fix(llama-cpp-server): fix rocm build by setting GGML_HIPBLAS

[richard-jfc]

Various LLAMA_ build flags have been renamed to GGML_ in llama-cpp-server (See ggerganov/llama.cpp@f3f6542) and need to be renamed in tabby to allow correctly targeting Metal/Rocm/Vulkan.

It appears that this was already done for Cuda.

May be related: #2811

[wsxiaoys]

Have you verified if the PR fixes your issue?

cmake handles the migration with macros defined in https://demo.tabbyml.com/files/git/llama.cpp/-/blob/7a221b672e49dfae459b1af27210ba3f2b5419b6/CMakeLists.txt?plain=1#L101

So my guess is it won't fix any build issues you are currently encountering.

[richard-jfc]

I built rocm v0.15.0 tag using docker. And have been running the built Rocm version via docker.

Without my change I was seeing a CPU spike for every chat request and no VRAM or GPU usage.

After my change I was seeing roughly 6GB VRAM and a GPU usage spike.

[wsxiaoys]

It does seem that GGML_HIPBLAS is the flag not handled by the llama_option_depr migration. In that case, would you consider updating the PR to contain only the GGML_HIPBLAS change, since it's the only one you've verified?

[richard-jfc]

Thanks, Done