Compilation with CLBlast in onboard graphics

[SilvaRaulEnriqueCJM]

In order to compile with CLBlast:
make LLAMA_CLBLAST=1

say error:

g++ --shared -I. -I./examples -O3 -std=c++11 -fPIC -DNDEBUG -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar -pthread -march=native -mtune=native -DGGML_USE_K_QUANTS -DGGML_USE_CLBLAST -I/usr/local/include -I/home/raul/Desktop/Librerias/OpenCL-SDK/install/include examples/embd-input/embd-input-lib.cpp ggml.o llama.o common.o k_quants.o ggml-opencl.o ggml-alloc.o -o libembdinput.so -L/usr/local/lib -lclblast -L/home/raul/Desktop/Librerias/OpenCL-SDK/install/lib -lOpenCL
/usr/bin/ld: /usr/local/lib/libclblast.a(xaxpy.cpp.o): warning: relocation against _ZN7clblast8database11XaxpyDoubleE' in read-only section .text.startup'
/usr/bin/ld: /usr/local/lib/libclblast.a(xdot.cpp.o): relocation R_X86_64_PC32 against symbol `ZTSZN7clblast6BufferISt7complexIdEEC4ERKNS_7ContextENS_12BufferAccessEmEUlPP7_cl_memE' can not be used when making a shared object; recompile con -fPIC
/usr/bin/ld: falló el enlace final: bad value
collect2: error: ld returned 1 exit status
make: *** [Makefile:382: libembdinput.so] Error 1

¿What is the problem?

[MichaelDays]

Silly suggestion: try running "sudo ldconfig"

The error suggests that the "libclblast.a" library file that is being picked up by g++ was not originally built to be used to link into relocatable object file. The "recompile con -fPIC" option seems to be used to suggest this.

There may already be a version of the ".a" file with a ".so" suffix on disc, but not being picked up by the ld tool's cache. ldconfig will rebuild that cache, and may allow g++ to link against that file instead.

[SlyEcho]

You need to compile CLBlast with -fPIC

Add it to CMAKE_C_FLAGS and CMAKE_CXX_FLAGS (in the file CMakeCache.txt file in the build directory of CLBlast). Recompile it and reinstall it.

[SilvaRaulEnriqueCJM]

Now CLBlast compile in AMD with CuDA and run fast. Thakyou.

But if compile in Intel (without graphics card), no errors in compilation time, but in runtime of main say:

main: build = 1015 (226255b)
main: seed = 1692672569
ggml_opencl: clGetPlatformIDs(NPLAT, platform_ids, &n_platforms) error -1001 at ggml-opencl.cpp:965

[SilvaRaulEnriqueCJM]

And in another Intel Core I3 (without graphics card), in runtime of main say:

main: build = 1018 (8e4364f)
main: seed = 1692703705
ggml_opencl: selecting platform: 'Portable Computing Language'
ggml_opencl: selecting device: 'pthread-haswell-Intel(R) Core(TM) i3-4100M CPU @ 2.50GHz'
ggml_opencl: warning, not a GPU: 'pthread-haswell-Intel(R) Core(TM) i3-4100M CPU @ 2.50GHz'.
ggml_opencl: device FP16 support: false
gguf_init_from_file: invalid magic number 67676a74
error loading model: llama_model_loader: failed to load model from /media/raul/f3a54992-da3d-4e46-93f6-e84c1180f25d/models/TheBloke/llama-2-7b-chat.ggmlv3.q3_K_L.bin

llama_load_model_from_file: failed to load model
llama_init_from_gpt_params: error: failed to load model '/media/raul/f3a54992-da3d-4e46-93f6-e84c1180f25d/models/TheBloke/llama-2-7b-chat.ggmlv3.q3_K_L.bin'
main: error: unable to load model

[SilvaRaulEnriqueCJM]

The parameters are:
-m /media/raul/f3a54992-da3d-4e46-93f6-e84c1180f25d/models/TheBloke/llama-2-7b-chat.ggmlv3.q3_K_L.bin -ngl 43 -i --prompt "Sabiendo que hay un celular azul sobre la mesa, ¿cuál letra corresponde a la respuesta correcta?: a) hay un celular azul sobre la mesa, b) hay un celular marrón sobre la mesa, c) hay un celular negro sobre la mesa, d) hay un celular gris sobre la mesa, e) hay un celular verde sobre la mesea. Respuesta: La letra que corresponde a la respuesta correcta es "

[SlyEcho]

Model format has changed, you need to convert files or download new files: #2398

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.