Support magic-T allocation via qp.allocate (MLIR + capture path) by Qottmann · Pull Request #3029 · PennyLaneAI/catalyst

Qottmann · 2026-07-16T15:08:25Z

Summary

Adds capture-path (qjit(capture=True)) support for qp.allocate(..., state="magic-T"|"magic-T-adj"), pairing with the merged PennyLaneAI/pennylane#9846. Magic states lower through qref/PBC to pbc.fabricate (H+T / H+T†).

This is PR 1 of 2. #3027 adds all legacy (capture=False) tracing on top of this PR.

User-facing state strings: "magic-T" and "magic-T-adj". MLIR/PBC enums remain magic / magic_conj; Catalyst maps at lowering time.

MLIR / PBC lowering

Add pbc.ref.fabricate for magic and magic-conjugate preparation
Lower pbc.ref.fabricate → pbc.fabricate in value-semantics conversion, with dead-anchor cleanup
Resource analysis + MLIR lit tests

Unified allocate/deallocate jaxpr

Capture frontend emits PennyLane allocate_prim / deallocate_prim
MLIR lowering in qref_jax_primitives.py selects qref.alloc vs pbc.ref.fabricate from state
Map "magic-T" / "magic-T-adj" → MLIR magic / magic_conj

Capture path only

qfunc_interpreter.py handles allocate/deallocate via unified jaxpr primitives
Device verification accepts new states

CI / deps

PennyLane 0.46.0.dev62 (.dep-versions, doc/requirements.txt)

Tests & docs (capture-only)

MLIR lit + Python lit (all capture=True)
Pytest: magic-T allocation, manual deallocate, MLIR lowering (capture=True only)
sharp_bits.rst + changelog

Review guide

Area	Files
MLIR	`mlir/include/PBC/IR/PBCRefOps.td`, `mlir/lib/QRef/Transforms/value_semantics_conversion.*`, `mlir/lib/PBC/IR/PBCOps.cpp`
JAX → MLIR	`frontend/catalyst/from_plxpr/qref_jax_primitives.py`
Capture	`frontend/catalyst/from_plxpr/qfunc_interpreter.py`, `frontend/catalyst/device/verification.py`

Test plan

MLIR lit tests
Capture lit + pytest magic-T tests
make format check=1 (clang-format, black, isort)
CI green before merging Complete legacy tracing for dynamic magic wires #3027

Own all jax_tracer legacy support on top of #3029: qref-mode dynamic wires, allocate/deallocate tracing, MCM/PPM routing, control-flow wire propagation, qref register lifetime, and ppr-to-ppm deallocation ordering. Includes legacy and dual-mode pytest coverage. Co-authored-by: Cursor <cursoragent@cursor.com>

Update SpecsResources field names and bump the PennyLane pin to 0.46.0.dev62 so CI uses the latest resource hierarchy. Co-authored-by: Cursor <cursoragent@cursor.com>

Keep #3029 focused on MLIR lowering and the capture path. Legacy tape tracing, non-capture tests, and jax_tracer qref-mode changes move to #3027. Co-authored-by: Cursor <cursoragent@cursor.com>

Own all jax_tracer legacy support on top of #3029: qref-mode dynamic wires, allocate/deallocate tracing, MCM/PPM routing, control-flow wire propagation, qref register lifetime, and ppr-to-ppm deallocation ordering. Includes legacy and dual-mode pytest coverage. Co-authored-by: Cursor <cursoragent@cursor.com>

Restore capture=False allocate error coverage, add negative deallocate tests, consolidate deallocate guard logic, document staged fabricate primitives for Co-authored-by: Cursor <cursoragent@cursor.com> #3027, and fix the sharp_bits.rst cross-reference syntax.

Own all jax_tracer legacy support on top of #3029: qref-mode dynamic wires, allocate/deallocate tracing, MCM/PPM routing, control-flow wire propagation, qref register lifetime, and ppr-to-ppm deallocation ordering. Includes legacy and dual-mode pytest coverage. Co-authored-by: Cursor <cursoragent@cursor.com>

Update resource specs tests, Operator2 PLXPR expectations, and decomposition graph handling for abstract Adjoint/Controlled resource reps introduced with magic-T and the new resource API. Co-authored-by: Cursor <cursoragent@cursor.com>

dime10

Nice work! I think we just need to sort out the jax primitive situation

Skip C(...) decomposition-graph nodes for Operator2 controlled reps, restore exact SpecsResources type equality in specs tests, and add a changelog entry. Co-authored-by: Cursor <cursoragent@cursor.com>

Test-suite and dependency-pin updates are not user-facing release notes. Co-authored-by: Cursor <cursoragent@cursor.com>

Own all jax_tracer legacy support on top of #3029: qref-mode dynamic wires, allocate/deallocate tracing, MCM/PPM routing, control-flow wire propagation, qref register lifetime, and ppr-to-ppm deallocation ordering. Includes legacy and dual-mode pytest coverage. Co-authored-by: Cursor <cursoragent@cursor.com>

When controls are applied via qfunc form (e.g. qp.ctrl(qp.S, wire)(0)), PennyLane stores control wires on the interpreter rather than in operator_p.n_ctrls. Merge self.control_wires into handle_operator and add regression tests from #3035. Co-authored-by: Cursor <cursoragent@cursor.com>

Lower state="magic" and state="magic_conj" to pbc.ref.fabricate with qref deallocation, and enable the legacy tape path for dynamic wire allocation. Co-authored-by: Cursor <cursoragent@cursor.com>

Erase dead pbc.ref.fabricate and qref.alloc_qb anchor ops at the end of conversion so they do not reach LLVM lowering, and update TestPBC checks. Co-authored-by: Cursor <cursoragent@cursor.com>

Fix import ordering, apply black formatting, and add a changelog entry. Co-authored-by: Cursor <cursoragent@cursor.com>

Co-authored-by: Cursor <cursoragent@cursor.com>

Both capture and legacy frontends now emit allocate/deallocate at the jaxpr level; MLIR lowering selects qref.alloc or pbc.ref.fabricate from the state argument. Co-authored-by: Cursor <cursoragent@cursor.com>

Map the new Python state strings to existing MLIR magic enums during allocate lowering and update tests and docs accordingly. Co-authored-by: Cursor <cursoragent@cursor.com>

Co-authored-by: Cursor <cursoragent@cursor.com>

Keep #3029 focused on MLIR lowering and the capture path. Legacy tape tracing, non-capture tests, and jax_tracer qref-mode changes move to #3027. Co-authored-by: Cursor <cursoragent@cursor.com>

Restore capture=False allocate error coverage, add negative deallocate tests, consolidate deallocate guard logic, document staged fabricate primitives for Co-authored-by: Cursor <cursoragent@cursor.com> #3027, and fix the sharp_bits.rst cross-reference syntax.

Describe the capture-path allocate feature in terms of user benefit rather than implementation details, per changelog guidelines. Co-authored-by: Cursor <cursoragent@cursor.com>

Graph decomposition tests used state="any" and restored=True, which now correctly raise CompileError per review feedback; zero-state allocation matches the previous Catalyst behaviour for these cases. Co-authored-by: Cursor <cursoragent@cursor.com>

Match the style of other tests in test_from_plxpr.py per review feedback. Co-authored-by: Cursor <cursoragent@cursor.com>

…AI/catalyst into qottmann/allocate-magic-core

dime10

Excellent work 💯

I'm happy with it, leaving some final comments. @paul0403 Could you provide a second review? :)

## Summary PennyLane **dev62** is the minimum pin that includes both [PennyLaneAI/pennylane#9846](PennyLaneAI/pennylane#9846) (`magic-T`) and [PennyLaneAI/pennylane#9841](PennyLaneAI/pennylane#9841) (resource API). This PR bumps Catalyst to that pin and fixes the resulting fallout. **Resource API migration (#9841)** - Bump PennyLane to `0.46.0.dev62` in `.dep-versions` and `doc/requirements.txt` - Replace `SymbolicSpecsResources` checks with `SpecsResources.is_symbolic` - Rename resource fields in tests: `gate_types`/`gate_counts` → `quantum_operations`, `measurements` → `measurement_processes`, `num_gates` → `total_quantum_operations` - Drop `gate_sizes` assertions (no longer tracked) - Handle `PBCSpecsResources` in PPR specs tests (depth fields) **dev62 fallout (Operator2 / magic-T)** - `decompose.py`: safe `.params` access for Operator2 ops; skip abstract `Adjoint(...)` / `Controlled(...)` decomposition-graph nodes that lack `op_type` - PLXPR tests: single-qubit ops like `S`/`T` now lower to `qref_operator` instead of `qref_qinst` - `test_adjoint.py`: replace removed `qp.Rotation` with `qp.Rot` - `test_quantum_control.py`: compare `ControlledOp2` decompositions with `qp.equal` - `test_specs.py`: normalize doubly-controlled `C(S)` counts in `test_complex` This is the base branch for the magic-state allocation stack (`#3029` → `#3027`). --------- Co-authored-by: Cursor <cursoragent@cursor.com>

…ottmann/allocate-magic-core

Document static num_wires in allocate lowering, drop redundant magic MLIR pytest covered by lit, and add value-semantics CHECKs for magic_conj fabricate. Co-authored-by: Cursor <cursoragent@cursor.com>

Mirror handleRefFabricate in reference_semantics_conversion and cover magic and magic_conj in Quantum SemanticsConversion lit tests. Co-authored-by: Cursor <cursoragent@cursor.com>

paul0403

Thanks! Almost there

Use quantum.custom result wires in reference-semantics conversion tests and tie ref-side CHECKs to the fabricated qubit in QRef fabricate lit tests. Co-authored-by: Cursor <cursoragent@cursor.com>

Co-authored-by: Cursor <cursoragent@cursor.com>

Deallocate the PauliX result wire after ref-to-value conversion, matching value-semantics gate behaviour. Co-authored-by: Cursor <cursoragent@cursor.com>

Use state="zero" in decomposition lit tests, ignore the restored flag in handle_allocate so PennyLane work-wire decomps compile, and align docs/tests. Co-authored-by: Cursor <cursoragent@cursor.com>

codecov · 2026-07-24T14:48:11Z

Codecov Report

❌ Patch coverage is 90.00000% with 6 lines in your changes missing coverage. Please review.
✅ Project coverage is 96.94%. Comparing base (2e39af6) to head (0ae7ec7).
⚠️ Report is 6 commits behind head on main.

Files with missing lines	Patch %	Lines
frontend/catalyst/device/verification.py	25.00%	2 Missing and 1 partial ⚠️
...rontend/catalyst/from_plxpr/qref_jax_primitives.py	95.83%	1 Missing and 1 partial ⚠️
frontend/catalyst/device/qjit_device.py	50.00%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3029      +/-   ##
==========================================
- Coverage   96.97%   96.94%   -0.03%     
==========================================
  Files         165      165              
  Lines       19111    19160      +49     
  Branches     1819     1831      +12     
==========================================
+ Hits        18532    18575      +43     
- Misses        424      427       +3     
- Partials      155      158       +3

☔ 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.

paul0403

Awesome! 🥳

albi3ro · 2026-08-05T13:34:20Z

+  - The ``state="any"`` configuration is not supported by Catalyst and will raise a
+    compilation error. The ``restored`` keyword is ignored. Catalyst supports
+    ``state="zero"`` initialization (the default), as well as magic state
+    initialization (see below).


Why can't we just continue treating state="any" as state="zero"? A little less efficient, but should still work?

This was referenced Jul 16, 2026

Complete legacy tracing for dynamic magic wires #3027

Open

Add frontend lowering for pbc.fabricate #3019

Closed

Qottmann changed the title ~~Support magic state allocation via qp.allocate (MLIR + capture path)~~ Support magic-T allocation via qp.allocate (MLIR + capture path) Jul 22, 2026

Migrate specs tests to PennyLane #9841 resource API.

1a671fb

Update SpecsResources field names and bump the PennyLane pin to 0.46.0.dev62 so CI uses the latest resource hierarchy. Co-authored-by: Cursor <cursoragent@cursor.com>

Qottmann force-pushed the qottmann/allocate-magic-core branch from efcea1a to 541804b Compare July 22, 2026 10:49

Qottmann changed the base branch from main to qottmann/fix-resources July 22, 2026 10:49

dime10 requested changes Jul 22, 2026

View reviewed changes

Qottmann and others added 3 commits July 23, 2026 12:02

Merge branch 'main' into qottmann/fix-resources

d59b449

Address PR #3050 review feedback.

e2735b4

Skip C(...) decomposition-graph nodes for Operator2 controlled reps, restore exact SpecsResources type equality in specs tests, and add a changelog entry. Co-authored-by: Cursor <cursoragent@cursor.com>

Remove internal CI changelog entry from #3050.

2864e3a

Test-suite and dependency-pin updates are not user-facing release notes. Co-authored-by: Cursor <cursoragent@cursor.com>

Qottmann and others added 11 commits July 23, 2026 12:20

Support magic state allocation via qp.allocate in Catalyst.

6d56c17

Lower state="magic" and state="magic_conj" to pbc.ref.fabricate with qref deallocation, and enable the legacy tape path for dynamic wire allocation. Co-authored-by: Cursor <cursoragent@cursor.com>

Fix ref.fabricate cleanup after value-semantics conversion.

3bfd753

Erase dead pbc.ref.fabricate and qref.alloc_qb anchor ops at the end of conversion so they do not reach LLVM lowering, and update TestPBC checks. Co-authored-by: Cursor <cursoragent@cursor.com>

Address review feedback for magic state allocation.

413e224

Fix import ordering, apply black formatting, and add a changelog entry. Co-authored-by: Cursor <cursoragent@cursor.com>

Apply clang-format formatting fixes.

7ee34be

Co-authored-by: Cursor <cursoragent@cursor.com>

Unify dynamic allocation on allocate/deallocate jaxpr primitives.

96a978b

Both capture and legacy frontends now emit allocate/deallocate at the jaxpr level; MLIR lowering selects qref.alloc or pbc.ref.fabricate from the state argument. Co-authored-by: Cursor <cursoragent@cursor.com>

Rename magic allocate states to magic-T and magic-T-adj.

a898e20

Map the new Python state strings to existing MLIR magic enums during allocate lowering and update tests and docs accordingly. Co-authored-by: Cursor <cursoragent@cursor.com>

Apply black and isort formatting fixes.

0b32030

Co-authored-by: Cursor <cursoragent@cursor.com>

Move legacy jax_tracer support to the stacked follow-up PR.

bf1a7c4

Keep #3029 focused on MLIR lowering and the capture path. Legacy tape tracing, non-capture tests, and jax_tracer qref-mode changes move to #3027. Co-authored-by: Cursor <cursoragent@cursor.com>

Rewrite magic-T changelog entry for user-facing style.

391268d

Describe the capture-path allocate feature in terms of user benefit rather than implementation details, per changelog guidelines. Co-authored-by: Cursor <cursoragent@cursor.com>

Qottmann force-pushed the qottmann/allocate-magic-core branch from bb4d012 to 391268d Compare July 23, 2026 10:34

Qottmann commented Jul 23, 2026

View reviewed changes

Comment thread frontend/test/pytest/from_plxpr/test_decompose_transform.py

Qottmann and others added 4 commits July 23, 2026 17:05

remove redundant test

721e754

Use direct jaxpr indexing in Operator2 control context test.

e50d016

Match the style of other tests in test_from_plxpr.py per review feedback. Co-authored-by: Cursor <cursoragent@cursor.com>

Merge branch 'qottmann/fix-resources' of https://github.com/PennyLane…

fc393a8

…AI/catalyst into qottmann/allocate-magic-core

remove pbc::FabricateOp again

3110552

dime10 approved these changes Jul 23, 2026

View reviewed changes

This was referenced Jul 23, 2026

3029 follow-up: Eliminate the rebinding of allocation primitives #3055

Open

3029 follow-up: Update device capabilities for dynamic allocation #3056

Open

dime10 requested a review from paul0403 July 23, 2026 15:30

Qottmann added 3 commits July 23, 2026 17:40

remove pbc::RefFabricateOp

da1d86a

changelog

2f62450

sharp bits2

a824160

paul0403 reviewed Jul 23, 2026

View reviewed changes

Comment thread mlir/lib/QRef/Transforms/value_semantics_conversion.h

Base automatically changed from qottmann/fix-resources to main July 23, 2026 18:16

Qottmann and others added 4 commits July 24, 2026 12:28

Merge branch 'main' of https://github.com/PennyLaneAI/catalyst into q…

a9bd6e8

…ottmann/allocate-magic-core

Address PR review follow-ups and merge main after #3050.

6b0e72d

Document static num_wires in allocate lowering, drop redundant magic MLIR pytest covered by lit, and add value-semantics CHECKs for magic_conj fabricate. Co-authored-by: Cursor <cursoragent@cursor.com>

remove unused import

222cfef

Add value-to-ref conversion for pbc.fabricate.

007c66d

Mirror handleRefFabricate in reference_semantics_conversion and cover magic and magic_conj in Quantum SemanticsConversion lit tests. Co-authored-by: Cursor <cursoragent@cursor.com>

paul0403 reviewed Jul 24, 2026

View reviewed changes

Comment thread mlir/test/Quantum/SemanticsConversion/TestPBC.mlir

Address TestPBC review feedback on value-semantics gate outputs.

a4b5759

Use quantum.custom result wires in reference-semantics conversion tests and tie ref-side CHECKs to the fabricated qubit in QRef fabricate lit tests. Co-authored-by: Cursor <cursoragent@cursor.com>

paul0403 reviewed Jul 24, 2026

View reviewed changes

Comment thread mlir/test/QRef/SemanticConversion/TestPBC.mlir Outdated

Qottmann and others added 3 commits July 24, 2026 15:21

Apply clang-format to reference_semantics_conversion.h.

754c69e

Co-authored-by: Cursor <cursoragent@cursor.com>

Fix QRef TestPBC CHECKs for value-semantics custom outputs.

ca22545

Deallocate the PauliX result wire after ref-to-value conversion, matching value-semantics gate behaviour. Co-authored-by: Cursor <cursoragent@cursor.com>

Fix frontend CI after qp.allocate validation changes.

0ae7ec7

Use state="zero" in decomposition lit tests, ignore the restored flag in handle_allocate so PennyLane work-wire decomps compile, and align docs/tests. Co-authored-by: Cursor <cursoragent@cursor.com>

paul0403 approved these changes Jul 24, 2026

View reviewed changes

albi3ro reviewed Aug 5, 2026

View reviewed changes

Uh oh!

Conversation

Qottmann commented Jul 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

MLIR / PBC lowering

Unified allocate/deallocate jaxpr

Capture path only

CI / deps

Tests & docs (capture-only)

Review guide

Test plan

Related

Uh oh!

dime10 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

dime10 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

paul0403 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

codecov Bot commented Jul 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

paul0403 left a comment

Choose a reason for hiding this comment

Uh oh!

albi3ro Aug 5, 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!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Qottmann commented Jul 16, 2026 •

edited

Loading

codecov Bot commented Jul 24, 2026 •

edited

Loading

albi3ro Aug 5, 2026 •

edited

Loading