-
Notifications
You must be signed in to change notification settings - Fork 12.5k
Description
I wanted to build binaries for ARMv6,7 on Linux but I'm getting the following two errors:
[19:53:33] /workspace/srcdir/llama.cpp/ggml-quants.c: In function ‘ggml_vec_dot_iq2_xxs_q8_K’:
| [19:53:33] /workspace/srcdir/llama.cpp/ggml-quants.c:7264:19: error: implicit declaration of function ‘vld1q_s8_x4’; did you mean ‘vld1q_s64’? [-Werror=implicit-function-declaration]
| [19:53:33] 7264 | q8b = vld1q_s8_x4(q8); q8 += 64;
| [19:53:33] | ^~~~~~~~~~~
| [19:53:33] | vld1q_s64
| [19:53:33] /workspace/srcdir/llama.cpp/ggml-quants.c:7264:19: error: incompatible types when assigning to type ‘int8x16x4_t’ from type ‘int’
| [19:53:33] cc1: some warnings being treated as errors
They both seem to stem from the changes made in SOTA 2-bit PR.
I know that there was a reference that ARM NEON isn't supported yet, but I thought the above would still build.
I haven't found any way to build without the quants (eg, equivalent of the previous LLAMA_K_QUANTS=OFF
).
Steps to reproduce:
cmake .. \
-DCMAKE_INSTALL_PREFIX=$prefix \
-DCMAKE_TOOLCHAIN_FILE=${CMAKE_TARGET_TOOLCHAIN} \
-DCMAKE_BUILD_TYPE=RELEASE \
-DBUILD_SHARED_LIBS=ON \
-DLLAMA_BUILD_TESTS=OFF \
-DLLAMA_BUILD_EXAMPLES=ON \
-DLLAMA_NATIVE=OFF \
-DLLAMA_ACCELERATE=ON \
-DLLAMA_AVX=ON \
-DLLAMA_AVX2=ON \
-DLLAMA_F16C=ON \
-DLLAMA_FMA=ON \
-DLLAMA_BLAS=OFF \
-DLLAMA_CUBLAS=OFF \
-DLLAMA_CLBLAST=OFF
Version: Latest master (b1794)
System: build system for armv7l-linux-gnueabihf-cxx03 (C compiler 10.2.0)
Build log: https://buildkite.com/julialang/yggdrasil/builds/7645#018ceaa2-2f2b-410c-8e4c-826c8b6a48a9
(all logs are here: https://buildkite.com/julialang/yggdrasil/builds/7645#_)
Thank you for your advice!
PS: The error is similar to this one, but I don't believe it's related (different arch)