mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2026-05-13 20:44:09 +00:00
Merge branch 'master' into compilade/batch-splits
This commit is contained in:
@@ -24,3 +24,18 @@ void llama_log_callback_default(ggml_log_level level, const char * text, void *
|
||||
#define LLAMA_LOG_INFO(...) llama_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__)
|
||||
#define LLAMA_LOG_WARN(...) llama_log_internal(GGML_LOG_LEVEL_WARN , __VA_ARGS__)
|
||||
#define LLAMA_LOG_ERROR(...) llama_log_internal(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
|
||||
|
||||
//
|
||||
// helpers
|
||||
//
|
||||
|
||||
static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
|
||||
if (search.empty()) {
|
||||
return; // Avoid infinite loop if 'search' is an empty string
|
||||
}
|
||||
size_t pos = 0;
|
||||
while ((pos = s.find(search, pos)) != std::string::npos) {
|
||||
s.replace(pos, search.length(), replace);
|
||||
pos += replace.length();
|
||||
}
|
||||
}
|
||||
|
||||
@@ -85,14 +85,14 @@ void llama_sample_top_k_impl(struct llama_sampling * smpl, llama_token_data_arra
|
||||
constexpr float bucket_low = -10.0f;
|
||||
constexpr float bucket_high = 10.0f;
|
||||
constexpr float bucket_scale = nbuckets/(bucket_high - bucket_low);
|
||||
constexpr float bucker_inter = -bucket_low * bucket_scale;
|
||||
constexpr float bucket_inter = -bucket_low * bucket_scale;
|
||||
|
||||
std::vector<int> bucket_idx(candidates->size);
|
||||
std::vector<int> histo(nbuckets, 0);
|
||||
|
||||
for (int i = 0; i < (int)candidates->size; ++i) {
|
||||
const float val = candidates->data[i].logit;
|
||||
int ib = int(bucket_scale * val + bucker_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low);
|
||||
int ib = int(bucket_scale * val + bucket_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low);
|
||||
ib = std::max(0, std::min(nbuckets-1, ib));
|
||||
bucket_idx[i] = ib;
|
||||
++histo[ib];
|
||||
|
||||
@@ -16,20 +16,6 @@
|
||||
// helpers
|
||||
//
|
||||
|
||||
static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
|
||||
std::string result;
|
||||
for (size_t pos = 0; ; pos += search.length()) {
|
||||
auto new_pos = s.find(search, pos);
|
||||
if (new_pos == std::string::npos) {
|
||||
result += s.substr(pos, s.size() - pos);
|
||||
break;
|
||||
}
|
||||
result += s.substr(pos, new_pos - pos) + replace;
|
||||
pos = new_pos;
|
||||
}
|
||||
s = std::move(result);
|
||||
}
|
||||
|
||||
LLAMA_ATTRIBUTE_FORMAT(1, 2)
|
||||
static std::string format(const char * fmt, ...) {
|
||||
va_list ap;
|
||||
@@ -816,6 +802,9 @@ struct llm_tokenizer_ugm {
|
||||
* the best tokenization.
|
||||
*/
|
||||
void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
|
||||
// get current size of output (for reversal later)
|
||||
size_t output_size = output.size();
|
||||
|
||||
// normalize the input first
|
||||
std::string normalized;
|
||||
normalize(text, &normalized);
|
||||
@@ -895,7 +884,7 @@ struct llm_tokenizer_ugm {
|
||||
}
|
||||
|
||||
// reverse the output since we added tokens starting from the end of the input
|
||||
std::reverse(output.begin(), output.end());
|
||||
std::reverse(output.begin() + output_size, output.end());
|
||||
}
|
||||
|
||||
private:
|
||||
@@ -1444,7 +1433,8 @@ llama_token_attr llama_token_get_attr_impl(const struct llama_vocab & vocab, lla
|
||||
bool llama_token_is_eog_impl(const struct llama_vocab & vocab, llama_token token) {
|
||||
return token != -1 && (
|
||||
token == llama_token_eos_impl(vocab) ||
|
||||
token == llama_token_eot_impl(vocab)
|
||||
token == llama_token_eot_impl(vocab) ||
|
||||
token == llama_token_eom_impl(vocab)
|
||||
);
|
||||
}
|
||||
|
||||
@@ -1500,6 +1490,10 @@ llama_token llama_token_eot_impl(const struct llama_vocab & vocab) {
|
||||
return vocab.special_eot_id;
|
||||
}
|
||||
|
||||
llama_token llama_token_eom_impl(const struct llama_vocab & vocab) {
|
||||
return vocab.special_eom_id;
|
||||
}
|
||||
|
||||
int32_t llama_tokenize_impl(
|
||||
const struct llama_vocab & vocab,
|
||||
const char * text,
|
||||
|
||||
@@ -45,6 +45,7 @@ struct llama_vocab {
|
||||
id special_suffix_id = -1;
|
||||
id special_middle_id = -1;
|
||||
id special_eot_id = -1; // TODO: move above after "eos_id", and here add "file separator" token
|
||||
id special_eom_id = -1;
|
||||
|
||||
// tokenizer flags
|
||||
bool tokenizer_add_space_prefix = false;
|
||||
@@ -101,6 +102,7 @@ llama_token llama_token_prefix_impl(const struct llama_vocab & vocab);
|
||||
llama_token llama_token_middle_impl(const struct llama_vocab & vocab);
|
||||
llama_token llama_token_suffix_impl(const struct llama_vocab & vocab);
|
||||
llama_token llama_token_eot_impl (const struct llama_vocab & vocab);
|
||||
llama_token llama_token_eom_impl (const struct llama_vocab & vocab);
|
||||
|
||||
int32_t llama_tokenize_impl(
|
||||
const struct llama_vocab & vocab,
|
||||
|
||||
838
src/llama.cpp
838
src/llama.cpp
File diff suppressed because it is too large
Load Diff
Reference in New Issue
Block a user