server: tests - slow inference causes timeout on the CI

[phymbert]

Context

Since we fixed // issue on server.cpp, step the server is busy is faster and the server is idle timeout after 3 seconds on the CI:
https://github.com/ggerganov/llama.cpp/actions/runs/8038306408/job/21954067563#step:11:131

Fix

This fix increases the timeout to 10s instead of 3s before considering the inference failed.

[phymbert]

@ggerganov My bad, I merged #5708 without waiting for CI tests 👎

[phymbert]

@ggerganov Can we remove this log in the llama.log, it's annoying:

llama.cpp/common/sampling.cpp

Line 269 in f1a98c5

LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());

[ggerganov]

Can we remove this log in the llama.log, it's annoying:

Yes, just comment the line

[phymbert]

Fixed and tested here: https://github.com/phymbert/llama.cpp/actions/runs/8040928813/job/21959594752
I am merging it now since it blocks master and there is so much job queued on the project.

@ggerganov each llama.cpp push on PRs is now triggering 110 workflow jobs, we need hours before to have CI checks passed. Do you really need to test all the arch on the Server ci workflow ? at least I can remove GPU based BLAS builds ? as we cannot start the server anyway on ubuntu GitHub CPU based runners.

[ggerganov]

@ggerganov each llama.cpp push on PRs is now triggering 110 workflow jobs, we need hours before to have CI checks passed. Do you really need to test all the arch on the Server ci workflow ? at least I can remove GPU based BLAS builds ? as we cannot start the server anyway on ubuntu GitHub CPU based runners.

We need just 4 builds:

• cmake --config Debug -DLLAMA_SANITIZE_ADDRESS
• cmake --config Debug -DLLAMA_SANITIZE_THREAD
• cmake --config Debug -DLLAMA_SANITIZE_UNDEFINED
• cmake --config Release

For now, we cannot test the GPU builds with Github CI. We can potentially add them to ggml-ci in the future