[Feature Request] Support using embeddings as input

[SanftMonster]

There used to be an issue asking for a way to use embeddings instead of tokens as input to generate the response. Now llama.cpp has supported using embeddings as input, as shown below.

typedef struct llama_batch {
      int32_t n_tokens;

      llama_token  *  token;
      float        *  embd; // set this member and keep token as null will take embeddings as input
      llama_pos    *  pos;
      int32_t      *  n_seq_id;
      llama_seq_id ** seq_id;
      int8_t       *  logits; 
      llama_pos    all_pos_0;  // used if pos == NULL
      llama_pos    all_pos_1;  // used if pos == NULL
      llama_seq_id all_seq_id; // used if seq_id == NULL
  } llama_batch;

Since there's already a binding of this struct in LLamaSharp, what needs to be done is to add an API for executors to accept embeddings as input.

[martindevans]

I've been planning to look into this for a while, since it's required for the BatchedExecutor to support llava. My plan has been to create a new batch class (LlamaBatchEmbeddings). This would probably be a lot simpler than the existing batch (LlamaBatch).

[SanftMonster]

That's great! I didn't notice that it's required for the BatchedExecutor to support llava so I set it as a good first issue. You could split the task into some sub-tasks and set easy and not urgent ones as good first issue, if you would like to. :)

[martindevans]

#770 added a new LLamaBatchEmbeddings which can be used to drive inference with embeddings instead of tokens. It can be used with any context (it's not tied to the BatchedExecutor) by calling Decode.