llama-server: Support --verbose-prompt

[a4lg]

llama-server has the option --verbose-prompt but was ineffective.

This commit recovers the logger for full prompt input (gated behind the non-default option --verbose-prompt) which was originally commented out.

[ngxson]

Use slots debug instead

[a4lg]

Okay, undocumented slot debugging feature...?

How to use it?

Invoking with Slots Debug enabled

LLAMA_SERVER_SLOTS_DEBUG=1 llama-server --port 11434 [ARGS...]

How to inspect the results?

After the process finishes, inspect slots. Here is an example:

# curl --silent 'http://localhost:11434/slots' | jq
[
  ...,
  {
    "id": 7,
    "n_ctx": 2048,
    "speculative": false,
    "is_processing": false,
    "id_task": 0,
    ...,
    "prompt": "<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n<|im_start|>user\nHello!<|im_end|>\n<|im_start|>assistant\n",
    "generated": "Hello! How can I assist you today?"
  }
]

You may need to specify the model name on the router mode.

curl --silent 'http://localhost:11434/slots?model=Qwen%2FQwen2.5-Coder-1.5B-Instruct' | jq

Alternatives

@ngxson
I'm half convinced. Although undocumented, it's a good feature which I have missed.

Then, I came up with one question. What's the point leaving --verbose-prompt which is completely ineffective on llama-server in the first place? If this is a valid question, I'll submit a PR to remove --verbose-prompt from llama-server instead.

What do you think?

[ngxson]

@a4lg It was undocumented because it was mean to be used for testing only. But feel free to add it to a new section at the end of server docs.

If this is a valid question, I'll submit a PR to remove --verbose-prompt from llama-server instead.

Yes, please, that would be helpful. You can push it with the same PR I suggested above