Add Windows wheel release job to nightly-wheels CI

[erwei-xilinx]

Summary

• Adds a parallel build-wheels-windows job to nightly-wheels.yml that builds and releases triton-xdna Windows x64 wheels alongside the existing Linux job. Both publish to the same latest-wheels release tag.
• Python matrix is 3.10 / 3.11 / 3.12 — narrower than the Linux matrix (3.10–3.14) because Xilinx publishes mlir-air Windows wheels only for those three Pythons today.
• XRT Windows SDK is downloaded from the pinned Xilinx/XRT 2.21.75 release (xrt_windows_sdk.zip) and extracted to C:\Program Files\AMD\xrt, where the existing Windows build infrastructure (utils/env_setup.ps1, setup.py) expects it.
• Fixes three stale README items uncovered while writing the workflow:
  • Quick Start referenced .\utils\build_windows.ps1 which was removed during the windows-build-minimal PR cleanup; replaced with the env_setup.ps1 path that does exist.
  • Manual Build claimed mlir-air must be built from source; replaced with the mlir_air[aie] pip-install command (matches what env_setup.ps1 actually does).
  • Python version requirement updated from "3.12+" to the accurate "3.10, 3.11, or 3.12" range.

Caveats reviewers should know

• Untested in CI. I can't run windows-latest from a Linux dev box. The first nightly is likely to need at least one fix — common things to watch for: MSVC version mismatch with triton-windows's expected toolchain, delvewheel repair failing on missing DLLs, or paths in setup.py:_build_triton_windows needing tweaks under cibuildwheel's environment.
• No repair-wheel-command set under [tool.cibuildwheel.windows] in pyproject.toml. cibuildwheel runs delvewheel repair by default; if our wheel bundles XRT-linked binaries we may need exclusions analogous to the Linux auditwheel step. Worth watching the first run's logs.
• setup_xrt_dev.ps1 is intentionally not invoked — that script is for stripping headers + generating .lib from a runtime-only xrt_coreutil.dll (the Ryzen AI SDK case). The full xrt_windows_sdk.zip already includes both, so it's not needed here.

Test plan

• CI: green Linux build job (regression check — should be unaffected)
• CI: green Windows build job for cp310/cp311/cp312
• Verify Windows wheels appear on the latest-wheels release alongside Linux wheels
• On a Windows machine, install the released wheel and run an example kernel

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR extends the nightly wheels CI workflow to build and publish Windows x64 wheels in parallel with the existing Linux wheel builds, and updates the README’s Windows instructions to match the intended Windows wheel availability and setup flow.

Changes:

• Adds a build-wheels-windows matrix job (cp310/cp311/cp312) to .github/workflows/nightly-wheels.yml, including XRT Windows SDK download/extract and wheel publishing to the shared latest-wheels release tag.
• Updates README Windows requirements and setup steps (Python version range, Quick Start script path, and mlir-air install instructions).
• Updates Windows “Known Limitations” documentation to reflect Python version constraints for Xilinx Windows wheels.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
README.md	Updates Windows requirements and build/setup instructions to align with Windows wheel support.
.github/workflows/nightly-wheels.yml	Adds a Windows wheel build + release job alongside the existing Linux workflow.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[astrelsky]

@erwei-xilinx fwiw I installed the released wheel an did not encounter any problems with msvc version (I only have the Visual Studio 2026 Build Tools installed now). I did encounter one unexpected failure, but when I reran that specific test it passed, so it's probably just a bit flaky and not much of a concern.

(venv) D:\winxdna>scripts\run_tests -v --device aie2p
Starting example test run...
Examples dir: D:\winxdna\examples
Target device: aie2p
Transform file: transform_aie2p.mlir
--------------------------------------------------
📁 Example: autotune-matmul
   ⏭️  SKIP: transform_aie2p.mlir not found for device aie2p

📁 Example: average_pool
   transform_aie2p.mlir detected; will set AIR_TRANSFORM_TILING_SCRIPT
   🔄 Running: average_pool.py
Command: D:\winxdna\venv\Scripts\python.exe average_pool.py
   ✅ PASS: average_pool.py

📁 Example: axpy
   transform_aie2p.mlir detected; will set AIR_TRANSFORM_TILING_SCRIPT
   🔄 Running: axpy.py
Command: D:\winxdna\venv\Scripts\python.exe axpy.py
   ✅ PASS: axpy.py

📁 Example: gelu
   transform_aie2p.mlir detected; will set AIR_TRANSFORM_TILING_SCRIPT
   🔄 Running: gelu.py
Command: D:\winxdna\venv\Scripts\python.exe gelu.py
   ✅ PASS: gelu.py

📁 Example: leaky_relu
   transform_aie2p.mlir detected; will set AIR_TRANSFORM_TILING_SCRIPT
   🔄 Running: leaky_relu.py
Command: D:\winxdna\venv\Scripts\python.exe leaky_relu.py
   ✅ PASS: leaky_relu.py

📁 Example: matmul_bf16_m64_n64_k64
   transform_aie2p.mlir detected; will set AIR_TRANSFORM_TILING_SCRIPT
   🔄 Running: matmul_bf16_m64_n64_k64.py
Command: D:\winxdna\venv\Scripts\python.exe matmul_bf16_m64_n64_k64.py
   ✅ PASS: matmul_bf16_m64_n64_k64.py

📁 Example: matmul_f32_m64_n32_k16_padded_atransposed
   transform_aie2p.mlir detected; will set AIR_TRANSFORM_TILING_SCRIPT
   🔄 Running: matmul_f32_m64_n32_k16_padded_atransposed.py
Command: D:\winxdna\venv\Scripts\python.exe matmul_f32_m64_n32_k16_padded_atransposed.py
stdout:
Mismatch at (470, 464): actual=0.0, expected=3947.583984375
Mismatch at (281, 388): actual=0.0, expected=4153.3486328125
Mismatch at (162, 92): actual=0.0, expected=4219.048828125
Mismatch at (375, 272): actual=0.0, expected=3944.739013671875
Mismatch at (47, 178): actual=0.0, expected=4247.2666015625
FAIL: 109/109 samples mismatched

   ❌ FAIL: matmul_f32_m64_n32_k16_padded_atransposed.py (exit code 1)

📁 Example: matmul_i8_m128_n64_k64
   transform_aie2p.mlir detected; will set AIR_TRANSFORM_TILING_SCRIPT
   🔄 Running: matmul_i8_m128_n64_k64.py
Command: D:\winxdna\venv\Scripts\python.exe matmul_i8_m128_n64_k64.py
   ✅ PASS: matmul_i8_m128_n64_k64.py

📁 Example: matmul_i8_m64_n64_k64
   transform_aie2p.mlir detected; will set AIR_TRANSFORM_TILING_SCRIPT
   🔄 Running: matmul_i8_m64_n64_k64.py
Command: D:\winxdna\venv\Scripts\python.exe matmul_i8_m64_n64_k64.py
   ✅ PASS: matmul_i8_m64_n64_k64.py

📁 Example: relu
   transform_aie2p.mlir detected; will set AIR_TRANSFORM_TILING_SCRIPT
   🔄 Running: relu.py
Command: D:\winxdna\venv\Scripts\python.exe relu.py
   ✅ PASS: relu.py

📁 Example: rms_norm
   transform_aie2p.mlir detected; will set AIR_TRANSFORM_TILING_SCRIPT
   🔄 Running: rms_norm.py
Command: D:\winxdna\venv\Scripts\python.exe rms_norm.py
stderr:
loc("-":83:11): error: application of transform.air.copy_to_dma expected to produce 1 results (actually produced 0).
loc("-":83:11): error: application of transform.air.copy_to_dma expected to produce 1 results (actually produced 0).

   ✅ PASS: rms_norm.py

📁 Example: sigmoid
   transform_aie2p.mlir detected; will set AIR_TRANSFORM_TILING_SCRIPT
   🔄 Running: sigmoid.py
Command: D:\winxdna\venv\Scripts\python.exe sigmoid.py
   ✅ PASS: sigmoid.py

📁 Example: silu
   transform_aie2p.mlir detected; will set AIR_TRANSFORM_TILING_SCRIPT
   🔄 Running: silu.py
Command: D:\winxdna\venv\Scripts\python.exe silu.py
   ✅ PASS: silu.py

📁 Example: swiglu
   transform_aie2p.mlir detected; will set AIR_TRANSFORM_TILING_SCRIPT
   🔄 Running: swiglu.py
Command: D:\winxdna\venv\Scripts\python.exe swiglu.py
   ✅ PASS: swiglu.py

📁 Example: test_layernorm
   transform_aie2p.mlir detected; will set AIR_TRANSFORM_TILING_SCRIPT
   🔄 Running: test_layernorm.py
Command: D:\winxdna\venv\Scripts\python.exe test_layernorm.py
   ✅ PASS: test_layernorm.py

📁 Example: test_softmax
   transform_aie2p.mlir detected; will set AIR_TRANSFORM_TILING_SCRIPT
   🔄 Running: test_softmax.py
Command: D:\winxdna\venv\Scripts\python.exe test_softmax.py
   ✅ PASS: test_softmax.py

📁 Example: vec-add
   transform_aie2p.mlir detected; will set AIR_TRANSFORM_TILING_SCRIPT
   🔄 Running: vec-add.py
Command: D:\winxdna\venv\Scripts\python.exe vec-add.py
   ✅ PASS: vec-add.py

📁 Example: weighted_rms_norm
   transform_aie2p.mlir detected; will set AIR_TRANSFORM_TILING_SCRIPT
   🔄 Running: weighted_rms_norm.py
Command: D:\winxdna\venv\Scripts\python.exe weighted_rms_norm.py
stderr:
Traceback (most recent call last):
  File "D:\winxdna\examples\weighted_rms_norm\weighted_rms_norm.py", line 95, in <module>
    bench_weighted_rms_norm(M, N, "test")
  File "D:\winxdna\examples\weighted_rms_norm\weighted_rms_norm.py", line 88, in bench_weighted_rms_norm
    torch.testing.assert_close(y, y_ref, atol=5e-1, rtol=1e-1)
  File "D:\winxdna\venv\Lib\site-packages\torch\testing\_comparison.py", line 1631, in assert_close
    raise error_metas[0].to_error(msg)
AssertionError: Tensor-likes are not close!

Mismatched elements: 1 / 16384 (0.0%)
Greatest absolute difference: 1.5625 at index (30, 128) (up to 0.5 allowed)
Greatest relative difference: 0.158203125 at index (30, 128) (up to 0.1 allowed)

   ❌ FAIL: weighted_rms_norm.py (exit code 1)

--------------------------------------------------
Test Results:
  ✅ Passed:  15
  ❌ Failed:  2
  ⏰ Timeouts: 0
  ⏭️  Skipped: 1
  📊 Total:   17
💔 2 failed, 0 timed out

[erwei-xilinx]

Thanks for the test run, @astrelsky! Two real bugs surfaced from this:

weighted_rms_norm (1/16384 elements) — Investigated and turned out to be a real systematic bug, not BF16 drift. Fixed by #66 — refactored the script to follow the mlir-air xrt 43_triton_layernorm prototype + a hybrid linalg_promote for the W operand.

rms_norm stderr — Visible in your output but the test passed, so it wasn't on your bug list. Was a real latent bug though: linalg_promote emits self-copies that crash transform.air.copy_to_dma's 1-result contract. Fixed by #65 with the same refactor pattern.

matmul_f32_padded_atransposed (109/109 zeros) — Expected on Windows for now; that's a known limitation while full-ELF support on Windows is still WIP.