Decomp/auto gateset from backend toml by yde773786 · Pull Request #3275 · PennyLaneAI/catalyst

yde773786 · 2026-09-25T16:55:04Z

Context:
Add a graph-decomposition pass based on the device toml backend.

Description of the Change:
Automatically run the graph-decomposition pass based on the device's target gateset.

Tests in frontend for ensuring pass is added successfully through device preprocessing.
xDSL path does not parse graph-decomposition correctly when offloading non-xDSL passes to Catalyst within the transform.sequence

Benefits:

User does not need to specify the target gateset to run graph-decomposition.

Possible Drawbacks:

pytests need to be updated to reflect when device preprocessing is necessary for the test.

Related GitHub Issues:

[sc-129198]

codecov · 2026-10-26T21:04:29Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.25%. Comparing base (b8f71a2) to head (cfa577a).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #3275   +/-   ##
=======================================
  Coverage   96.25%   96.25%           
=======================================
  Files         177      177           
  Lines       21699    21734   +35     
  Branches     2162     2168    +6     
=======================================
+ Hits        20886    20921   +35     
  Misses        629      629           
  Partials      184      184

☔ View full report in Codecov by Harness.
📢 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.

albi3ro · 2026-10-27T13:39:27Z

Now that there's more decomposition transforms, things started checking the wrong qnode. So just need to anchor each qnode.

albi3ro · 2026-10-27T13:40:23Z

Just using skip_preprocess on all the drawing tests. XDSL can be a follow up feature if needed. But drawing can't handle all the new classical processing introduced, like SliceOp.

albi3ro · 2026-10-27T13:40:44Z

Same as above.

albi3ro · 2026-10-27T13:42:15Z

Something about functions with both quantum and classical instructions causes differentiation to fail.

albi3ro · 2026-10-27T13:43:37Z

So gridsynth currently gets run on all the decomposition rules in addition to the qnode body. This makes it so mixed-fallback introduces mcm's inside adjoint regions, which then causes errors in the decomposition pass.

albi3ro · 2026-10-27T13:44:23Z

    pipe = [("pipe", ["quantum-compilation-stage"])]

-    @qjit(pipelines=pipe, target="mlir", capture=True)
+    @qjit(pipelines=pipe, target="mlir", capture=True, skip_preprocess=True)


So none of the PPR ops are actually supported by null qubit 🤷

https://xanaduhq.slack.com/archives/C030UB33QH3/p1790965280121669

I think this is relevant here, although I know @albi3ro had concerns on whether this was the most appropriate convention.

albi3ro · 2026-10-27T13:45:05Z



-def test_controlled_pauli_rot_failure():
+def test_controlled_pauli_rot_works():


🎉 my favourite type of test update.

mudit2812

Looks great to me! Just one comment regarding the gridsynth tests

mudit2812

🎉

maliasadi · 2026-10-29T19:37:46Z


-        pipeline = _create_mlir_cli_schedule(pass_ops)
-        modified = _quantum_opt(*pipeline, "-mlir-print-op-generic", stdin=buffer.getvalue())
+        modified = _quantum_opt(


Compared to the previous (no-capture) pipeline, decomposing circuits after measurement transforms is right, but is decomposing after validation and after dynamic-one- shot okay? Validation is a no-op under capture rn, so that part doesn't break anything for now. But any rule that adds mcms would bypass the one-shot handling.

maliasadi · 2026-10-29T19:42:16Z

 def test_paramshift_with_gates(gate, state, capture_mode):
    """Test parameter shift works with a variety of gates present in the circuit."""

+    if gate is qp.BasisState and capture_mode:


Any plans to fix this? BasisState doesn't need to be decomposed as Lightning devices can support these ops re to capabilities.initial_state_prep. We also only test it with param-shift rn, could you test the device adjoint?

So the problem is that BasisState had different interpretation in pennylane and catalyst. One was more of a "xor bit flip" interpretation, and one was a more "prepare this state" interpretation. Since existing templates where using it in the "xor bit flip" interpretation, we couldn't lower it to a state prep instruction.

As for fixing gradients, something about functions with both quantum and classical stuff in it is causing issues. I looked into fixing it via an inlining pass, but we can only inline functions with quantum instructions if the function has a target gate attribute. As far as I can tell, no one knows why and how this restriction got added.

BasisState had different interpretation in pennylane and catalyst

I don't think this is a difference in PL vs Catalyst, Catalyst simply reimplemented the distinction already in PL with regards to the "initial state prep" in a program. Catalyst relied on the decomposition in the frontend to select when to decompose and when not, and any remaining state prep were assumed to be of the "initial" type.

With the decomposition moved to the compiler we'd have to replicate this decision somehow, but the graph isn't contextual at the moment. But in principal the quantum.operator version of BasisState could be the generic one (i.e. unitary semantics) while the dedicated ops could be reserved for the initial kind (i.e. set state semantics).

maliasadi · 2026-10-29T19:44:53Z



+# pylint: disable=unused-argument
+def _gateset_preprocessing(


the gateset literally ignores the gradient method / compatible gates. In the current no-capture pipeline), when a gradient is requested, the old frontend catalyst_acceptance also decomposes ops the device can't differentiate. On lightning those are Rot, QubitUnitary, PauliRot, etc.

Yep. This also literally breaks gradient support as a whole. So we are going to have to come back to gradients again whenever product decides it is a priority.

maliasadi · 2026-10-29T19:45:24Z

    """

-    @qjit(capture=True, collect_decomp_rules=False)
+    @qjit(capture=True, collect_decomp_rules=False, skip_preprocess=True)


Why is this required?

Because otherwise it will try to decompose BasisState, and we can't handle a decomposition rule on more than one quantum register. hence why have to go back to the drawing board on how allocation works.

maliasadi · 2026-10-29T19:45:55Z

    object code or IR. The default is stdout for IR and `<module-name>.o` for object code.
    Stdout (i.e. `-o -`) is rejected when the output is object code.

+* Decomposition to the device's gateset is now applied by default with `capture=True`.


Suggested change

* Decomposition to the device's gateset is now applied by default with `capture=True`.

* Decomposition to the device's gate set is now applied by default with `capture=True`.

maliasadi · 2026-10-29T19:46:48Z

+  This may introduce performance slowdowns or failures with cases that are not yet supported by the
+  decomposition pass. In the event of issues, this can be turned off with `skip_preprocess=True` being
+  passed to `qjit`.
+  [(#3275)](https://github.com/PennyLaneAI/catalyst/pull/3275)


should we add it to "Breaking Changes" as it breaks the default behaviour? 🤔

added to breaking changes.

… args from xdsl

…has no following block

…support with device gateset

…ified Compiler (#3276) ### Before submitting Please complete the following checklist when submitting a PR: - [ ] All new functions and code must be clearly commented and documented. - [ ] Ensure that code is properly formatted by running `make format`. The latest version of black and `clang-format-20` are used in CI/CD to check formatting. - [ ] All new features must include a unit test. Integration and frontend tests should be added to [`frontend/test`](../frontend/test/), Quantum dialect and MLIR tests should be added to [`mlir/test`](../mlir/test/), and Runtime tests should be added to [`runtime/tests`](../runtime/tests/). When all the above are checked, delete everything above the dashed line and fill in the pull request template. ------------------------------------------------------------------------------------------------------------ **Context:** **Description of the Change:** **Benefits:** **Possible Drawbacks:** **Related GitHub Issues:** Used Cursor --------- Co-authored-by: Nikhil Sreekumar <nikhil.sreekumar@xanadu.ai> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: Christina Lee <chrissie.c.l@gmail.com>

yde773786 mentioned this pull request Sep 30, 2026

Decomp/populate gateset from backend toml #3221

Closed

yde773786 marked this pull request as ready for review October 15, 2026 15:18

yde773786 force-pushed the decomp/auto-gateset-from-backend-toml branch from 098d5bf to 380dafb Compare October 16, 2026 17:06

albi3ro reviewed Oct 26, 2026

View reviewed changes

Comment thread frontend/test/pytest/test_decomposition.py Resolved

albi3ro reviewed Oct 27, 2026

View reviewed changes

Comment thread frontend/test/pytest/python_interface/inspection/test_mlir_graph.py

albi3ro Oct 27, 2026

Copy link
Copy Markdown

Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Same as above.

albi3ro reviewed Oct 27, 2026

View reviewed changes

mudit2812 reviewed Oct 28, 2026

View reviewed changes

Comment thread frontend/test/pytest/test_gridsynth.py Resolved

mudit2812 reviewed Oct 28, 2026

View reviewed changes

albi3ro added this pull request to stack #3415 October 29, 2026 15:05

albi3ro force-pushed the decomp/auto-gateset-from-backend-toml branch from 3f63e4a to 8b7e753 Compare October 29, 2026 15:25

mudit2812 approved these changes Oct 29, 2026

View reviewed changes

Comment thread frontend/test/pytest/test_gridsynth.py Resolved

maliasadi requested changes Oct 29, 2026

View reviewed changes

albi3ro requested a review from maliasadi October 29, 2026 20:06

albi3ro force-pushed the decomp/auto-gateset-from-backend-toml branch 2 times, most recently from 880a02e to 8790fae Compare October 30, 2026 17:32

yde773786 and others added 9 commits October 30, 2026 13:35

draft of implementation + tests

5056ac0

lazily apply transform sequence in catalyst rather than converting to…

1866910

… args from xdsl

transform block insertion point should remain valid even if original …

93c58a6

…has no following block

parity_synth filechecks shouldn't be affected by grph_decomp

12742cb

apply transformation sequence through xdsl run successfully now

2d157d7

intersect with runtime operations

4369999

use existing utility (from former frontend) for intersecting runtime …

6620053

…support with device gateset

accept incoming change

91d70ac

avoid inplace mutation of device

5ff05ad

albi3ro and others added 24 commits October 30, 2026 13:35

fix some xdsl tests

fe81fe0

more test fixes

a1a2722

accept incoming change

209f91d

update a few more tests

19ca54c

accept incoming change

e65d976

analyze and level=all-user, better xfailing for gridsynth and gradient

1ee6121

remove unnecessary control values patch

35ed814

maybe fixing something

661b696

test update

bd43ea4

accept incoming change

33d8f4c

fix lit test

189a331

format

150e4d8

accept incoming change

f7c88de

Update frontend/test/pytest/test_decomposition.py

e65aabf

test updates

d92fc1e

removing xfails now

b5ad19d

hopefully final test fixes

e81d87b

revert what I messed up

6f803a8

changelog merge

86a8940

Codex 6.1: remove collect_decomp_rules=False where needed

653c172

move symbol-dce pass till later so decomp rules not cleaned up

a5abd31

xfail more gridsynth tests

d7d2571

all the xfailing and skipping

578bc84

albi3ro force-pushed the decomp/auto-gateset-from-backend-toml branch from 8790fae to 578bc84 Compare October 30, 2026 17:39



		def test_controlled_pauli_rot_failure():
		def test_controlled_pauli_rot_works():



		# pylint: disable=unused-argument
		def _gateset_preprocessing(

	* Decomposition to the device's gateset is now applied by default with `capture=True`.
	* Decomposition to the device's gate set is now applied by default with `capture=True`.

Conversation

yde773786 commented Sep 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

codecov Bot commented Oct 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

albi3ro Oct 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

albi3ro Oct 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

mudit2812 left a comment

Choose a reason for hiding this comment

Uh oh!

mudit2812 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

dime10 Oct 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

albi3ro Oct 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

yde773786 commented Sep 25, 2026 •

edited

Loading

codecov Bot commented Oct 26, 2026 •

edited

Loading

albi3ro Oct 27, 2026 •

edited

Loading

albi3ro Oct 27, 2026 •

edited

Loading

dime10 Oct 29, 2026 •

edited

Loading

albi3ro Oct 29, 2026 •

edited

Loading