Hi, I'm not familiar with llama-cpp-python (actually not familiar with cpp) but I have to use gguf model for my project.

I want to generate answer from pre-computed embedding vectors(torch.Tensor) with size (1, n_tokens, 4096), not from query text. Here I mean the embedding vectors are text embeddings that generated from torch.nn.Embedding()
(Just like inputs_embeds argument of generate() function of transformers model)

What I want to do is just skip process 1 and 2:

1. tokenize input string
2. make text embeddings from tokens
3. model inference
4. get output token
5. detokenize

Is this feature already implemented? If not, please anyone help me where should I begin.