Enable OpenMP pragmas for gate kernels on Linux Wheels

[maliasadi]

For context, the default threading behavior of lightning.qubit assumes single-threaded execution at the gate level, while allowing multi-threaded execution over observables when using the adjoint differentiation pipeline. This design ensures optimal (SIMD) gate kernel performance but can limit performance for deep circuits on large CPUs with many cores and large caches.

The flag LQ_ENABLE_KERNEL_OMP enables OpenMP support across all kernel types (LM, AVX2, and AVX512) which allows for better performance tuning on HPC systems while preserving the default behavior for standard releases. This PR makes this default on LQ wheels and from-src installations.

[latest/latest]

[sc-101572]
[sc-104235]

[github-actions]

Hello. You may have forgotten to update the changelog!
Please edit .github/CHANGELOG.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]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.26%. Comparing base (e9353ef) to head (99a1036).

❗ There is a different number of reports uploaded between BASE (e9353ef) and HEAD (99a1036). Click for more details.

HEAD has 24 uploads less than BASE

Flag	BASE (e9353ef)	HEAD (99a1036)
unit_tests	31	7

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1133      +/-   ##
==========================================
- Coverage   95.84%   89.26%   -6.59%     
==========================================
  Files         243      187      -56     
  Lines       40695    28938   -11757     
==========================================
- Hits        39005    25831   -13174     
- Misses       1690     3107    +1417     

Flag	Coverage Δ
unit_tests	89.26% <100.00%> (-6.59%)	⬇️

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

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[maliasadi]

/benchmark

[LuisAlfredoNu]

/benchmark

[maliasadi]

/benchmark

[maliasadi]

/benchmark

[ringo-but-quantum]

PennyLane Benchmarks

Benchmarks Report for lightning.qubit

Weighted Average Scores

Direct execution (without Catalyst)

• Time: 1.53 🟢

• Virtual Memory: 1.01

Aggregated Results

Time score

Workflow	Python
QPE	1.57	🟢
QSVT	0.96	🔴
XAS	2.93	🟢
shor	-
molecular_hamiltonian	1.01
sampling	1.10	🟢
stateprep	1.34	🟢
grover	2.68	🟢
QAOA_layers_scaling	1.33	🟢
QML	-
QML_jaxjit	1.02
UCCSD	0.94	🔴
VQE	1.05	🟢

Memory score

Workflow	Python virtual
QPE	1.00
QSVT	1.00
XAS	1.00
shor	-
molecular_hamiltonian	1.00
sampling	1.00
stateprep	1.07	🟢
grover	1.00
QAOA_layers_scaling	1.00
QML	-
QML_jaxjit	1.02
UCCSD	1.00
VQE	1.00

Complexity score

No complexity data available.

Detailed Per-Workflow Results

• The assumed noise level for runtime improvements/regressions is 3.0%
• ⚠️ marks workflows with runtime fluctuations greater than 5.0% (std/mean)

Direct execution (without Catalyst)

Workflow	time [s]	std/mean		time score		virt mem [MB]	virt mem score
QPE[12-12]	0.971	0.4%		1.57	🟢	268.690	1.00
QPE[13-12]	1.978	0.1%		1.57	🟢	537.121	1.00
QSVT[10]	1.057	5.6%	⚠️	0.96	🔴	2.282	1.00
XAS[2-1-9]	1.386	5.2%	⚠️	3.04	🟢	15.057	1.00
XAS[2-2-9]	3.020	7.6%	⚠️	2.82	🟢	15.048	1.00
molecular_hamiltonian[H2O-STO-3G]	3.109	1.0%		1.01		136.276	1.00
molecular_hamiltonian[NH3-STO-3G]	9.016	0.3%		1.01		334.763	1.00
sampling[24-2]	1.204	0.2%		1.10	🟢	812.612	1.00
sampling[25-2]	2.469	1.1%		1.10	🟢	1,623.308	1.00
stateprep[14-MottonenStatePreparation]	2.935	2.8%		0.82	🔴	26.530	1.08	🟢
stateprep[15-ArbitraryStatePreparation]	2.940	5.2%	⚠️	1.86	🟢	18.416	1.06	🟢
grover[18]	0.610	3.9%		2.68	🟢	13.020	1.00
QAOA_layers_scaling[19-4]	0.866	0.6%		1.33	🟢	310.716	1.00
QML_jaxjit[IQPKernelClassifier-12-10]	4.761	18.2%	⚠️	0.97	🔴	26.091	1.04	🟢
QML_jaxjit[IQPVariationalClassifier-20-10]	3.890	8.2%	⚠️	1.07	🟢	34.584	1.00
UCCSD[H2O-STO-3G]	0.915	8.3%	⚠️	0.81	🔴	136.468	1.00
UCCSD[NH3-STO-3G]	2.834	3.5%		1.07	🟢	334.777	1.00
VQE[H2O-STO-3G]	2.475	2.6%		1.05	🟢	136.148	1.00
VQE[NH3-STO-3G]	14.167	0.3%		1.05	🟢	334.743	1.00

System Information

Hardware and Software Specifications

System	Current	Reference
CPU Name	Intel(R) Xeon(R) E-2388G CPU @ 3.20GHz	Intel(R) Xeon(R) E-2388G CPU @ 3.20GHz
Architecture	X86_64	X86_64
Cache L2	4096.0 KiB	4096.0 KiB
Cache L3	16384.0 KiB	16384.0 KiB
Cache L1	256.0 KiB I + 384.0 KiB D	256.0 KiB I + 384.0 KiB D
Max MHz	0.0	0.0
Min MHz	0.0	0.0
Nominal	3287.5	3287.5
Enabled	8 cores, 16 threads	8 cores, 16 threads
Memory	62.64 GB	62.64 GB
Storage	435.52 GB on /	435.52 GB on /
OS	Linux	Linux
OS Version	Linux-5.15.0-142-generic-x86_64-with-glibc2.35	Linux-5.15.0-142-generic-x86_64-with-glibc2.35
Kernel	5.15.0-142-generic	5.15.0-142-generic
Hostname	agassi	agassi
Python Version	3.11.9	3.11.9
Python Compiler	GCC 11.4.0	GCC 11.4.0
PennyLane Version	0.44.0.dev7	0.44.0.dev7
Lightning Version	0.44.0.dev3	0.44.0.dev3
Catalyst Version	Unknown	Unknown
NumPy Version	2.3.4	2.3.4
SciPy Version	1.16.2	1.16.2
Lightning Compiler	Unknown	Unknown

Report generated on Thu, 16 Oct 2025 at 01:32:38 (UTC)

[josephleekl]

seems to work well! Some suggestions for things to update:

• docs to mention this is on by default, and can be toggled on/off with OMP_NUM_THREADS, and might be a good idea to set to 1 if using e.g. concurrency
• update changelog
• in `.github/workflows/tests_lqcpu_python.yml we set it to on, with this change, we don't need to explicitly do that anymore?

[maliasadi]

@josephleekl Now, it's ready to review ⚡

[LuisAlfredoNu]

One question. Doing the OMP ON as default, will not interfere with the Parallel adjoint differentiation functionality? We will have the chance of running each gate with threads and each differentiation with more threads, producing a slowdown due to thread pool flood 🤔

[josephleekl]

Thanks @maliasadi , LGTM! 🚀

[LuisAlfredoNu]

Thanks @maliasadi
🐊

[mlxd]

Just adding the red to prevent this from accidentally merging.

We need to do more intensive eval, including across the python package ecosystem and OSs. Will do a proper review next week

[mlxd]

Nice one @maliasadi
Just a few queries/suggestions, but will be happy to approve pending your response.