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

llama : add handling of byte tokens in UGM tokenizer (same as in SPM)

Browse source 
llama : fix preventing crashes when precompiled_charsmap is not present
This commit is contained in:
Stanisław Szymczyk 2024-06-24 17:39:41 +02:00
parent c2c799cefa
commit f4c03c0966
1 changed files with 39 additions and 35 deletions
Download patch file Download diff file Expand all files Collapse all files

16
llama.cpp
View file

@ -13335,7 +13335,8 @@ static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) {
GGML_ASSERT(llama_is_byte_token(vocab, id)); GGML_ASSERT(llama_is_byte_token(vocab, id));
const auto & token_data = vocab.id_to_token.at(id); const auto & token_data = vocab.id_to_token.at(id);
switch (llama_vocab_get_type(vocab)) { switch (llama_vocab_get_type(vocab)) {
case LLAMA_VOCAB_TYPE_SPM: { case LLAMA_VOCAB_TYPE_SPM:
case LLAMA_VOCAB_TYPE_UGM: {
auto buf = token_data.text.substr(3, 2); auto buf = token_data.text.substr(3, 2);
return strtol(buf.c_str(), NULL, 16); return strtol(buf.c_str(), NULL, 16);
} }
@ -13355,7 +13356,8 @@ static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) {
GGML_ASSERT(llama_vocab_get_type(vocab) != LLAMA_VOCAB_TYPE_NONE); GGML_ASSERT(llama_vocab_get_type(vocab) != LLAMA_VOCAB_TYPE_NONE);
static const char * hex = "0123456789ABCDEF"; static const char * hex = "0123456789ABCDEF";
switch (llama_vocab_get_type(vocab)) { switch (llama_vocab_get_type(vocab)) {
case LLAMA_VOCAB_TYPE_SPM: { case LLAMA_VOCAB_TYPE_SPM:
case LLAMA_VOCAB_TYPE_UGM: {
const char buf[7] = { '<', '0', 'x', hex[ch >> 4], hex[ch & 15], '>', 0 }; const char buf[7] = { '<', '0', 'x', hex[ch >> 4], hex[ch & 15], '>', 0 };
auto token = vocab.token_to_id.find(buf); auto token = vocab.token_to_id.find(buf);
if (token != vocab.token_to_id.end()) { if (token != vocab.token_to_id.end()) {
@ -14242,6 +14244,7 @@ private:
size_t longest_prefix_length = 0; size_t longest_prefix_length = 0;
size_t longest_prefix_offset = 0; size_t longest_prefix_offset = 0;
if (xcda_array_size > 0) {
struct xcda_array_view xcda_view(xcda_array, xcda_array_size); struct xcda_array_view xcda_view(xcda_array, xcda_array_size);
// Find the longest normalized sequence matching the input prefix by walking // Find the longest normalized sequence matching the input prefix by walking
@ -14274,6 +14277,7 @@ private:
longest_prefix_offset = xcda_view.get_value(node_index); longest_prefix_offset = xcda_view.get_value(node_index);
} }
} }
}
if (longest_prefix_length > 0) { if (longest_prefix_length > 0) {
// we have a match, so return the replacement sequence // we have a match, so return the replacement sequence
@ -14299,11 +14303,11 @@ private:
// escaped space symbol - U+2581 (Lower One Eighth Block) // escaped space symbol - U+2581 (Lower One Eighth Block)
const std::string escaped_space = "\xE2\x96\x81"; const std::string escaped_space = "\xE2\x96\x81";
char * prefix_replacements; char * prefix_replacements = NULL;
size_t prefix_replacements_size; size_t prefix_replacements_size = 0;
uint32_t * xcda_array; uint32_t * xcda_array = NULL;
size_t xcda_array_size; size_t xcda_array_size = 0;
struct naive_trie user_defined_token_matcher; struct naive_trie user_defined_token_matcher;