[Windows ARM64] Native CUDA serving enablement stack

[khmyznikov]

Purpose

Contribute the validated native Windows ARM64 CUDA 13.4 enablement as a reviewable stack. GitHub native stacks require every branch to live in one repository; because khmyznikov does not have push access to SystemPanic/vllm-windows, the six-layer native stack lives in the contributor fork and this draft is the cross-fork umbrella.

Review stack

1. [Build][Windows] Add native ARM64 build plumbing khmyznikov/vllm-windows#1 - build and packaging foundation
2. [Rust][Windows] Support native process and listener APIs khmyznikov/vllm-windows#2 - Rust Windows process/listener support
3. [Kernel][Windows] Launch SM120 CUTLASS kernels indirectly khmyznikov/vllm-windows#3 - SM120 CUTLASS indirect launch
4. [Build][Windows] Compile native SM120 FlashAttention SASS khmyznikov/vllm-windows#4 - native SM120 FA2 SASS
5. [Attention] Gate FA4 on runtime and SM12x support khmyznikov/vllm-windows#5 - FA4 runtime/SM12x gating
6. [Model Loader][Windows] Synchronize UMA safetensors copies khmyznikov/vllm-windows#6 - Windows UMA safetensors lifetime

Each stacked PR shows only its layer. This umbrella shows the complete end-to-end diff against vllm-for-windows.

Scope

Generated Windows ARM64 documentation is intentionally omitted. The local NVML capability memoization and missing-FlashInfer sampler fallback are also omitted because active upstream vLLM PRs vllm-project#50393 and vllm-project#48956 cover those areas.

Validation

• Rust workspace compiles on native Windows ARM64; rustfmt passes.
• Focused FA4 tests: 6 passed; Ruff passes.
• Focused UMA loader test: 1 passed; Ruff passes.
• CUTLASS source passes clang-format and the native CUDA build compiles the changed SM120 translation unit.
• Full wheel build and final PE/SASS inspection are running from C:\Dev\vllm-windows-stack; results will be added before marking ready.

All commits carry DCO sign-off. AI assistance was used; every changed line and validation result was reviewed by the contributor.

[khmyznikov]

Top-of-stack native Windows ARM64 validation is complete.

• Built vllm-0.26.1.dev6+g00334106a.cu134-cp313-cp313-win_arm64.whl from commit 00334106a with CUDA 13.4, CPython 3.13 ARM64, and native ARM64 MSVC.
• CMake/Ninja completed all 189 targets, including _vllm_fa2_C.pyd.
• The wheel contains seven .pyd files plus vllm/vllm-rs.exe; dumpbin /headers reports AA64 machine (ARM64) for all eight binaries.
• cuobjdump --list-elf reports 76 native sm_120.cubin entries in FA2 (152 cubins total).
• Smoke-imported the exact wheel artifact: vllm, _C_stable_libtorch, _moe_C_stable_libtorch, and _vllm_fa2_C all load successfully.
• cargo check --workspace --locked and cargo fmt --all -- --check pass on native Windows ARM64.
• FA4 runtime gate: 6 focused tests pass; Ruff passes.
• UMA safetensors synchronization: focused test passes; Ruff passes.
• Generated documentation and the two changes overlapping upstream PRs [Bugfix][Platform] Stop re-initializing NVML on every device-capability check (fixes #50381) vllm-project/vllm#50393 and [Bugfix] Fall back to native sampling when FlashInfer cannot target the GPU vllm-project/vllm#48956 remain excluded.

The six same-repository draft PRs remain linked in native stack vllm-project#7. This umbrella remains draft pending maintainer direction on the cross-fork/native-stack permission constraint.

[khmyznikov]

Superseded by #82, which uses a clean descriptive head branch and includes the completed native ARM64 validation results.