server: add "schema" and validation

[ngxson]

Overview

Add the notion of server_schema that will define the input schema in a more systematic way.

This is actually something I wanted to do since a long time ago. Motivations for this proposal:

1. To address bugs like GHSA-8947-pfff-2f3c due to lack of input validation
2. Generalize the validation logic
3. Inline documentation, can be exported to markdown in the future

TODO in follow-up PRs:

• Migrate other schema (for example, chat completion) to using this system
• Export schema to markdown
• Allow linking arg.cpp <--> server-schema.cpp via an enum? (not sure how it will be useful)

Example of the code:

// old way
params.sampling.top_k              = json_value(data, "top_k",               defaults.sampling.top_k);

// new way
add((new field_num("top_k", params.sampling.top_k))
    ->set_limits(0, INT32_MAX)
    ->set_desc("Limit the next token selection to the K most probable tokens (0 = disabled)"));

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: to spot and fix mistakes during the migration

[ServeurpersoCom]

Testing this on my pod. It seems to me that the OAI spec can have sampling values less than 0 even if it is rarely useful in practice, such as encouraging repetitions with frequency_penalty. ( https://developers.openai.com/api/reference/python/resources/completions/methods/create search for "-2")

[ServeurpersoCom]

Type errors return the raw nlohmann message without the field name, maybe wrap the eval loop in a try/catch to prepend it so the client knows which param failed?

[ngxson]

yes I added it in the last commit(s), along with some corrections for the numerical limits. PTAL

example error message:

"message": "Field 'min_keep': Value must be between 0 <= value <= 2147483647, but got -100"

"message": "Field 'min_keep': [json.exception.type_error.302] type must be number, but is string",

[ServeurpersoCom]

std::set<const char *> compares pointer addresses, not strings, so the alias order is not the insertion order. I got max_tokens winning over n_predict on my pod just by changing TU link order. A std::vector fixes it and it's a 2 line change, tested here.

Suggested change

	std::set<const char *> name;
	std::vector<const char *> name;

And :

-        name.insert(n);
+        name.push_back(n);

[ngxson]

yup, nice catch