---------- Forwarded message --------- From: "Nathalie Chan King Choy (Windows on Arm)" <woa@hub.corecollective.dev> Date: Sunday, 5 July 2026 at 19:23:27 UTC-7 Hi all, The notes from the 6/30 call are below. Please let us know if you spot any errors or important omissions. Thanks & regards, Nathalie -- CoreCollective Community Manager Android & WoA working groups *Attended:* Saket J John S Randy L Anthony R Mayank A Dave W Nathalie CKC Gleb K Evgeny K Omair J Jillian B *AI Notes:* Meeting Overview - The meeting focused on Python upstreaming progress, along with LLVM and GCC toolchain work for Windows on Arm. - The group discussed ongoing challenges with unresponsive upstream project maintainers delaying pull request merges. Key Decisions Made - Python Upstreaming Prioritization: Idled down active LLVM optimization work & Works on WoA website work to redirect engineering efforts toward Python upstreaming and dependency enablement. - Numba Model Workaround: Using a small code model configuration within LLVM Lite to manage Numba just-in-time compilation , bypassing an immediate requirement for native COFF large code model support from Microsoft. - ARMv7 Linux Release Discontinuation: Planning to discontinue ARMv7 Linux releases in subsequent LLVM release cycles, unless someone decides to sponsor continuation. - Repository Access Guidelines: [Deferred] Documenting specific criteria for collaborator access levels on the Core Collective repository was deferred since very few users currently require write access. Action ItemsGleb Khmyznikov - Contact the internal Microsoft team that previously worked on Torch for Windows on Arm to determine why the packages are missing from PyPI. - Inquire internally at Microsoft about potential engineering contacts or options to advance the OpenAI TikToken pull request. - Share the technical details of the LLVM code model issue internally to discuss native support requirements. - Receive the list of stalled packages from Mayank Agrawal to forward to Anaconda contacts once their infrastructure is ready. Omair Javaid - Linaro - Send detailed technical notes regarding the LLVM code model and COFF file format limitations to Gleb Khmyznikov. - Monitor ongoing testing and provide support if needed for the small code model implementation. - Collaborate with the group in the coming days to troubleshoot compiler and LLVM-specific issues for Python libraries. Qualcomm Team / Mayank Agrawal - Conduct testing from the Qualcomm side to ensure the small code model works cleanly across the test suite. - Begin working on the Windows on Arm enablement for Librosa and Shap now that Numba is available. Dave Whaley - Provide links from the Microsoft Build conference announcements to add to the Working Group repository landing page. For Future - LLVM 23 Release: Release Candidates expected in two weeks and the final release out in about a month. - Seek help from Anaconda: We should create a list of packages that are stuck for months without response from maintainers. Anaconda might be interested in supporting this effort. Gleb is this WG’s interface with Anaconda. NotesPython Upstreaming & Unresponsive Maintainers - Anthony noted that many project maintainers are completely ignoring all pull requests across their repositories, rather than ignoring Windows on Arm specific submissions. - The pool of open projects has shrunk, making the remaining unresponsive maintainers a more noticeable bottleneck for enablement work. - Anthony clarified that the Jellyfish pull request was mistakenly submitted to a dead GitHub mirror instead of the author's primary repository. SGLang & PyTorch Blockers - Anthony is currently trying to get SGLang to build on Windows on Arm. - Torch builds Windows on Arm nightlies successfully, but these packages are not uploaded to PyPI. - Manually downloading and installing the nightly Torch wheel allows the SGLang installation to proceed. LLVM Code Models & Numba Resolution - Numba acts as a just-in-time compiler that utilizes LLVM Lite to translate Python into LLVM IR. - The toolchain previously encountered errors because LLVM Lite used a large code model, which lacks support in the COFF object file format for AArch64. - Moving to a small code model resolved the build issue and has been officially approved by the PR reviewer. - Evgeny Karpov noted that using the small code model functions as a workaround, and the issue might reoccur if code differences scale beyond a 4 GB address space. GCC & Toolchain Pipelines - The LLVM Windows on Arm release pipeline has been fully automated via GitHub Actions and is funded by the LLVM Foundation. - Evgeny Karpov reported that a recent regression in the MinGW library was caught and successfully validated with a patch that will be merged upstream soon. Working Group Repository Governance - https://github.com/CoreCollective-dev/windows-on-arm-wg - Gleb Khmyznikov has been granted write access - No one else in the WG has requested write access yet - let Nathalie know if needed - Once we have more folks requesting write access we should discuss governance for external contributor roles