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

Refactor + add 'jina-v2' for testing 'lstrip'

Browse source
This commit is contained in:
jaime-m-p 2024-06-01 21:22:57 +02:00
parent ada961cec2
commit 01c9229186
2 changed files with 44 additions and 35 deletions
Download patch file Download diff file Expand all files Collapse all files

77
llama.cpp
View file

@ -4872,9 +4872,29 @@ static void llm_load_vocab(
//NOTE: Each model customizes per token attributes. //NOTE: Each model customizes per token attributes.
//NOTE: Per token attributes are missing from the GGUF file. //NOTE: Per token attributes are missing from the GGUF file.
//TODO: Merge llama_token_type and llama_token_attrib. //TODO: Merge llama_token_type and llama_token_attrib.
//TODO: Extract attribs from GGUF file.
{ {
auto _contains_any = [] (const std::string &str, const std::vector<std::string> &substrs) -> bool {
for (auto substr : substrs) {
if (str.find(substr) < std::string::npos) {
return true;
}
}
return false;
};
auto _set_tokenid_attrib = [&] (const llama_vocab::id id, llama_token_attrib attrib, bool value) {
uint32_t attribs = vocab.id_to_token.at(id).attribs;
attribs = value ? (attribs | attrib) : (attribs & ~attrib);
vocab.id_to_token[id].attribs = (llama_token_attrib) attribs;
};
auto _set_token_attrib = [&] (const std::string & token, llama_token_attrib attrib, bool value) {
_set_tokenid_attrib(vocab.token_to_id.at(token), attrib, value);
};
// convert token type as an attribute // convert token type as an attribute
for (auto data : vocab.id_to_token) { for (auto &data : vocab.id_to_token) {
uint32_t attrib = LLAMA_TOKEN_ATTRIB_UNDEFINED; uint32_t attrib = LLAMA_TOKEN_ATTRIB_UNDEFINED;
attrib |= LLAMA_TOKEN_ATTRIB_UNKNOWN * (data.type == LLAMA_TOKEN_TYPE_UNKNOWN); attrib |= LLAMA_TOKEN_ATTRIB_UNKNOWN * (data.type == LLAMA_TOKEN_TYPE_UNKNOWN);
attrib |= LLAMA_TOKEN_ATTRIB_UNUSED * (data.type == LLAMA_TOKEN_TYPE_UNUSED); attrib |= LLAMA_TOKEN_ATTRIB_UNUSED * (data.type == LLAMA_TOKEN_TYPE_UNUSED);
@ -4885,44 +4905,31 @@ static void llm_load_vocab(
data.attribs = (llama_token_attrib) attrib; data.attribs = (llama_token_attrib) attrib;
} }
// set attributes by model name
std::string model_name; std::string model_name;
if (ml.get_key(LLM_KV_GENERAL_NAME, model_name, false)) { std::string tokenizer_pre;
std::transform(model_name.begin(), model_name.end(), model_name.begin(),
[] (const std::string::value_type x) {
return std::tolower(x);
}
);
auto _contains_any = [&model_name] (const std::vector<std::string> &substrs) -> bool { ml.get_key(LLM_KV_GENERAL_NAME, model_name, false);
for (auto substr : substrs) { ml.get_key(LLM_KV_TOKENIZER_PRE, tokenizer_pre, false);
if (model_name.find(substr) < std::string::npos) {
return true;
}
}
return false;
};
auto _set_tokenid_attrib = [&] (const llama_vocab::id id, llama_token_attrib attrib, bool value) { // model name to lowercase
uint32_t attribs = vocab.id_to_token[id].attribs; std::transform(model_name.begin(), model_name.end(), model_name.begin(),
attribs = value ? (attribs | attrib) : (attribs & ~attrib); [] (const std::string::value_type x) {
vocab.id_to_token[id].attribs = (llama_token_attrib) attribs; return std::tolower(x);
}; }
);
auto _set_token_attrib = [&] (const std::string & token, llama_token_attrib attrib, bool value) { // set attributes by model/tokenizer name
_set_tokenid_attrib(vocab.token_to_id.at(token), attrib, value); if (_contains_any(tokenizer_pre, {"jina-v2-es", "jina-v2-de"})) {
}; _set_token_attrib("<mask>", LLAMA_TOKEN_ATTRIB_LSTRIP, true);
} else if (_contains_any(model_name, {"phi-3", "phi3"})) {
if (_contains_any({"phi-3", "phi3"})) { for (auto id : vocab.cache_special_tokens) {
for (auto id : vocab.cache_special_tokens) { _set_tokenid_attrib(id, LLAMA_TOKEN_ATTRIB_RSTRIP, true);
_set_tokenid_attrib(id, LLAMA_TOKEN_ATTRIB_RSTRIP, true); }
} for (auto token : {"</s>"}) {
for (auto token : {"</s>"}) { _set_token_attrib(token, LLAMA_TOKEN_ATTRIB_RSTRIP, true);
_set_token_attrib(token, LLAMA_TOKEN_ATTRIB_RSTRIP, true); }
} for (auto token : {"<unk>", "<s>", "<|endoftext|>"}) {
for (auto token : {"<unk>", "<s>", "<|endoftext|>"}) { _set_token_attrib(token, LLAMA_TOKEN_ATTRIB_RSTRIP, false);
_set_token_attrib(token, LLAMA_TOKEN_ATTRIB_RSTRIP, false);
}
} }
} }
} }

2
tests/test-tokenizer-random.py
View file

@ -156,6 +156,8 @@ def generator_custom_text_edge_cases() -> Iterator[str]:
'<s>a', # Phi-3 fail '<s>a', # Phi-3 fail
'<unk><|endoftext|><s>', # Phi-3 fail '<unk><|endoftext|><s>', # Phi-3 fail
'a\na', # TODO: Bert fail 'a\na', # TODO: Bert fail
'a </s> b', # rstrip phi-3
'a <mask> b', # lstrip jina-v2
] ]