Windows ARM64 (Snapdragon X Elite) Build: Three Blockers and Fixes

[caprion]

System

OS	Windows 11 ARM64 Build 26200
CPU	Qualcomm Snapdragon X Elite
VS	2022 Community 17.14, ClangCL 19.1.5
CMake	4.2.3
Python	3.13 ARM64

Summary

Native Windows ARM64 build is possible on Snapdragon X Elite with three fixes. After applying them, llama-cli.exe builds and runs at 28.59 tok/s (i2_s kernel, 8 threads, 2.41B model). MATMUL_INT8 = 1 confirms i8mm is active.

---

Blocker 1: Wrong -march flags (i8mm falsely believed unsupported)

The existing documentation and community issues mark i8mm support on Snapdragon X Elite as "unknown." It is not unknown — the CPU fully supports it. WSL2 /proc/cpuinfo confirms:

Features: ... i8mm bf16 dotprod asimddp ...

The error when building with -march=armv8.2-a+fp16:

error: always_inline function 'vmmlaq_s32' requires target feature 'i8mm',
but would be inlined into function 'ggml_vec_dot_q4_0_q8_0' that is compiled
without support for 'i8mm'

ClangCL detects the ARM64 target and conditionally enables the i8mm code path, but the -march flag doesn't authorise the instructions. Fix: use -march=armv8.6-a+fp16 — the Snapdragon X Elite is ARMv8.6-A capable.

---

Blocker 2: C++ exceptions disabled by default in ClangCL

error: cannot use 'throw' with exceptions disabled

ClangCL defaults to /EHs-. llama.cpp uses throw throughout. Fix: add /EHsc to CMAKE_CXX_FLAGS.

---

Blocker 3: Missing <chrono> include in common/common.cpp and common/log.cpp

error: no type named 'system_clock' in namespace 'std::chrono'
error: 'clock' is not a class, namespace, or enumeration

On Linux/Mac, <chrono> is pulled in transitively through <thread>. On Windows with ClangCL it is not. Both common.cpp and log.cpp use std::chrono without explicitly including <chrono>. Fix: add #include <chrono> to both files.

---

Working Build Command

From a VS 2022 Developer PowerShell with ClangCL on PATH:

# Kernel generation (required first)
python utils/codegen_tl1.py --model bitnet_b1_58-3B --BM 160,320,320 --BK 64,128,64 --bm 32,64,32

# Configure
cmake -B build `
  -T ClangCL `
  -DBITNET_ARM_TL1=OFF `
  -DCMAKE_BUILD_TYPE=Release `
  -DCMAKE_C_FLAGS="-march=armv8.6-a+fp16" `
  -DCMAKE_CXX_FLAGS="-march=armv8.6-a+fp16 /EHsc"

# Build
cmake --build build --config Release --target llama-cli

---

Confirmed Performance — Snapdragon X Elite, i2_s kernel

Metric	Value
Model	BitNet-b1.58-2B-4T (i2_s)
Load time	844 ms
Prompt eval	210.27 tok/s
Generation	28.59 tok/s
Threads	8
MATMUL_INT8	active

---

Suggested Repo Changes

1. CMakeLists.txt — detect ClangCL on ARM64 Windows and add /EHsc and -march=armv8.6-a+fp16 automatically
2. common/common.cpp and common/log.cpp — add #include <chrono> explicitly (transitive include is not portable)
3. Documentation — note that Snapdragon X Elite supports i8mm, bf16, dotprod; -march=armv8.6-a is the correct target
4. CI — consider adding a Windows ARM64 build check (GitHub Actions now supports ARM64 runners)

---

Additional Note

gguf-py install fails on Windows ARM64 due to a CMake 4.x incompatibility in the sentencepiece submodule and a hardcoded -A x64 arch flag. This blocks setup_env.py but is not needed for inference if GGUF models are downloaded directly from HuggingFace.