ggml-cpu.c: `src1_cont` undeclared → build fails on ARM CPUs with i8mm (Apple M2/M3/M4, Graviton3, ...)

Summary

Compilation of 3rdparty/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c fails on any ARM CPU that has the i8mm extension when GGML_NATIVE=ON (the default used by setup_env.py). The I2_S fast path uses a variable that only exists when llamafile is enabled, but the file explicitly disables llamafile on i8mm/SVE targets. Independent of the model and of the OS (found on macOS; a Linux ARM host with i8mm, e.g. Graviton3 or Neoverse V1/N2, should hit the same line).

Environment

• MacBook Pro, Apple M2 Pro, macOS 26.5.2, Apple clang 21.0.0 (clang-2100.1.1.101), CMake 4.4.3
• microsoft/BitNet at 0b341e5 (current main), submodule 3rdparty/llama.cpp at 390c3077 (branch release-bitnet-embedding-0.6b-270m)
• Default setup_env.py configuration (-DBITNET_ARM_TL1=OFF), plus -DCMAKE_SHARED_LINKER_FLAGS=-Wl,-undefined,dynamic_lookup to get past the macOS link error of #611 / #595 (otherwise the build stops earlier and this error is never reached)

Error

3rdparty/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c:1495:40: error: use of undeclared identifier 'src1_cont'
 1495 |         const size_t src1_col_stride = src1_cont || src1->type != vec_dot_type ? i2s_row_size : nb11;
      |                                        ^~~~~~~~~
1 error generated.

Compile flags for that translation unit (from compile_commands.json): -mcpu=native+dotprod+i8mm+nosve+nosme ... -DGGML_USE_LLAMAFILE.

Root cause

ggml-cpu.c line 48:

#if defined(__ARM_FEATURE_SVE) || defined(__ARM_FEATURE_MATMUL_INT8)
#undef GGML_USE_LLAMAFILE
#endif

With i8mm enabled (-mcpu=native+...+i8mm, so __ARM_FEATURE_MATMUL_INT8 is defined), the llamafile block in ggml_compute_forward_mul_mat is compiled out. src1_cont is declared only inside that block (line 1361, #if GGML_USE_LLAMAFILE). The I2_S fast path added at lines 1491-1495 uses src1_cont unconditionally, so it only compiles when llamafile happens to be on: x86, and ARM without i8mm (e.g. Apple M1). Every ARM CPU with i8mm fails.

Suggested fix

Compute the flag locally where it is needed, e.g. in the I2_S block:

const bool src1_cont = ggml_is_contiguous(src1);

or hoist the existing declaration out of the #if GGML_USE_LLAMAFILE region.

Workarounds

• -DCMAKE_C_FLAGS=-U__ARM_FEATURE_MATMUL_INT8 -DCMAKE_CXX_FLAGS=-U__ARM_FEATURE_MATMUL_INT8 (same mechanism ggml itself uses when a feature check fails), or
• -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8.2-a+dotprod+fp16 (build without i8mm).

Both make the build complete. Note that I2_S inference on the resulting arm64 binary is still wrong (garbage output even with the official BitNet-b1.58-2B-4T-gguf), which is the separate, already reported #585 / #600.

Start in 3rdparty/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c, reading the feature guards near line 48 and the I2_S path in ggml_compute_forward_mul_mat around lines 1361 and 1491-1495. Build with the default setup_env.py configuration and GGML_NATIVE=ON on an i8mm ARM target; done means compilation completes without the src1_cont error.