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

llama : support Llama 3 HF conversion (#6745)

Browse source 
* Support Llama 3 conversion

The tokenizer is BPE.

* style

* Accept suggestion

Co-authored-by: Sourab Mangrulkar <13534540+pacman100@users.noreply.github.com>

* llama : add llama_token_is_eog()

ggml-ci

* llama : auto-detect more EOT tokens when missing in KV data

* convert : replacing EOS token is a hack

* llama : fix codegemma EOT token + add TODOs

* llama : fix model type string for 8B model

---------

Co-authored-by: Sourab Mangrulkar <13534540+pacman100@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Pedro Cuenca 2024-04-21 13:50:41 +02:00 • committed by GitHub
parent b8109bc013
commit b97bc3966e
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
20 changed files with 123 additions and 64 deletions
Download patch file Download diff file Expand all files Collapse all files

2
examples/batched.swift/Sources/main.swift
View file

@ -153,7 +153,7 @@ while n_cur <= n_len {
// const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
// is it an end of stream? -> mark the stream as finished
if new_token_id == llama_token_eos(model) || n_cur == n_len {
if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
i_batch[i] = -1
// print("")
if n_parallel > 1 {

4
examples/batched/batched.cpp
View file

@ -191,8 +191,8 @@ int main(int argc, char ** argv) {
//const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
// is it an end of stream? -> mark the stream as finished
if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
// is it an end of generation? -> mark the stream as finished
if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
i_batch[i] = -1;
LOG_TEE("\n");
if (n_parallel > 1) {

2
examples/beam-search/beam-search.cpp
View file

@ -47,7 +47,7 @@ struct beam_search_callback_data {
// In this case, end-of-beam (eob) is equivalent to end-of-sentence (eos) but this need not always be the same.
// For example, eob can be flagged due to maximum token length, stop words, etc.
static bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, size_t n_tokens) {
return n_tokens && tokens[n_tokens-1] == llama_token_eos(llama_get_model(callback_data.ctx));
return n_tokens && llama_token_is_eog(llama_get_model(callback_data.ctx), tokens[n_tokens-1]);
}
// Function matching type llama_beam_search_callback_fn_t.

10
examples/infill/infill.cpp
View file

@ -586,7 +586,7 @@ int main(int argc, char ** argv) {
// deal with eot token in infill mode
if ((llama_sampling_last(ctx_sampling) == llama_token_eot(model) || is_interacting) && params.interactive){
if(is_interacting && !params.interactive_first) {
if (is_interacting && !params.interactive_first) {
// print an eot token
printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
}
@ -651,8 +651,8 @@ int main(int argc, char ** argv) {
// LOG_TEE("took new input\n");
is_interacting = false;
}
// deal with end of text token in interactive mode
else if (llama_sampling_last(ctx_sampling) == llama_token_eos(model)) {
// deal with end of generation tokens in interactive mode
else if (llama_token_is_eog(model, llama_sampling_last(ctx_sampling))) {
LOG("found EOS token\n");
if (params.interactive) {
@ -731,8 +731,8 @@ int main(int argc, char ** argv) {
}
}
// end of text token
if (!embd.empty() && embd.back() == llama_token_eos(model) && !params.interactive) {
// end of generation
if (!embd.empty() && llama_token_is_eog(model, embd.back()) && !params.interactive) {
break;
}

2
examples/llama.android/app/src/main/cpp/llama-android.cpp
View file

@ -408,7 +408,7 @@ Java_com_example_llama_Llm_completion_1loop(
const auto new_token_id = llama_sample_token_greedy(context, &candidates_p);
const auto n_cur = env->CallIntMethod(intvar_ncur, la_int_var_value);
if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
return env->NewStringUTF("");
}

2
examples/llama.swiftui/llama.cpp.swift/LibLlama.swift
View file

@ -158,7 +158,7 @@ actor LlamaContext {
new_token_id = llama_sample_token_greedy(context, &candidates_p)
}
if new_token_id == llama_token_eos(model) || n_cur == n_len {
if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
print("\n")
let new_token_str = String(cString: temporary_invalid_cchars + [0])
temporary_invalid_cchars.removeAll()

2
examples/llava/llava-cli.cpp
View file

@ -45,7 +45,7 @@ static const char * sample(struct llama_sampling_context * ctx_sampling,
const llama_token id = llama_sampling_sample(ctx_sampling, ctx_llama, NULL);
llama_sampling_accept(ctx_sampling, ctx_llama, id, true);
static std::string ret;
if (id == llama_token_eos(llama_get_model(ctx_llama))) {
if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
ret = "</s>";
} else {
ret = llama_token_to_piece(ctx_llama, id);

2
examples/lookahead/lookahead.cpp
View file

@ -299,7 +299,7 @@ int main(int argc, char ** argv) {
}
fflush(stdout);
if (id == llama_token_eos(model)) {
if (llama_token_is_eog(model, id)) {
has_eos = true;
}

2
examples/lookup/lookup.cpp
View file

@ -141,7 +141,7 @@ int main(int argc, char ** argv){
printf("%s", token_str.c_str());
}
if (id == llama_token_eos(model)) {
if (llama_token_is_eog(model, id)) {
has_eos = true;
}

8
examples/main/main.cpp
View file

@ -795,8 +795,8 @@ int main(int argc, char ** argv) {
}
}
// deal with end of text token in interactive mode
if (llama_sampling_last(ctx_sampling) == llama_token_eos(model)) {
// deal with end of generation tokens in interactive mode
if (llama_token_is_eog(model, llama_sampling_last(ctx_sampling))) {
LOG("found EOS token\n");
if (params.interactive) {
@ -920,8 +920,8 @@ int main(int argc, char ** argv) {
}
}
// end of text token
if (!embd.empty() && embd.back() == llama_token_eos(model) && !(params.instruct || params.interactive || params.chatml)) {
// end of generation
if (!embd.empty() && llama_token_is_eog(model, embd.back()) && !(params.instruct || params.interactive || params.chatml)) {
LOG_TEE(" [end of text]\n");
break;
}

2
examples/parallel/parallel.cpp
View file

@ -359,7 +359,7 @@ int main(int argc, char ** argv) {
// client.id, client.seq_id, id, client.n_decoded, client.i_batch, token_str.c_str());
if (client.n_decoded > 2 &&
(id == llama_token_eos(model) ||
(llama_token_is_eog(model, id) ||
(params.n_predict > 0 && client.n_decoded + client.n_prompt >= params.n_predict) ||
client.response.find("User:") != std::string::npos ||
client.response.find('\n') != std::string::npos)) {

4
examples/passkey/passkey.cpp
View file

@ -252,8 +252,8 @@ int main(int argc, char ** argv) {
// sample the most likely token
const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
// is it an end of stream?
if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
// is it an end of generation?
if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
LOG_TEE("\n");
break;

2
examples/server/server.cpp
View file

@ -1201,7 +1201,7 @@ struct server_context {
});
}
if (result.tok == llama_token_eos(model)) {
if (llama_token_is_eog(model, result.tok)) {
slot.stopped_eos = true;
slot.has_next_token = false;

4
examples/server/utils.hpp
View file

@ -381,10 +381,6 @@ static json oaicompat_completion_params_parse(
} else {
llama_params["stop"] = json_value(body, "stop", json::array());
}
// Some chat templates don't use EOS token to stop generation
// We must add their end sequences to list of stop words
llama_params["stop"].push_back("<|im_end|>"); // chatml
llama_params["stop"].push_back("<end_of_turn>"); // gemma
// Handle "response_format" field
if (body.contains("response_format")) {

4
examples/simple/simple.cpp
View file

@ -133,8 +133,8 @@ int main(int argc, char ** argv) {
// sample the most likely token
const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
// is it an end of stream?
if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
// is it an end of generation?
if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
LOG_TEE("\n");
break;

2
examples/speculative/speculative.cpp
View file

@ -360,7 +360,7 @@ int main(int argc, char ** argv) {
}
}
if (token_id == llama_token_eos(model_tgt)) {
if (llama_token_is_eog(model_tgt, token_id)) {
has_eos = true;
}
++n_predict;