vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Update llama.cpp to use instruct mode

Browse source
This commit is contained in:
Thomas Antony 2023-03-19 13:00:02 -07:00
parent 660a4d5365
commit edb52ab13c
2 changed files with 14 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

12
llama.cpp
View file

@ -1176,6 +1176,18 @@ void llama_update_input(llama_context& ctx, const std::string& text)
state.embd_inp.insert(state.embd_inp.end(), line_inp.begin(), line_inp.end()); state.embd_inp.insert(state.embd_inp.end(), line_inp.begin(), line_inp.end());
state.remaining_tokens -= line_inp.size(); state.remaining_tokens -= line_inp.size();
} }
/// @brief Updates the context and appends new input tokens (overloaded version)
/// @param ctx
/// @param tokens
void llama_update_input(llama_context& ctx, const std::vector<gpt_vocab::id>& tokens)
{
llama_state& state = *ctx.state;
llama_model& model = ctx.model;
const gpt_params& params = ctx.params;
state.embd_inp.insert(state.embd_inp.end(), tokens.begin(), tokens.end());
state.remaining_tokens -= tokens.size();
}
/// @brief Ingests a batch of input tokens into the context /// @brief Ingests a batch of input tokens into the context
/// @param ctx /// @param ctx

2
llama.h
View file

@ -50,6 +50,8 @@ const std::vector<gpt_vocab::id> llama_tokenize_text(const llama_context& ctx, c
void llama_add_bos(llama_context& ctx); void llama_add_bos(llama_context& ctx);
// Queues up input text to the model input // Queues up input text to the model input
void llama_update_input(llama_context& ctx, const std::string& text); void llama_update_input(llama_context& ctx, const std::string& text);
// Queues up input tokens to the model input
void llama_update_input(llama_context& ctx, const std::vector<gpt_vocab::id>& tokens);
// Ingests input previously added using llama_update_input() // Ingests input previously added using llama_update_input()
void llama_ingest_input_batch(llama_context& ctx); void llama_ingest_input_batch(llama_context& ctx);
// Ingests all input previously added using llama_update_input() in multiple batches // Ingests all input previously added using llama_update_input() in multiple batches