Fix/win ninja tl2 q6k

[manuelmen2a-blip]

Remove dead/non-building code + add portable Windows/Ninja CI

Removed (dead code)

• gpu/bitnet_kernels/gemm_native.cu (69 lines): no references anywhere in
  the tree (not in any CMakeLists.txt, compile.sh/setup.py; grep for
  gemm_native/native_gemm/ggml_bitnet_gemm under gpu/ → 0 matches).
  Never compiled, never called; its kernel also produced numerically wrong
  results in local testing.
• src/ggml-bitnet-lut.cpp — TL2 x86 ggml_bitnet_mul_mat (46 lines):
  calls ggml_bitnet_mul_mat_task_init/ggml_bitnet_mul_mat_task_compute,
  which are declared in include/ggml-bitnet.h but defined nowhere in the repo,
  so any build with BITNET_X86_TL2 fails to link. The working path is I2_S,
  not TL2.

Changed (build portability)

• CMakeLists.txt: ggml-base arch is now a cache variable
  BITNET_GGML_ARCH (default native) instead of a hardcoded
  -march=native. -march=native is not portable across CI runners (one
  runner hit undeclared AVX-512 intrinsics in ggml.c); CI builds with
  -DBITNET_GGML_ARCH=x86-64-v3 (the AVX2 microarchitecture level).
• .github/workflows/ci.yml: new GitHub Actions workflow
  (windows-latest, clang + Ninja) that configures, builds, and smoke-tests
  llama-cli --version (which loads ggml-base.dll, so a link failure
  surfaces). Configure flags: -DBITNET_GGML_ARCH=x86-64-v3 -DLLAMA_BUILD_COMMON=ON -DLLAMA_BUILD_EXAMPLES=ON -DLLAMA_BUILD_TOOLS=ON.

Kept (unchanged)

• The Win/Ninja I2_S link fix: src/ggml-bitnet-mad.cpp and
  src/ggml-bitnet-shim.c compiled directly into ggml-base via
  target_sources (fixes the quantize_i2_s / dequantize_row_i2_s
  undefined-symbol link error caused by GGML_SOURCES_BITNET never being
  consumed).
• bitlinear_int8xint2_batched in gpu/bitnet_kernels/bitnet_kernels.cu
  (referenced by gpu/model.py:107 for the batched prefill path).

Verification

• CI green on the fork (manuelmen2a-blip/BitNet, run
  36258669883): Configure → Build → Smoke test all pass (7m47s).
• llama-cli.exe --version exits 0 and loads ggml-base.dll.
• The set of compiled translation units is unchanged versus the previous
  working I2_S build (only unreferenced code was removed).

Test plan

• cmake --build build configures, compiles and links on Windows/Ninja
• ggml-bitnet-mad.cpp compiled with an AVX2-level ISA (x86-64-v3)
• BITNET_X86_TL2 path no longer carries a file that cannot link
• llama-cli --version smoke test passes
• Cross-repo PR CI needs maintainer approval to run on
  microsoft/BitNet (GitHub runs base-branch workflows only for fork PRs)

🤖 Generated with Claude Code

[manuelmen2a-blip]

fix: win/ninja I2_S build + TL2 x86 mul_mat + GPU int2 prefill + GEMM M>1

[manuelmen2a-blip]

@microsoft-github-policy-service agree