Misc. bug: Re-check from version b5028 and above, b5026 seems to have valid DLLs compared the new versions???

[spyrman]

Name and Version

b5026 seems to be the last valid
b5028 and above need some review
test project created for that please feel free to test it yourself.

Operating systems

No response

Which llama.cpp modules do you know to be affected?

No response

Command line

Problem description & steps to reproduce

Hello, i am using a wrapper to llama.cpp in Pascal Delphi and i noticed that from b5028 some DLLs are not working as expected i have created a demo project for you to test with Lazarus the project is here
https://github.com/spyrman/Test_llamacpp_dlls
you can compile it yourself or download the executable and try to load the model with various DLLs and see which one will work for you. On my tests this works up to version b5026 while from that version and above it stops working for example it might work with CPU alone but when i try to make it work with the CUDA or the GPUs DLLs it does not seem to work. In other words if i use the n_gpu_layers > 0 please have a look at it.

First Bad Commit

No response

Relevant log output

[JohannesGaessler]

If at all possible, please reproduce the issue using llama.cpp code only.

[slaren]

Some of the releases are now built with dynamically loadable backends, which are not loaded automatically. The easiest way to fix this would be to call ggml_backend_load_all from ggml.dll to load the backends before calling llama.cpp functions. Alternatively, you can make your own build with GGML_BACKEND_DL disabled.