llamafile : tmp disable + build sgemm.o when needed

[ggerganov]

MoE ppl is currently abnormally high indicating some issue:

# build without LLAMA_NO_LLAMAFILE or right before this commit
make -j perplexity && ./perplexity -f build/wikitext-2-raw/wiki.test.raw -m models/mixtral-8x7b-32k-fast/ggml-model-f16.gguf -ngl 0

...

perplexity: tokenizing the input ..
perplexity: tokenization took 573.843 ms
perplexity: calculating perplexity over 642 chunks, n_ctx=512, batch_size=2048, n_seq=4
perplexity: 71.87 seconds per pass - ETA 3 hours 12.23 minutes
[1]923.5180,[2]1248.7441,[3]1137.4388,[4]1598.2650,^C

Need to fix this before re-enabling by default

[cebtenzzre]

This error is because vaddvq_f32 is only available on armv8/aarch64, but the build targets armv7 (armv7-none-linux-androideabi33). I doubt this is specific to Android, I'm sure you would hit the same error when building for a 32-bit Raspberry Pi.

Aren't most Android devices armv8 these days, anyway?

[ggerganov]

Yes, I think so. On armv7 people can always disable the code manually via LLAMA_NO_LLAMAFILE. For now I'm just interested in fixing the CI and was hoping that CMAKE_SYSTEM_NAME would do the job, but unfortunately it does not

[cebtenzzre]

The CI is still failing because the Android build pulls from llama.cpp master: https://github.com/ggerganov/llama.cpp/blob/8dd1ec8b3ffbfa2d26e82e672cea89f5eeb2f141/examples/llama.android/app/src/main/cpp/CMakeLists.txt#L16-L20

When CMAKE_SYSTEM_NAME is Android, you can check CMAKE_ANDROID_ARCH_ABI to determine the target architecture.