cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils
@Sebastian-Larsson @robell @rascani

Signed-off-by: Sebastian Larsson <sebastian.larsson@arm.com>
Co-authored-by: Sebastian Larsson <38941629+Sebastian-Larsson@users.noreply.github.com>

## Summary

- pin the tokenizers submodule to v1.5.0-rc2 while keeping the wheel
dependency at the latest published version, pytorch-tokenizers 1.4.1
- pin torchao to v0.18.0 and update the submodule to its release tag
- move TorchCodec, QNN, and ROCm release installs to stable packages and
production indexes
- restore the Qwen CUDA export matmul mode that TorchAO 0.18 no longer
sets globally
- update source, SwiftPM, and Raspberry Pi documentation for release/1.5
- accommodate the current Transformers Qwen3-VL annotations in mypy

## Validation

- python3 -m py_compile setup.py examples/models/qwen3_5_moe/export.py
backends/arm/test/models/Qwen3_VL/test_qwen3_vl_model.py
- Black 24.4.2 check for the changed Python files
- git diff --check
- verified pytorch-tokenizers 1.4.1 publishes wheels for Python
3.10-3.14 on Windows, Linux x86-64/aarch64, and macOS arm64
- verified the tokenizers and torchao gitlinks resolve exactly to
v1.5.0-rc2 and v0.18.0
- verified torch 2.14.0, torchao 0.18.0, and TorchCodec 0.16.0 resolve
from their configured indexes

## Release status

This stays draft while pytorch/test-infra#8780 blocks persistence of the
tokenizers RC2 wheel artifacts. Raise the package dependency to
pytorch-tokenizers>=1.5.0 after the final tokenizer wheels are promoted,
then verify that the final tag points to the RC2 commit or update the
submodule gitlink.

This PR was authored with Codex assistance.

…face (#22911)

Cherry-pick of #22908 onto `release/1.5`.

This exposes the thread count through the installed public parallel
interface, allowing downstream TorchAO code to avoid depending on the
private `extension/threadpool/threadpool.h` header.

Test plan:
- Original PR #22908 CI
- `python3 -m py_compile .ci/scripts/wheel/test_cpp_sdk.py`
- `git diff --check`

### Summary
Update CMSIS-NN to v8.0.0 and align the CMSIS pack dependency and
smoke-test versions. Stop Arm runner scripts from automatically
substituting the Ethos-U SDK checkout, so fresh runner builds use the
backend CMake pin. Remove the transpose-convolution ReLU and Hardtanh
expected failures resolved by the release's padding-row alignment fix.

### Test plan
The original update was validated with an explicitly selected v8.0.0 M55
runner: 51 passed, 7 xfailed in the focused Cortex-M suite. For the
runner-script correction, shell syntax, ShellCheck error checks, and git
diff checks pass. The CMake-selected v8 library cross-compiles for M55,
M7, and M0+ with the expected MVE, DSP, and scalar paths. Complete
runner builds and FVP tests have not been rerun for this correction.

Authored with OpenAI Codex.

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils
@Sebastian-Larsson @robell @rascani

Co-authored-by: RJ Ascani <rja@meta.com>

Allow TOSA support checks to distinguish between symbolic tensor
metadata and symbolic shape materialization.

Without the TOSA shape extension, symbolic tensor input/output shapes
can still be accepted for ops whose lowering does not need materialized
shape values. Reject symbolic SymInt arguments and the known shape-
dependent edge cases instead.

For targets that require fully resolved tensor shapes, add a separate
partitioner-level check that rejects nodes with unresolved tensor input
or output shapes.


cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils
@Sebastian-Larsson @robell @rascani

Signed-off-by: Oscar Andersson <oscar.andersson@arm.com>
Co-authored-by: Oscar Andersson <87121123+oscarandersson8218@users.noreply.github.com>

Eval-mode dropout can disappear before shared no-op cleanup runs,
leaving adjacent dequantize/quantize conversions. Promote NXP's
`FoldRedundantDequantizeQuantizePass` from #21016 into the shared
lowering sequence used by `to_edge` and `to_edge_transform_and_lower`.
Cortex-M MLPerf Tiny DS-CNN then finishes with one quantize and one
dequantize operation in both layouts, down from three of each.
MobileNetV2 similarly drops from two to one of each.

The shared `eliminate_dq_q` matcher remains unchanged, keeping this
release patch focused on promoting the existing cleanup. Remove the
stale dropout target from `RemoveNoopPass`. NXP retains a later
invocation of the shared pass because its auxiliary-operator splitting
can introduce new adjacent pairs. Update the Cortex-M model counts and
correct the clone regression's obsolete Q/DQ signatures so the test also
checks numerical equivalence.

Validation on PyTorch 2.14 CPU: 77 EXIR and clone-transform tests
passed, with 28 passing subtests and one expected failure. The dropout
regression covers both edge APIs, per-tensor and per-channel
quantization, and int8/uint8/int16/int32. All 25 selected Cortex-M
DS-CNN, MobileNetV2, and explicit-layout tests passed, including
Corstone-300 FVP implementation tests. Additional explicit-layout DS-CNN
and legacy `ExirExportedProgram.to_edge` numerical checks passed.
Formatting, flake8, and `git diff --check` passed.

Authored with AI assistance from OpenAI Codex.

Co-authored-by: RJ Ascani <rja@meta.com>

## Summary

Bump the ExecuTorch release version from 1.5.0 to 1.5.1 and update the
stable SwiftPM documentation examples. This mirrors the version and
documentation changes made for the 1.4.1 patch release. The dated
nightly SwiftPM example remains unchanged.

## Validation

`git diff --check`

Authored with AI assistance from OpenAI Codex.

Support dynamic convolutions and pooling operations without the TOSA
shape extension when partitioning proves lowering needs neither input-
size adjustment nor dynamic padding.

Recognize value-only max pooling in its exported with-indices form
before backend preprocessing canonicalizes it.

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils
@Sebastian-Larsson @robell @rascani

Signed-off-by: Oscar Andersson <oscar.andersson@arm.com>
Co-authored-by: Oscar Andersson <87121123+oscarandersson8218@users.noreply.github.com>