Skip to content

More checks before assuming FIM tokens #7644

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 4 commits into from
Jun 14, 2024
Merged

More checks before assuming FIM tokens #7644

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
998d208
More checks before assuming FIM tokens for Llama arch
CISC May 30, 2024
29a9884
Merge branch 'ggerganov:master' into non-llama-fim-fix
CISC May 30, 2024
f75eedf
extensive token check
CISC Jun 3, 2024
f4df3ff
Merge branch 'ggerganov:master' into non-llama-fim-fix
CISC Jun 9, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
68 changes: 39 additions & 29 deletions llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4553,35 +4553,6 @@ static void llm_load_vocab(
vocab.special_cls_id = -1;
vocab.special_mask_id = -1;

// For Fill-In-the-Middle (FIM)/infill models which where converted
// prior to support of FIM special tokens in GGUF, the following
// will allow those models to continue to work. The general names
// of the known models are currently CodeLlama (LLM_ARCH_LLAMA) and
// CodeGemma (LLM_ARCH_GEMMA). This can potentially be removed once
// new versions of these models have been published.
std::string gen_name;
ml.get_key(LLM_KV_GENERAL_NAME, gen_name, false);

std::transform(gen_name.begin(), gen_name.end(), gen_name.begin(),
[](unsigned char c){ return std::tolower(c); });

if (gen_name.find("code") != std::string::npos) {
if (model.arch == LLM_ARCH_LLAMA) {
vocab.special_prefix_id = 32007;
vocab.special_suffix_id = 32008;
vocab.special_middle_id = 32009;
vocab.special_eot_id = 32010;
} else if (model.arch == LLM_ARCH_GEMMA) {
vocab.special_prefix_id = 67;
vocab.special_suffix_id = 69;
vocab.special_middle_id = 68;
// TODO: this is not EOT, it is "file separator" token, needs fix
// https://huggingface.co/google/codegemma-7b-it/blob/9b1d9231388358c04d90bd003458f5070d97db44/tokenizer_config.json#L565-L572
//vocab.special_eot_id = 70;
vocab.special_eot_id = 107;
}
}

const int add_space_prefix_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_ADD_PREFIX).c_str());
if (add_space_prefix_keyidx != -1) {
vocab.add_space_prefix = gguf_get_val_bool(ctx, add_space_prefix_keyidx);
Expand Down Expand Up @@ -4762,6 +4733,45 @@ static void llm_load_vocab(

// determine the newline token: LLaMA "<0x0A>" == 10 == '\n', Falcon 193 == '\n'
if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
// For Fill-In-the-Middle (FIM)/infill models which where converted
// prior to support of FIM special tokens in GGUF, the following
// will allow those models to continue to work. The general names
// of the known models are currently CodeLlama (LLM_ARCH_LLAMA) and
// CodeGemma (LLM_ARCH_GEMMA). This can potentially be removed once
// new versions of these models have been published.
std::string gen_name;
ml.get_key(LLM_KV_GENERAL_NAME, gen_name, false);

std::transform(gen_name.begin(), gen_name.end(), gen_name.begin(),
[](unsigned char c){ return std::tolower(c); });

if (gen_name.find("code") != std::string::npos) {
if (model.arch == LLM_ARCH_LLAMA
&& 32010 < vocab.id_to_token.size()
&& vocab.id_to_token[32007].text == "<PRE>"
&& vocab.id_to_token[32008].text == "<SUF>"
&& vocab.id_to_token[32009].text == "<MID>"
&& vocab.id_to_token[32010].text == "<EOT>") {
vocab.special_prefix_id = 32007;
vocab.special_suffix_id = 32008;
vocab.special_middle_id = 32009;
vocab.special_eot_id = 32010;
} else if (model.arch == LLM_ARCH_GEMMA
&& 107 < vocab.id_to_token.size()
&& vocab.id_to_token[67].text == "<|fim_prefix|>"
&& vocab.id_to_token[69].text == "<|fim_suffix|>"
&& vocab.id_to_token[68].text == "<|fim_middle|>"
&& vocab.id_to_token[107].text == "<end_of_turn>") {
vocab.special_prefix_id = 67;
vocab.special_suffix_id = 69;
vocab.special_middle_id = 68;
// TODO: this is not EOT, it is "file separator" token, needs fix
// https://huggingface.co/google/codegemma-7b-it/blob/9b1d9231388358c04d90bd003458f5070d97db44/tokenizer_config.json#L565-L572
//vocab.special_eot_id = 70;
vocab.special_eot_id = 107;
}
}

try {
vocab.linefeed_id = llama_byte_to_token(vocab, '\n');
} catch (const std::exception & e) {
Expand Down
Loading