Compile bug: Emulated Linux ARM64 CPU build fails

[SamuelTallet]

Git commit

5cd85b5

Operating systems

Linux

GGML backends

CPU

Problem description & steps to reproduce

My Docker ARM64 build (based on Alpine Linux 3.20) fails while compiling the CPU flavor of LLaMA.cpp.
Additional context: It's a GitHub Action using QEMU.

First Bad Commit

Between 0a11f8b and 5cd85b5, I suspect 21ae3b9

Relevant log output

# cmake -B build -DLLAMA_BUILD_SERVER=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
-- The C compiler identification is GNU 13.2.1
-- The CXX compiler identification is GNU 13.2.1
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /usr/bin/cc - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Found Git: /usr/bin/git (found version "2.45.2")
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD - Success
-- Found Threads: TRUE
-- Warning: ccache not found - consider installing it for faster compilation or disable this warning with GGML_CCACHE=OFF
-- CMAKE_SYSTEM_PROCESSOR: aarch64
-- Including CPU backend
-- Found OpenMP_C: -fopenmp (found version "4.5")
-- Found OpenMP_CXX: -fopenmp (found version "4.5")
-- Found OpenMP: TRUE (found version "4.5")
-- ARM detected
-- Performing Test GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E
-- Performing Test GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E - Failed
-- ARM -mcpu not found, -mcpu=native will be used
-- Performing Test GGML_MACHINE_SUPPORTS_dotprod
-- Performing Test GGML_MACHINE_SUPPORTS_dotprod - Failed
-- Performing Test GGML_MACHINE_SUPPORTS_i8mm
-- Performing Test GGML_MACHINE_SUPPORTS_i8mm - Failed
-- Performing Test GGML_MACHINE_SUPPORTS_sve
-- Performing Test GGML_MACHINE_SUPPORTS_sve - Failed
cc1: error: unknown value 'native+nodotprod+noi8mm+nosve' for '-mcpu'
cc1: note: valid arguments are: cortex-a34 cortex-a35 cortex-a53 cortex-a57 cortex-a72 cortex-a73 thunderx thunderxt88p1 thunderxt88 octeontx octeontx81 octeontx83 thunderxt81 thunderxt83 ampere1 ampere1a emag xgene1 falkor qdf24xx exynos-m1 phecda thunderx2t99p1 vulcan thunderx2t99 cortex-a55 cortex-a75 cortex-a76 cortex-a76ae cortex-a77 cortex-a78 cortex-a78ae cortex-a78c cortex-a65 cortex-a65ae cortex-x1 cortex-x1c neoverse-n1 ares neoverse-e1 octeontx2 octeontx2t98 octeontx2t96 octeontx2t93 octeontx2f95 octeontx2f95n octeontx2f95mm a64fx tsv110 thunderx3t110 neoverse-v1 zeus neoverse-512tvb saphira cortex-a57.cortex-a53 cortex-a72.cortex-a53 cortex-a73.cortex-a35 cortex-a73.cortex-a53 cortex-a75.cortex-a55 cortex-a76.cortex-a55 cortex-r82 cortex-a510 cortex-a710 cortex-a715 cortex-x2 cortex-x3 neoverse-n2 cobalt-100 neoverse-v2 demeter generic
CMake Error at ggml/src/ggml-cpu/CMakeLists.txt:145 (message):
Failed to get ARM features
Call Stack (most recent call first):
ggml/src/CMakeLists.txt:298 (ggml_add_cpu_backend_variant_impl)

-- Configuring incomplete, errors occurred!

[SamuelTallet]

Issue is related to ARM emulation with QEMU.
I fixed my build pipeline with these CMake parameters -DGGML_NATIVE=OFF
and -DGGML_CPU_ARM_ARCH=armv8-a (because it matches my requirements).

[slaren]

We should still avoid passing incompatible flags to the compiler with GGML_NATIVE enabled, so let's keep this open.

[Nurb4000]

For what its worth, same thing happened to me this morning on NVIDIA Jetson hardware ( an older Xavier NX ). First time im attempting this. Not real sure what version their CPU would be considered since its their own flavor, so i passed in "-DGGML_CPU_ARM_ARCH=generic" instead of a specific flavor. Its compiling now, will see how it goes.

Update: Compile finished rather quickly so tried a really small model ( just under 2gb ) and while it loads so slow i thought it had died, it did eventually finish and runs at just under 7t/s .. ( i'm sure load time is due to the hardware )

Update 2: seems its not using GPU for inference ( but cuda is detected ).. i guess i have some more research to do. But hey, its at least running now :)