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

Added string debugging, removed bos token from end, added mlock.

Browse source
This commit is contained in:
apaz-cli 2023-09-05 01:42:20 -05:00
parent ca588a3c39
commit 2636a8bad4
1 changed files with 9 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

11
examples/fill-in-middle/FIM.c
View file

@ -12,6 +12,7 @@ For a quick summary of what's going on here, see issue #2818.
static inline struct llama_context* static inline struct llama_context*
codellama_create_fim_context(const char* model_path, const char** error_message) { codellama_create_fim_context(const char* model_path, const char** error_message) {
struct llama_context_params params = llama_context_default_params(); struct llama_context_params params = llama_context_default_params();
params.use_mlock = 1;
struct llama_model* model = llama_load_model_from_file(model_path, params); struct llama_model* model = llama_load_model_from_file(model_path, params);
if (!model) { if (!model) {
*error_message = "Failed to load model."; *error_message = "Failed to load model.";
@ -43,7 +44,7 @@ codellama_fill_in_middle(struct llama_context* ctx, const char* prefix, const ch
// Append first part of prompt // Append first part of prompt
*tokens_end++ = spm ? llama_token_suffix(ctx) : llama_token_prefix(ctx); *tokens_end++ = spm ? llama_token_suffix(ctx) : llama_token_prefix(ctx);
tokens_end += num_tokens = llama_tokenize(ctx, spm ? suffix : prefix, tokens_end, n_max_tokens, 1); tokens_end += num_tokens = llama_tokenize(ctx, spm ? suffix : prefix, tokens_end, n_max_tokens, 0);
if (num_tokens < 0) { if (num_tokens < 0) {
*error_message = "Failed to tokenize the prompt."; *error_message = "Failed to tokenize the prompt.";
free(tokens); free(tokens);
@ -52,7 +53,7 @@ codellama_fill_in_middle(struct llama_context* ctx, const char* prefix, const ch
// Append second part of prompt // Append second part of prompt
*tokens_end++ = spm ? llama_token_prefix(ctx) : llama_token_suffix(ctx); *tokens_end++ = spm ? llama_token_prefix(ctx) : llama_token_suffix(ctx);
tokens_end += num_tokens = llama_tokenize(ctx, spm ? prefix : suffix, tokens_end, n_max_tokens, 1); tokens_end += num_tokens = llama_tokenize(ctx, spm ? prefix : suffix, tokens_end, n_max_tokens, 0);
if (num_tokens < 0) { if (num_tokens < 0) {
*error_message = "Failed to tokenize the prompt."; *error_message = "Failed to tokenize the prompt.";
free(tokens); free(tokens);
@ -112,6 +113,12 @@ codellama_fill_in_middle(struct llama_context* ctx, const char* prefix, const ch
// Append the token, so it's there for subsequent evaluations. // Append the token, so it's there for subsequent evaluations.
generated_tokens[num_generated_tokens++] = likeliest_token; generated_tokens[num_generated_tokens++] = likeliest_token;
// Translate the token to a string.
char cs[20] = {0};
int token_length = llama_token_to_piece(ctx, likeliest_token, cs, 20);
cs[token_length] = '\0';
printf("%s\n", cs);
} }
// Allocate memory for the final result // Allocate memory for the final result