How to build it on Apple M1?

[evstarshov]

Works fine with my intel mac, but on M1 i got these errors when I launch make file:

clang: warning: argument unused during compilation: '-s' [-Wunused-command-line-argument]
otherarch/ggml_v1.c:1152:21: error: always_inline function '_mm256_cvtph_ps' requires target feature 'f16c', but would be inlined into function 'ggml_v1_vec_dot_f16' that is compiled without support for 'f16c'
ax[j] = GGML_V1_F16_VEC_LOAD(x + i + j*GGML_V1_F16_EPR, j);
^
otherarch/ggml_v1.c:800:40: note: expanded from macro 'GGML_V1_F16_VEC_LOAD'
#define GGML_V1_F16_VEC_LOAD(p, i) GGML_V1_F32Cx8_LOAD(p)
^
otherarch/ggml_v1.c:790:36: note: expanded from macro 'GGML_V1_F32Cx8_LOAD'
#define GGML_V1_F32Cx8_LOAD(x) _mm256_cvtph_ps(_mm_loadu_si128((__m128i )(x)))
^
otherarch/ggml_v1.c:1153:21: error: always_inline function '_mm256_cvtph_ps' requires target feature 'f16c', but would be inlined into function 'ggml_v1_vec_dot_f16' that is compiled without support for 'f16c'
ay[j] = GGML_V1_F16_VEC_LOAD(y + i + jGGML_V1_F16_EPR, j);
^
otherarch/ggml_v1.c:800:40: note: expanded from macro 'GGML_V1_F16_VEC_LOAD'
#define GGML_V1_F16_VEC_LOAD(p, i) GGML_V1_F32Cx8_LOAD(p)
^
otherarch/ggml_v1.c:790:36: note: expanded from macro 'GGML_V1_F32Cx8_LOAD'
#define GGML_V1_F32Cx8_LOAD(x) _mm256_cvtph_ps(_mm_loadu_si128((__m128i )(x)))
^
otherarch/ggml_v1.c:1530:21: error: always_inline function '_mm256_cvtph_ps' requires target feature 'f16c', but would be inlined into function 'ggml_v1_vec_mad_f16' that is compiled without support for 'f16c'
ax[j] = GGML_V1_F16_VEC_LOAD(x + i + jGGML_V1_F16_EPR, j);
^
otherarch/ggml_v1.c:800:40: note: expanded from macro 'GGML_V1_F16_VEC_LOAD'
#define GGML_V1_F16_VEC_LOAD(p, i) GGML_V1_F32Cx8_LOAD(p)
^
otherarch/ggml_v1.c:790:36: note: expanded from macro 'GGML_V1_F32Cx8_LOAD'
#define GGML_V1_F32Cx8_LOAD(x) _mm256_cvtph_ps(_mm_loadu_si128((__m128i )(x)))
^
otherarch/ggml_v1.c:1531:21: error: always_inline function '_mm256_cvtph_ps' requires target feature 'f16c', but would be inlined into function 'ggml_v1_vec_mad_f16' that is compiled without support for 'f16c'
ay[j] = GGML_V1_F16_VEC_LOAD(y + i + jGGML_V1_F16_EPR, j);
^

[LostRuins]

Please try the latest commit and see if the issue is fixed.

[evstarshov]

Can you give me advice how to change compiler? I tried to add to Makefile line 174: -arch arm64. It compiles, but after start window and selecting model I have another error:

Welcome to KoboldCpp - Version 1.15
For command line arguments, please refer to --help
Otherwise, please manually select ggml file:
2023-04-28 12:56:09.007 python3[22414:754319] +[CATransaction synchronize] called within transaction
Warning: OpenBLAS library file not found. Non-BLAS library will be used.
Initializing dynamic library: koboldcpp.so
Traceback (most recent call last):
File "/Volumes/KINGSTON/M1/koboldcpp/koboldcpp.py", line 588, in
main(args)
File "/Volumes/KINGSTON/M1/koboldcpp/koboldcpp.py", line 499, in main
init_library() # Note: if blas does not exist and is enabled, program will crash.
File "/Volumes/KINGSTON/M1/koboldcpp/koboldcpp.py", line 122, in init_library
handle.load_model.argtypes = [load_model_inputs]
File "/Users/evgenijstarsov/miniconda3/lib/python3.10/ctypes/init.py", line 387, in getattr
func = self.getitem(name)
File "/Users/evgenijstarsov/miniconda3/lib/python3.10/ctypes/init.py", line 392, in getitem
func = self._FuncPtr((name_or_ordinal, self))
AttributeError: dlsym(0x85df4bf0, load_model): symbol not found

[Myrminki]

Holy cow, now I can compile it on my old xeon e5 1620 0 on Linux, the other versions gave me similar errors, so I tried again and now it works.

[LostRuins]

nice

[evstarshov]

I found my problem why it won't work on M1 mac after compiling - my terminal was opened with enabled Rosetta option. I enabled it to run another project and forgot about it. Now everything compiles and works perfectly.