wait a bit before this gets merged into master. However you can follow the PR to see if we have any updates there. |
|
Thanks for the update, I've had to remove penny lane from my own repos for the time being. I'll keep an eye on the PR and add it back in when the time comes. If there's jobs along the way that I can help out on I'd be happy to |
|
Oh that's unfortunate. Hopefully we'll be able to unpin it soon. We'll keep you updated and let you know if we can use your help! At the very least we can use your help flagging any issues once the change is done 😄 |
## Context: Previously, PennyLane pinned NumPy to version `numpy<2.1` (see the [first commit](449ae04) of this PR). This restriction created compatibility challenges for users building their own projects with PennyLane as a dependency, as reported in #7033. We've determined there are no fundamental blockers preventing PennyLane from supporting newer NumPy versions - only historical integration issues, particularly with TensorFlow, that have caused cascading CI problems and potential ecosystem compatibility concerns. With NumPy 1.26's end-of-support approaching in September 2025 ([reference](https://endoflife.date/numpy)), it's critical we thoroughly investigate and address these obstacles. This PR aims to explore unpinning NumPy, document issues encountered, implement fixes, and adjust our CI pipeline accordingly.  Besides, we can check the NEP 29 Drop Schedule: https://numpy.org/neps/nep-0029-deprecation_policy.html ## Description of the Change: This PR follows a systematic approach: 1. Initially unpinning NumPy~1.26 but setting `>= 2.0.1` to ensure the main environment uses NumPy 2 2. Documenting and analyzing all issues encountered with NumPy 2 compatibility 3. Methodically resolving each identified issue 4. Fully unpinning NumPy to allow maximum flexibility 5. Updating CI workflows as needed to accommodate version differences 6. Adding demonstrations showing PennyLane works safely with projects requiring NumPy 2 Specific changes include: - [x] Updated `tensorflow` to version `2.18` for NumPy 2 compatibility across multiple files: - `.github/workflows/interface-dependency-versions.yml` - `docker/interfaces/install-interface-gpu.sh` - `docker/interfaces/install-interface.sh` - `requirements.txt` - `doc/requirements.txt` - `setup.py` - [x] Updated `jax` and `jaxlib` to `0.4.28` for NumPy 2 compatibility in `doc/requirements.txt` - [x] Pinned `ml_dtypes` for compatibility in `requirements.txt` and `setup.py` - [x] Added conditional xfail decorator for `KerasLayer` tests that won't pass with NumPy 2 (pending deprecation) - [x] Added conditional xfail for `qchem` tests requiring `openfermion` based on NumPy version Extra manual test: - [x] Locally installation of `numpy-2.2` works without conflicts with PL - [x] Ci tests with other affected packages in PL ecosys passed: - [x] lightning: PennyLaneAI/pennylane-lightning#1095 - [x] pennylane-cirq: PennyLaneAI/pennylane-cirq#211 - [x] pennylane-qiskit: PennyLaneAI/pennylane-qiskit#622 ## Benefits: - Identified and resolved root causes of NumPy 2.x CI failures - Users can now use PennyLane with NumPy 2.x environments without compatibility issues - Improved future maintainability by removing artificial version constraints - Better aligned with the Python ecosystem's forward progress ## Remaining Challenges: We encountered several issues that warrant further investigation, though not crucially important for now: - Sphinx documentation generation succeeded with NumPy 1.x but fails with NumPy 2.x despite no obvious connection, since the documentation building should have their own dependencies in principle, as `doc/requirements.txt` exists. However, this was reflected possibly resulting from that there used to be some spaghetti between main requiremnts and doc/requirements. - Inconsistent test results between CI environment and local development environment when using NumPy 2 + TensorFlow 2.18 + Keras 3. Basically, if one tested Keras on their local machine, they should see ~half~ one third of them to fail; however in our CI, actually only one of them fails; this is also why the xfail macro was defined by not widely used in the affected test file. This phenomenon could be caused by some special environment variables in workflow setup, s.t. eventually the actual Keras used is still Keras2. ## Related GitHub Issues: [sc-85628] --------- Co-authored-by: JerryChen97 <chenys13@outlook.com>
|
Hi @MIWdlB , our team was able to remove the NumPy pin! This change is available in the Master version of PennyLane, and will be available in the next stable release of PennyLane. It would help a lot if you could try using the master version of PennyLane and let us know of any issues you encounter! Thanks again for opening this issue and enjoy using PennyLane! |
Feature details
Pinning numpy to <2.1 is now becoming restrictive. What's the issue that stops this from being <3?
Implementation
No response
How important would you say this feature is?
3: Very important! Blocking work.
Additional information
No response
The text was updated successfully, but these errors were encountered: