sampling : refactor init to use llama_sampling_params

[ggerganov]

Change llama_sampling_init()

Step towards integrating the common/sampling into llama.cpp

// old
struct llama_sampling_context * llama_sampling_init(const struct gpt_params & params);

// new
struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params);

Change repetition penalty API

This is more efficient since the repetition, frequency and presence penalties are applied in one go instead of 2

// old
    LLAMA_API void llama_sample_repetition_penalty(
            struct llama_context * ctx,
          llama_token_data_array * candidates,
               const llama_token * last_tokens,
                          size_t   last_tokens_size,
                          float    penalty);

    LLAMA_API void llama_sample_frequency_and_presence_penalties(
             struct llama_context * ctx,
           llama_token_data_array * candidates,
                const llama_token * last_tokens,
                           size_t   last_tokens_size,
                            float   alpha_frequency,
                            float   alpha_presence);

// new
    LLAMA_API void llama_sample_repetition_penalties(
             struct llama_context * ctx,
           llama_token_data_array * candidates,
                const llama_token * last_tokens,
                           size_t   penalty_last_n,
                            float   penalty_repeat,
                            float   penalty_freq,
                            float   penalty_present);

Sampling context keeps fewer previous tokens

The prev array used to have a size of n_ctx. Since it is used just for applying repetition penalties and checking for antiprompts, it does not have to be larger than params.penalty.last_n (a.k.a. repeat_n_last)

Delete obsolete examples

• embd-input
• gptneox-wip

Other changes

• Moved the grammar string from gpt2_params to llama_sampling_params
• Renamed repetition, frequency and presence penalty params in the code. API names are unchanged
• Hided the prev vector in llama_sampling_context behind API calls for future-proofness
• Applied server patch from Fixed segfault when prompt was too long #3661 (bbc0b7c)
• OCD indentations and renaming as usual

---

🤖 Generated by Copilot at 56ba00b

Summary

🗑️🛠️✨

This pull request refactors the sampling module of the llama library to use a new struct for the sampling parameters, to simplify the interface and the penalty logic, and to make the grammar rules optional. It also removes the outdated and incompatible embd-input example and its related files.

We've cleaned up the code and made it run faster
We've dropped embd-input and sorted the examples
We've refactored the sampling and the gpt_params
So heave away, me hearties, heave away on three

Walkthrough

• Removed the embd-input example and its related files, as it was no longer needed (link, link, link, link, link, link, link, link, link, link, link, link, link, link)
• Renamed and moved some fields in the llama_sampling_params and gpt_params structs, and updated the functions that used them, to improve clarity and consistency (link, link, link, link, link, link, link, link, link, link, link, link, link, link, link, link, link, link, link, link, link, link, link, link, link)
• Added new functions and parameters to the common/sampling.cpp and common/sampling.h files, to provide more functionality and flexibility for the sampling process (link, link, link)
• Changed some comments and subdirectory order in the common/sampling.cpp and examples/CMakeLists.txt files, as minor stylistic changes (link, link, link, link)