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

Add validation for special token ids to llama.cpp

Browse source 
Small optimization for llama_byte_to_token SPM mode
This commit is contained in:
KerfuffleV2 2023-10-15 07:25:21 -06:00
parent 8402566a7c
commit d1075f6e08
1 changed files with 25 additions and 9 deletions
Download patch file Download diff file Expand all files Collapse all files

34
llama.cpp
View file

@ -2235,15 +2235,32 @@ static void llm_load_vocab(
if (vocab.type == LLAMA_VOCAB_TYPE_SPM) { if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
vocab.linefeed_id = llama_byte_to_token(vocab, '\n'); vocab.linefeed_id = llama_byte_to_token(vocab, '\n');
} else { } else {
vocab.linefeed_id = llama_tokenize_internal(vocab, "\u010A", false)[0]; const std::vector<int> ids = llama_tokenize_internal(vocab, "\u010A", false);
GGML_ASSERT(ids.size() == 1 && "model vocab missing newline token");
vocab.linefeed_id = ids[0];
} }
// special tokens // special tokens
GGUF_GET_KEY(ctx, vocab.special_bos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_BOS_ID)); {
GGUF_GET_KEY(ctx, vocab.special_eos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_EOS_ID)); const std::vector<std::tuple<enum llm_kv, int32_t *>> special_token_types = {
GGUF_GET_KEY(ctx, vocab.special_unk_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_UNK_ID)); { LLM_KV_TOKENIZER_BOS_ID, &vocab.special_bos_id },
GGUF_GET_KEY(ctx, vocab.special_sep_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_SEP_ID)); { LLM_KV_TOKENIZER_EOS_ID, &vocab.special_eos_id },
GGUF_GET_KEY(ctx, vocab.special_pad_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_PAD_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 },
};
for (auto & it : special_token_types ) {
int32_t id = -1;
const std::string kstr = kv(std::get<0>(it));
GGUF_GET_KEY(ctx, id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kstr);
if (id != -1 && (id < 0 || size_t(id) >= vocab.id_to_token.size())) {
LLAMA_LOG_WARN("%s: bad special token value %d for key '%s' -- ignoring\n", __func__, id, kstr.c_str());
continue;
}
*(std::get<1>(it)) = id;
}
}
// build special tokens cache // build special tokens cache
{ {
@ -6084,11 +6101,10 @@ static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) {
} }
static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) { static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) {
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: {
char buf[7]; const char buf[7] = { '<', '0', 'x', hex[ch >> 4], hex[ch & 15], '>', 0 };
int result = snprintf(buf, sizeof(buf), "<0x%02X>", ch);
GGML_ASSERT(0 <= result && result < 7);
return vocab.token_to_id.at(buf); return vocab.token_to_id.at(buf);
} }
case LLAMA_VOCAB_TYPE_BPE: { case LLAMA_VOCAB_TYPE_BPE: {