ggml : move more prints to the ggml log system

[slaren]

Use the ggml log system more consistently. Some of the trace prints that previously were shown in debug builds unconditionally now use GGML_LOG_DEBUG and won't be shown unless verbose mode (-v) is used.

[slaren]

This causes some debug prints in ggml-backend (backend registering messages) to be shown without -v because they are generated during common_params_parse, which is called before the logging callback is set in common_init. Should common_init be called first instead? Or even called automatically from common_params_parse.

[ggerganov]

Yes, let's call it automatically in common_params_parse. I was hesitating to do it and the only reason to call it explicitly after common_params_parse was a very minor/esthetic one:

Currently on master with log prefix and timestamps the output at the start of the program looks like this:

# env
LLAMA_LOG_COLORS=1
LLAMA_LOG_PREFIX=1
LLAMA_LOG_TIMESTAMPS=1

./llama-cli

0.00.000.192 I build: 3910 (36815404) with Apple clang version 16.0.0 (clang-1600.0.26.3) for arm64-apple-darwin24.0.0
0.00.000.978 I main: llama backend init
0.00.001.901 I main: load the model and apply lora adapter, if any
...

While if common_init is called before common_params_parse it will look like this:

build: 3910 (36815404) with Apple clang version 16.0.0 (clang-1600.0.26.3) for arm64-apple-darwin24.0.0
0.00.001.902 I main: llama backend init
0.00.002.943 I main: load the model and apply lora adapter, if any
...

I.e. the first message with the build info does not have the log prefix since the env variables haven't been parsed yet.

Anyway, this is a very minor detail that I don't mind to ignore.

[slaren]

On a second though, calling common_init first won't fix this problem either for the same reason that you mentioned, the command line arguments are not parsed yet when these prints happens, so the verbose flag has no effect. I have reverted that change so that these messages are shown with debug builds only.