Bump mlir-aie to 1.4.4.dev18 and llvm-aie to 2026091401

[erwei-xilinx]

The pin has sat at 10767b50 (2026-09-03) while mlir-aie moved on. This takes it to f5196d8 — the newest released wheel, 1.4.4.dev18, 2026-09-18 — and moves llvm-aie with it, as the comment beside that pin asks (2026090201+a36c62b9 → 2026091401+a5d028e8, from mlir-aie's own utils/peano-requirements.txt).

LLVM does not move. mlir-aie at f5196d8 still pins LLVM_PROJECT_COMMIT 56bcc187 / DATETIME 2026080106, exactly what utils/clone-llvm.sh already carries. So this is a pure mlir-aie + llvm-aie step, four lines across four files.

MLIR_PYTHON_EXTRAS_SHORTHASH is left alone on purpose — nothing in the tree reads it.

Validation

No AIR source changes were needed. Built against the new wheel:

• check-air-mlir 539 passed / 0 failed (7 expected-fail, 11 unsupported)
• check-air-cpp 1/1
• both designs on the memtile chain-lock path — programming_examples/fused_decode and fused_decode_ple — compile clean

One behaviour change worth naming

mlir-aie #3702 ("Avoid sharing a stream-switch arbiter between packet flows that can deadlock", merged 2026-09-10) lands in this range, and it interacts with the open gemma4-e2b decode hang.

It diagnoses that hang precisely. Run its pathfinder over gemma4's device region and it reports the hazard at tile (2, 0), and again at (2, 1) and (2, 2):

at tile (2, 0), packet flow 12 shares arbiter 1 with packet flow 29,
  which it can deadlock against; no independent arbiter is free

It cannot repair it: column 2 is oversubscribed — 9 masters against 6 arbiters — so the pass takes its documented "no independent arbiter is free" fallback, keeps the assignment and warns.

What it does do is reshuffle. At (2, 0) the four projection-gather flows were paired arb1={W0,W2}, arb2={W1,W3}; now it is arb1={W0, pkt12} and arb2={W1,W2,W3}. The gather's lock chain runs W0 → W1 → W2 → W3, so one 3-way hazard has more ways to close a cycle than two 2-way ones — and it measures that way on an NPU2 Krackan:

toolchain	default pmode	turbo
pinned (1.4.3.dev55)	3–7 / 20	4 / 20
bumped (1.4.4.dev18)	0 / 20	0 / 20

0 of 60 dispatches across both power modes, against a 15–35% pass rate before. That path is already --expect-fail at every context in gemma4_e2b_q4nx/run_npu2_sweep.lit and is tracked separately, so no gate changes state — but a reviewer weighing this bump should know it moves an open defect the wrong way, and that what it buys in exchange is a compile-time warning naming the tile instead of a silent hang.

I'd still take the bump: the regression is confined to an already-failing path, and the diagnostics are the thing that makes that path fixable. Happy to hold it if you'd rather land the gemma4 fix first.

🤖 Generated with Claude Code

[Copilot]

Copilot review overview

🔵 Needs a closer look

The bump changes stream-switch behavior on an already-failing gemma4 path, warranting final human review.

Review effort: Lite
Findings: None

What changed in this PR

Updates MLIR-AIE and LLVM-AIE pins for local builds and CI while keeping LLVM unchanged.

Changes:

• Bumps MLIR-AIE to 1.4.4.dev18.
• Updates LLVM-AIE to 2026091401.
• Synchronizes build scripts and CI workflows.

File	Description
utils/​clone-mlir-aie.sh	Updates MLIR-AIE source and wheel pins.
utils/​build-mlir-air-using-wheels.sh	Updates the LLVM-AIE wheel pin.
.github/​workflows/​nightlyPerfBenchmark.yml	Aligns nightly benchmarking with the new dependency.
.github/​workflows/​buildAndTestRyzenAI.yml	Aligns Ryzen AI CI with the new dependency.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.