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

BERT tokenizer fixes (#6498)

Browse source 
Key changes:
* BERT conversion: fix abuse of LlamaHfVocab, do not set BOS or EOS
* Nomic Embed conversion: pad vocab instead of slicing embedding tensor
* llama_tokenize: handle added special tokens like HF does
This commit is contained in:
Jared Van Bortel 2024-04-09 13:44:08 -04:00 • committed by GitHub
parent c4a3a4ff47
commit 1b67731e18
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
20 changed files with 221 additions and 194 deletions
Download patch file Download diff file Expand all files Collapse all files

167
llama.cpp
View file

@ -318,6 +318,8 @@ enum llm_kv {
LLM_KV_TOKENIZER_UNK_ID,
LLM_KV_TOKENIZER_SEP_ID,
LLM_KV_TOKENIZER_PAD_ID,
LLM_KV_TOKENIZER_CLS_ID,
LLM_KV_TOKENIZER_MASK_ID,
LLM_KV_TOKENIZER_ADD_BOS,
LLM_KV_TOKENIZER_ADD_EOS,
LLM_KV_TOKENIZER_ADD_PREFIX,
@ -388,6 +390,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_TOKENIZER_UNK_ID, "tokenizer.ggml.unknown_token_id" },
{ LLM_KV_TOKENIZER_SEP_ID, "tokenizer.ggml.seperator_token_id" },
{ LLM_KV_TOKENIZER_PAD_ID, "tokenizer.ggml.padding_token_id" },
{ LLM_KV_TOKENIZER_CLS_ID, "tokenizer.ggml.cls_token_id" },
{ LLM_KV_TOKENIZER_MASK_ID, "tokenizer.ggml.mask_token_id" },
{ LLM_KV_TOKENIZER_ADD_BOS, "tokenizer.ggml.add_bos_token" },
{ LLM_KV_TOKENIZER_ADD_EOS, "tokenizer.ggml.add_eos_token" },
{ LLM_KV_TOKENIZER_ADD_PREFIX, "tokenizer.ggml.add_space_prefix" },
@ -2018,11 +2022,13 @@ struct llama_vocab {
std::map<std::pair<std::string, std::string>, int> bpe_ranks;
// default LLaMA special tokens
id special_bos_id = 1;
id special_eos_id = 2;
id special_unk_id = 0;
id special_sep_id = -1;
id special_pad_id = -1;
id special_bos_id = 1;
id special_eos_id = 2;
id special_unk_id = 0;
id special_sep_id = -1;
id special_pad_id = -1;
id special_cls_id = -1;
id special_mask_id = -1;
int special_add_bos = -1; // -1 unknown, 1 add, 0 don't add.
int special_add_eos = -1; // -1 unknown, 1 add, 0 don't add.
@ -3978,7 +3984,9 @@ static void llm_load_hparams(
}
// TODO: This should probably be in llama.h
static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool bos, bool special = false);
static std::vector<llama_vocab::id> llama_tokenize_internal(
const llama_vocab & vocab, std::string raw_text, bool add_special, bool parse_special = false
);
static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch);
static void llm_load_vocab(
@ -4000,23 +4008,27 @@ static void llm_load_vocab(
vocab.type = LLAMA_VOCAB_TYPE_NONE;
// default special tokens
vocab.special_bos_id = -1;
vocab.special_eos_id = -1;
vocab.special_unk_id = -1;
vocab.special_sep_id = -1;
vocab.special_pad_id = -1;
vocab.linefeed_id = -1;
vocab.special_bos_id = -1;
vocab.special_eos_id = -1;
vocab.special_unk_id = -1;
vocab.special_sep_id = -1;
vocab.special_pad_id = -1;
vocab.special_cls_id = -1;
vocab.special_mask_id = -1;
vocab.linefeed_id = -1;
return;
} else if (tokenizer_name == "llama") {
vocab.type = LLAMA_VOCAB_TYPE_SPM;
// default special tokens
vocab.special_bos_id = 1;
vocab.special_eos_id = 2;
vocab.special_unk_id = 0;
vocab.special_sep_id = -1;
vocab.special_pad_id = -1;
vocab.special_bos_id = 1;
vocab.special_eos_id = 2;
vocab.special_unk_id = 0;
vocab.special_sep_id = -1;
vocab.special_pad_id = -1;
vocab.special_cls_id = -1;
vocab.special_mask_id = -1;
const int add_space_prefix_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_ADD_PREFIX).c_str());
if (add_space_prefix_keyidx != -1) {
@ -4051,20 +4063,24 @@ static void llm_load_vocab(
}
// default special tokens
vocab.special_bos_id = 11;
vocab.special_eos_id = 11;
vocab.special_unk_id = -1;
vocab.special_sep_id = -1;
vocab.special_pad_id = -1;
vocab.special_bos_id = 11;
vocab.special_eos_id = 11;
vocab.special_unk_id = -1;
vocab.special_sep_id = -1;
vocab.special_pad_id = -1;
vocab.special_cls_id = -1;
vocab.special_mask_id = -1;
} else if (tokenizer_name == "bert") {
vocab.type = LLAMA_VOCAB_TYPE_WPM;
// default special tokens
vocab.special_bos_id = 101;
vocab.special_eos_id = 102;
vocab.special_unk_id = 100;
vocab.special_sep_id = -1;
vocab.special_pad_id = -1;
vocab.special_bos_id = -1;
vocab.special_eos_id = -1;
vocab.special_unk_id = 100;
vocab.special_sep_id = 102;
vocab.special_pad_id = 0;
vocab.special_cls_id = 101;
vocab.special_mask_id = 103;
vocab.add_space_prefix = false;
} else {
LLAMA_LOG_WARN("%s: unknown tokenizer: '%s'", __func__, tokenizer_name.c_str());
@ -4127,11 +4143,13 @@ static void llm_load_vocab(
// special tokens
{
const std::vector<std::pair<enum llm_kv, int32_t &>> special_token_types = {
{ LLM_KV_TOKENIZER_BOS_ID, vocab.special_bos_id },
{ LLM_KV_TOKENIZER_EOS_ID, vocab.special_eos_id },
{ LLM_KV_TOKENIZER_UNK_ID, vocab.special_unk_id },
{ LLM_KV_TOKENIZER_SEP_ID, vocab.special_sep_id },
{ LLM_KV_TOKENIZER_PAD_ID, vocab.special_pad_id },
{ LLM_KV_TOKENIZER_BOS_ID, vocab.special_bos_id },
{ LLM_KV_TOKENIZER_EOS_ID, vocab.special_eos_id },
{ LLM_KV_TOKENIZER_UNK_ID, vocab.special_unk_id },
{ LLM_KV_TOKENIZER_SEP_ID, vocab.special_sep_id },
{ LLM_KV_TOKENIZER_PAD_ID, vocab.special_pad_id },
{ LLM_KV_TOKENIZER_CLS_ID, vocab.special_cls_id },
{ LLM_KV_TOKENIZER_MASK_ID, vocab.special_mask_id },
};
for (const auto & it : special_token_types) {
const std::string & key = kv(std::get<0>(it));
@ -4323,12 +4341,14 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
LLAMA_LOG_INFO("%s: general.name = %s\n", __func__, model.name.c_str());
// special tokens
if (vocab.special_bos_id != -1) { LLAMA_LOG_INFO( "%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].text.c_str() ); }
if (vocab.special_eos_id != -1) { LLAMA_LOG_INFO( "%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].text.c_str() ); }
if (vocab.special_unk_id != -1) { LLAMA_LOG_INFO( "%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].text.c_str() ); }
if (vocab.special_sep_id != -1) { LLAMA_LOG_INFO( "%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].text.c_str() ); }
if (vocab.special_pad_id != -1) { LLAMA_LOG_INFO( "%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].text.c_str() ); }
if (vocab.linefeed_id != -1) { LLAMA_LOG_INFO( "%s: LF token = %d '%s'\n", __func__, vocab.linefeed_id, vocab.id_to_token[vocab.linefeed_id].text.c_str() ); }
if (vocab.special_bos_id != -1) { LLAMA_LOG_INFO( "%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].text.c_str() ); }
if (vocab.special_eos_id != -1) { LLAMA_LOG_INFO( "%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].text.c_str() ); }
if (vocab.special_unk_id != -1) { LLAMA_LOG_INFO( "%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].text.c_str() ); }
if (vocab.special_sep_id != -1) { LLAMA_LOG_INFO( "%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].text.c_str() ); }
if (vocab.special_pad_id != -1) { LLAMA_LOG_INFO( "%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].text.c_str() ); }
if (vocab.special_cls_id != -1) { LLAMA_LOG_INFO( "%s: CLS token = %d '%s'\n", __func__, vocab.special_cls_id, vocab.id_to_token[vocab.special_cls_id].text.c_str() ); }
if (vocab.special_mask_id != -1) { LLAMA_LOG_INFO( "%s: MASK token = %d '%s'\n", __func__, vocab.special_mask_id, vocab.id_to_token[vocab.special_mask_id].text.c_str() ); }
if (vocab.linefeed_id != -1) { LLAMA_LOG_INFO( "%s: LF token = %d '%s'\n", __func__, vocab.linefeed_id, vocab.id_to_token[vocab.linefeed_id].text.c_str() ); }
}
// Returns false if cancelled by progress_callback
@ -11358,9 +11378,6 @@ struct llm_tokenizer_wpm {
output.push_back(vocab.special_unk_id);
}
}
// append eos token
output.push_back(vocab.special_eos_id);
}
std::vector<std::string> preprocess(const std::string & text) {
@ -11565,30 +11582,28 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
}
}
static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool bos, bool special) {
static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool add_special, bool parse_special) {
std::vector<llama_vocab::id> output;
// OG tokenizer behavior:
//
// tokenizer.encode('', add_bos=True) returns [1]
// tokenizer.encode('', add_bos=False) returns []
if (bos && vocab.special_bos_id != -1) {
output.push_back(vocab.special_bos_id);
}
if (raw_text.empty()) {
return output;
}
std::forward_list<fragment_buffer_variant> fragment_buffer;
fragment_buffer.emplace_front(raw_text, 0, raw_text.length());
if (special) tokenizer_st_partition(vocab, fragment_buffer);
if (!raw_text.empty()) {
fragment_buffer.emplace_front(raw_text, 0, raw_text.length());
if (parse_special) tokenizer_st_partition(vocab, fragment_buffer);
}
switch (vocab.type) {
case LLAMA_VOCAB_TYPE_SPM:
{
// OG tokenizer behavior:
//
// tokenizer.encode('', add_special_tokens=True) returns [1]
// tokenizer.encode('', add_special_tokens=False) returns []
if (add_special && vocab.special_add_bos != 0) {
GGML_ASSERT(vocab.special_bos_id != -1);
output.push_back(vocab.special_bos_id);
}
for (const auto & fragment : fragment_buffer) {
if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
// without adding this leading whitespace, we do not get the same results as the original tokenizer
@ -11614,9 +11629,19 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
output.push_back(fragment.token);
}
}
if (add_special && vocab.special_add_eos == 1) {
GGML_ASSERT(vocab.special_eos_id != -1);
output.push_back(vocab.special_eos_id);
}
} break;
case LLAMA_VOCAB_TYPE_BPE:
{
if (add_special && vocab.special_add_bos == 1) {
GGML_ASSERT(vocab.special_bos_id != -1);
output.push_back(vocab.special_bos_id);
}
for (const auto & fragment : fragment_buffer) {
if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
@ -11630,9 +11655,16 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
output.push_back(fragment.token);
}
}
GGML_ASSERT(vocab.special_add_eos != 1);
} break;
case LLAMA_VOCAB_TYPE_WPM:
{
if (add_special) {
GGML_ASSERT(vocab.special_cls_id != -1);
output.push_back(vocab.special_cls_id);
}
for (const auto & fragment : fragment_buffer) {
if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
@ -11646,6 +11678,11 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
output.push_back(fragment.token);
}
}
if (add_special) {
GGML_ASSERT(vocab.special_sep_id != -1);
output.push_back(vocab.special_sep_id);
}
} break;
case LLAMA_VOCAB_TYPE_NONE:
GGML_ASSERT(false);
@ -16104,6 +16141,14 @@ llama_token llama_token_eos(const struct llama_model * model) {
return model->vocab.special_eos_id;
}
llama_token llama_token_cls(const struct llama_model * model) {
return model->vocab.special_cls_id;
}
llama_token llama_token_sep(const struct llama_model * model) {
return model->vocab.special_sep_id;
}
llama_token llama_token_nl(const struct llama_model * model) {
return model->vocab.linefeed_id;
}
@ -16138,9 +16183,9 @@ int32_t llama_tokenize(
int32_t text_len,
llama_token * tokens,
int32_t n_tokens_max,
bool add_bos,
bool special) {
auto res = llama_tokenize_internal(model->vocab, std::string(text, text_len), add_bos, special);
bool add_special,
bool parse_special) {
auto res = llama_tokenize_internal(model->vocab, std::string(text, text_len), add_special, parse_special);
if (n_tokens_max < (int) res.size()) {
// LLAMA_LOG_ERROR("%s: too many tokens\n", __func__);