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

formatting, remove copying iterator on delete

Browse source
This commit is contained in:
staviq 2023-10-12 00:26:44 +02:00
parent eac5f5442c
commit f7b1205a51
1 changed files with 23 additions and 89 deletions
Download patch file Download diff file Expand all files Collapse all files

88
llama.cpp
View file

@ -2294,7 +2294,6 @@ static void llm_load_vocab(
is_tokenizable = true; is_tokenizable = true;
break; break;
} }
i++; i++;
} }
else else
@ -2321,7 +2320,6 @@ static void llm_load_vocab(
if (utf8_str_len > 1) if (utf8_str_len > 1)
{ {
// At this point what we have left are special tokens only // At this point what we have left are special tokens only
vocab.special_tokens_cache[token] = id; vocab.special_tokens_cache[token] = id;
// Count manually found special tokens // Count manually found special tokens
@ -2337,7 +2335,7 @@ static void llm_load_vocab(
if( special_tokens_definition_mismatch || special_tokens_count_from_verification != special_tokens_count_by_type ) if( special_tokens_definition_mismatch || special_tokens_count_from_verification != special_tokens_count_by_type )
{ {
fprintf(stderr, "%s: WARNING: Mismatch in special tokens definition ( %u/%zu vs %u/%zu ).\n", fprintf(stderr, "warning: %s: Mismatch in special tokens definition ( %u/%zu vs %u/%zu ).\n",
__func__, __func__,
special_tokens_count_from_verification, vocab.id_to_token.size(), special_tokens_count_from_verification, vocab.id_to_token.size(),
special_tokens_count_by_type, vocab.id_to_token.size() special_tokens_count_by_type, vocab.id_to_token.size()
@ -6608,7 +6606,7 @@ struct fragment_buffer_variant{
const uint64_t length; const uint64_t length;
}; };
#define PRETOKENIZERDEBUG // #define PRETOKENIZERDEBUG
static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<fragment_buffer_variant> & buffer) static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<fragment_buffer_variant> & buffer)
{ {
@ -6619,7 +6617,6 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
const auto & special_id = st.second; const auto & special_id = st.second;
// for each text fragment // for each text fragment
//for (auto & fragment: buffer)
std::forward_list<fragment_buffer_variant>::iterator it = buffer.begin(); std::forward_list<fragment_buffer_variant>::iterator it = buffer.begin();
while (it != buffer.end()) while (it != buffer.end())
{ {
@ -6640,22 +6637,14 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
auto match = raw_text->find( special_token, raw_text_base_offset ); auto match = raw_text->find( special_token, raw_text_base_offset );
// no occurences found, stop processing this fragment for a given special token // no occurences found, stop processing this fragment for a given special token
if (match == std::string::npos) if (match == std::string::npos) break;
{
break;
}
// check if match is within bounds of offset <-> length // check if match is within bounds of offset <-> length
if( match + special_token.length() > raw_text_base_offset + raw_text_base_length ) if (match + special_token.length() > raw_text_base_offset + raw_text_base_length) break;
{
// match is out of bounds
break;
}
#ifdef PRETOKENIZERDEBUG #ifdef PRETOKENIZERDEBUG
fprintf(stderr,"FF: (%ld %ld %ld) '%s'\n", raw_text->length(), raw_text_base_offset, raw_text_base_length, raw_text->substr(raw_text_base_offset, raw_text_base_length).c_str()); fprintf(stderr,"FF: (%ld %ld %ld) '%s'\n", raw_text->length(), raw_text_base_offset, raw_text_base_length, raw_text->substr(raw_text_base_offset, raw_text_base_length).c_str());
#endif #endif
auto source = std::distance(buffer.begin(), it); auto source = std::distance(buffer.begin(), it);
// if match is further than base offset // if match is further than base offset
@ -6663,7 +6652,6 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
if (match > raw_text_base_offset) if (match > raw_text_base_offset)
{ {
// left // left
//buffer.emplace_after(it, raw_text->substr(0, match));
const int64_t left_reminder_offset = raw_text_base_offset + 0; const int64_t left_reminder_offset = raw_text_base_offset + 0;
const int64_t left_reminder_length = match - raw_text_base_offset; const int64_t left_reminder_length = match - raw_text_base_offset;
buffer.emplace_after(it, (*raw_text), left_reminder_offset, left_reminder_length); buffer.emplace_after(it, (*raw_text), left_reminder_offset, left_reminder_length);
@ -6671,7 +6659,6 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
#ifdef PRETOKENIZERDEBUG #ifdef PRETOKENIZERDEBUG
fprintf(stderr,"FL: (%ld %ld) '%s'\n", left_reminder_offset, left_reminder_length, raw_text->substr(left_reminder_offset, left_reminder_length).c_str()); fprintf(stderr,"FL: (%ld %ld) '%s'\n", left_reminder_offset, left_reminder_length, raw_text->substr(left_reminder_offset, left_reminder_length).c_str());
#endif #endif
it++; it++;
} }
@ -6679,16 +6666,9 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
buffer.emplace_after(it, special_id); buffer.emplace_after(it, special_id);
it++; it++;
// right // right
if (match + special_token.length() < raw_text_base_offset + raw_text_base_length) if (match + special_token.length() < raw_text_base_offset + raw_text_base_length)
{ {
/*
| |
-------------------------------------------------------------------------
. |ttttt| |
*/
//buffer.emplace_after(it, raw_text->substr(match + special_token.length()));
const int64_t right_reminder_offset = match + special_token.length(); const int64_t right_reminder_offset = match + special_token.length();
const int64_t right_reminder_length = raw_text_base_length - ((match - raw_text_base_offset) + special_token.length()); const int64_t right_reminder_length = raw_text_base_length - ((match - raw_text_base_offset) + special_token.length());
buffer.emplace_after(it, (*raw_text), right_reminder_offset, right_reminder_length); buffer.emplace_after(it, (*raw_text), right_reminder_offset, right_reminder_length);
@ -6699,66 +6679,25 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
it++; it++;
if (source == 0) if (source == 0) buffer.erase_after(buffer.before_begin());
{ else buffer.erase_after(std::next(buffer.begin(), (source-1)));
// TODO? It might not be needed to store/restore the iterator like this
// but this gives me the peace of mind I'm not causing some
// accidental undefined behaviour.
auto it_backup = std::distance( buffer.begin(), it );
buffer.erase_after(buffer.before_begin());
it = std::next( buffer.begin(), it_backup-1 );
}
else
{
auto it_backup = std::distance( buffer.begin(), it );
//auto prev = std::prev( buffer.begin(), -(source-1) );
auto prev = std::next( buffer.begin(), (source-1) );
buffer.erase_after(prev);
it = std::next( buffer.begin(), it_backup-1 );
}
//it = std::prev( it, 1 );
// repeat for the right side // repeat for the right side
raw_text_base_offset = right_reminder_offset; //match + special_token.length(); raw_text_base_offset = right_reminder_offset;
raw_text_base_length = right_reminder_length; //right_reminder_length - ( ( match + special_token.length() ) - raw_text_base_offset ); raw_text_base_length = right_reminder_length;
//raw_text = &((*it).raw_text);
#ifdef PRETOKENIZERDEBUG #ifdef PRETOKENIZERDEBUG
fprintf(stderr,"RR: (%ld %ld) '%s'\n", raw_text_base_offset, raw_text_base_length, raw_text->substr(raw_text_base_offset, raw_text_base_length).c_str()); fprintf(stderr,"RR: (%ld %ld) '%s'\n", raw_text_base_offset, raw_text_base_length, raw_text->substr(raw_text_base_offset, raw_text_base_length).c_str());
#endif #endif
} }
else else
{ {
if (source == 0) if (source == 0) buffer.erase_after(buffer.before_begin());
{ else buffer.erase_after(std::next(buffer.begin(), (source-1)));
auto it_backup = std::distance( buffer.begin(), it );
buffer.erase_after(buffer.before_begin());
it = std::next( buffer.begin(), it_backup-1 );
}
else
{
auto it_backup = std::distance( buffer.begin(), it );
//auto prev = std::prev( buffer.begin(), -(source) );
auto prev = std::next( buffer.begin(), (source-1) );
buffer.erase_after(prev);
it = std::next( buffer.begin(), it_backup-1 );
}
//it = std::prev( it, 1 );
break; break;
} }
} }
} }
it++; it++;
} }
} }
@ -6781,12 +6720,9 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
} }
std::forward_list<fragment_buffer_variant> fragment_buffer; std::forward_list<fragment_buffer_variant> fragment_buffer;
fragment_buffer.emplace_front( raw_text, 0, raw_text.length() ); fragment_buffer.emplace_front( raw_text, 0, raw_text.length() );
if (special) { if (special) tokenizer_st_partition( vocab, fragment_buffer );
tokenizer_st_partition( vocab, fragment_buffer );
}
switch (vocab.type) { switch (vocab.type) {
case LLAMA_VOCAB_TYPE_SPM: case LLAMA_VOCAB_TYPE_SPM:
@ -6806,7 +6742,6 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
#ifdef PRETOKENIZERDEBUG #ifdef PRETOKENIZERDEBUG
fprintf(stderr,"TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str()); fprintf(stderr,"TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
#endif #endif
llm_tokenizer_spm tokenizer(vocab); llm_tokenizer_spm tokenizer(vocab);
llama_escape_whitespace(raw_text); llama_escape_whitespace(raw_text);
tokenizer.tokenize(raw_text, output); tokenizer.tokenize(raw_text, output);
@ -6828,7 +6763,6 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
#ifdef PRETOKENIZERDEBUG #ifdef PRETOKENIZERDEBUG
fprintf(stderr,"TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str()); fprintf(stderr,"TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
#endif #endif
llm_tokenizer_bpe tokenizer(vocab); llm_tokenizer_bpe tokenizer(vocab);
tokenizer.tokenize(raw_text, output); tokenizer.tokenize(raw_text, output);
} }