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

llama : avoid hardcoded special tokens

Browse source
This commit is contained in:
Georgi Gerganov 2023-08-18 17:29:20 +03:00
parent 035d511457
commit 5d2656d670
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
11 changed files with 61 additions and 65 deletions
Download patch file Download diff file Expand all files Collapse all files

6
examples/embd-input/embd-input-lib.cpp
View file

@ -167,7 +167,7 @@ llama_token sampling_id(struct MyModel* mymodel) {
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
// TODO: Apply penalties
// float nl_logit = logits[llama_token_nl()];
// float nl_logit = logits[llama_token_nl(ctx)];
// auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx);
// llama_sample_repetition_penalty(ctx, &candidates_p,
// last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
@ -176,7 +176,7 @@ llama_token sampling_id(struct MyModel* mymodel) {
// last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
// last_n_repeat, alpha_frequency, alpha_presence);
// if (!penalize_nl) {
// logits[llama_token_nl()] = nl_logit;
// logits[llama_token_nl(ctx)] = nl_logit;
// }
if (temp <= 0) {
@ -211,7 +211,7 @@ const char * sampling(struct MyModel * mymodel) {
llama_context * ctx = mymodel->ctx;
int id = sampling_id(mymodel);
static std::string ret;
if (id == llama_token_eos()) {
if (id == llama_token_eos(ctx)) {
ret = "</s>";
} else {
ret = llama_token_to_str(ctx, id);

4
examples/llama-bench/llama-bench.cpp
View file

@ -851,7 +851,7 @@ struct sql_printer : public printer {
};
static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
std::vector<llama_token> tokens(n_batch, llama_token_bos());
std::vector<llama_token> tokens(n_batch, llama_token_bos(ctx));
int n_processed = 0;
while (n_processed < n_prompt) {
int n_tokens = std::min(n_prompt - n_processed, n_batch);
@ -861,7 +861,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat
}
static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
llama_token token = llama_token_bos();
llama_token token = llama_token_bos(ctx);
for (int i = 0; i < n_gen; i++) {
llama_eval(ctx, &token, 1, n_past + i, n_threads);
}

19
examples/main/main.cpp
View file

@ -143,7 +143,7 @@ int main(int argc, char ** argv) {
{
fprintf(stderr, "%s: testing memory usage for n_batch = %d, n_ctx = %d\n", __func__, params.n_batch, params.n_ctx);
const std::vector<llama_token> tmp(params.n_batch, llama_token_bos());
const std::vector<llama_token> tmp(params.n_batch, llama_token_bos(ctx));
llama_eval(ctx, tmp.data(), tmp.size(), params.n_ctx, params.n_threads);
}
@ -345,10 +345,9 @@ int main(int argc, char ** argv) {
fprintf(stderr, "\n");
{
auto it = params.logit_bias.find(llama_token_eos());
auto it = params.logit_bias.find(llama_token_eos(ctx));
if (it != params.logit_bias.end() && it->second == -INFINITY) {
fprintf(stderr,
"%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
fprintf(stderr, "%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
}
}
@ -398,7 +397,7 @@ int main(int argc, char ** argv) {
// do one empty run to warm up the model
{
const std::vector<llama_token> tmp = { llama_token_bos(), };
const std::vector<llama_token> tmp = { llama_token_bos(ctx), };
llama_eval(ctx, tmp.data(), tmp.size(), 0, params.n_threads);
llama_reset_timings(ctx);
}
@ -582,7 +581,7 @@ int main(int argc, char ** argv) {
}
// Apply penalties
float nl_logit = logits[llama_token_nl()];
float nl_logit = logits[llama_token_nl(ctx)];
auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx);
llama_sample_repetition_penalty(ctx, &candidates_p,
last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
@ -591,7 +590,7 @@ int main(int argc, char ** argv) {
last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
last_n_repeat, alpha_frequency, alpha_presence);
if (!penalize_nl) {
logits[llama_token_nl()] = nl_logit;
logits[llama_token_nl(ctx)] = nl_logit;
}
if (grammar != NULL) {
@ -697,7 +696,7 @@ int main(int argc, char ** argv) {
}
// deal with end of text token in interactive mode
if (last_n_tokens.back() == llama_token_eos()) {
if (last_n_tokens.back() == llama_token_eos(ctx)) {
if (params.interactive) {
if (params.antiprompt.size() != 0) {
// tokenize and inject first reverse prompt
@ -721,7 +720,7 @@ int main(int argc, char ** argv) {
}
if (params.input_prefix_bos) {
embd_inp.push_back(llama_token_bos());
embd_inp.push_back(llama_token_bos(ctx));
}
std::string buffer;
@ -786,7 +785,7 @@ int main(int argc, char ** argv) {
}
// end of text token
if (!embd.empty() && embd.back() == llama_token_eos() && !(params.instruct || params.interactive)) {
if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !(params.instruct || params.interactive)) {
fprintf(stderr, " [end of text]\n");
break;
}

2
examples/perplexity/perplexity.cpp
View file

@ -63,7 +63,7 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
// add BOS token for the first batch of each chunk
if (j == 0) {
tokens[batch_start] = llama_token_bos();
tokens[batch_start] = llama_token_bos(ctx);
}
if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) {

14
examples/server/server.cpp
View file

@ -279,7 +279,7 @@ struct llama_server_context
grammar_parser::print_grammar(stderr, parsed_grammar);
{
auto it = params.logit_bias.find(llama_token_eos());
auto it = params.logit_bias.find(llama_token_eos(ctx));
if (it != params.logit_bias.end() && it->second == -INFINITY) {
LOG_WARNING("EOS token is disabled, which will cause most grammars to fail", {});
}
@ -402,7 +402,7 @@ struct llama_server_context
if (params.n_predict == 0)
{
has_next_token = false;
result.tok = llama_token_eos();
result.tok = llama_token_eos(ctx);
return result;
}
@ -442,7 +442,7 @@ struct llama_server_context
llama_token_data_array candidates_p = {candidates.data(), candidates.size(), false};
// Apply penalties
float nl_logit = logits[llama_token_nl()];
float nl_logit = logits[llama_token_nl(ctx)];
auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), params.n_ctx);
llama_sample_repetition_penalty(ctx, &candidates_p,
last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
@ -452,7 +452,7 @@ struct llama_server_context
last_n_repeat, alpha_frequency, alpha_presence);
if (!penalize_nl)
{
logits[llama_token_nl()] = nl_logit;
logits[llama_token_nl(ctx)] = nl_logit;
}
if (grammar != nullptr) {
@ -515,7 +515,7 @@ struct llama_server_context
// decrement remaining sampling budget
--n_remain;
if (!embd.empty() && embd.back() == llama_token_eos())
if (!embd.empty() && embd.back() == llama_token_eos(ctx))
{
// stopping_word = llama_token_to_str(ctx, embd.back());
has_next_token = false;
@ -949,7 +949,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
static json format_generation_settings(llama_server_context &llama)
{
const auto eos_bias = llama.params.logit_bias.find(llama_token_eos());
const auto eos_bias = llama.params.logit_bias.find(llama_token_eos(llama.ctx));
const bool ignore_eos = eos_bias != llama.params.logit_bias.end() &&
eos_bias->second < 0.0f && std::isinf(eos_bias->second);
@ -1084,7 +1084,7 @@ static void parse_options_completion(const json &body, llama_server_context &lla
llama.params.logit_bias.clear();
if (body.value("ignore_eos", false))
{
llama.params.logit_bias[llama_token_eos()] = -INFINITY;
llama.params.logit_bias[llama_token_eos(llama.ctx)] = -INFINITY;
}
const auto &logit_bias = body.find("logit_bias");

2
examples/simple/simple.cpp
View file

@ -106,7 +106,7 @@ int main(int argc, char ** argv) {
new_token_id = llama_sample_token_greedy(ctx , &candidates_p);
// is it an end of stream ?
if (new_token_id == llama_token_eos()) {
if (new_token_id == llama_token_eos(ctx)) {
fprintf(stderr, " [end of text]\n");
break;
}

16
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -1996,7 +1996,7 @@ void print_tokens_batch(struct llama_context* ctx, struct ggml_tensor * tokens)
}
}
void get_example_targets(const int * train_samples, size_t n_train_samples, const llama_token * train_data, size_t n_train_data, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * target_logits, struct ggml_tensor * target_probs) {
void get_example_targets(struct llama_context * lctx, const int * train_samples, size_t n_train_samples, const llama_token * train_data, size_t n_train_data, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * target_logits, struct ggml_tensor * target_probs) {
int n_tokens = tokens_input->ne[0];
int n_vocab = target_logits->ne[0];
@ -2005,7 +2005,7 @@ void get_example_targets(const int * train_samples, size_t n_train_samples, cons
ggml_set_f32(target_logits, -1.0f/n_vocab);
ggml_set_f32(target_probs, 0.0f);
ggml_set_i32_1d(tokens_input, 0, llama_token_bos());
ggml_set_i32_1d(tokens_input, 0, llama_token_bos(lctx));
for (int i=1; i<n_tokens+1; ++i) {
int token = clamp(train_data[sample+i-1], 0, n_vocab-1);
set_f32_2d(target_logits, token, i-1, +1.0f);
@ -2016,7 +2016,7 @@ void get_example_targets(const int * train_samples, size_t n_train_samples, cons
}
}
void get_example_targets_batch(struct llama_context * /*lctx*/, const int * train_samples, size_t n_train_samples, const llama_token * train_data, size_t n_train_data, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * target_logits, struct ggml_tensor * target_probs) {
void get_example_targets_batch(struct llama_context * lctx, const int * train_samples, size_t n_train_samples, const llama_token * train_data, size_t n_train_data, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * target_logits, struct ggml_tensor * target_probs) {
GGML_ASSERT(tokens_input->n_dims == 2);
GGML_ASSERT(target_logits->n_dims == 3);
GGML_ASSERT(target_probs->n_dims == 3);
@ -2036,7 +2036,7 @@ void get_example_targets_batch(struct llama_context * /*lctx*/, const int * trai
size_t sample = train_samples[(example_id*n_batch + k) % n_train_samples];
GGML_ASSERT(sample+n_tokens-1 < n_train_data);
set_i32_2d(tokens_input, 0, k, llama_token_bos());
set_i32_2d(tokens_input, 0, k, llama_token_bos(lctx));
for (int i=1; i<n_tokens+1; ++i) {
int token = clamp(train_data[sample+i-1], 0, n_vocab-1);
// print_token(lctx, token);
@ -2294,7 +2294,7 @@ llama_token sample(struct my_llama_sampler * sampler, float * logits, const llam
const auto params = sampler->params;
// Apply penalties
const float nl_logit = logits[llama_token_nl()];
const float nl_logit = logits[llama_token_nl(ctx)];
const int n_last = std::min(std::min(n_last_tokens, params.repeat_last_n), sampler->n_ctx);
@ -2313,7 +2313,7 @@ llama_token sample(struct my_llama_sampler * sampler, float * logits, const llam
params.alpha_presence);
if (!params.penalize_nl) {
logits[llama_token_nl()] = nl_logit;
logits[llama_token_nl(ctx)] = nl_logit;
}
llama_token token = 0;
@ -3181,7 +3181,7 @@ int main(int argc, char ** argv) {
std::vector<int> train_samples;
train_samples.push_back(0);
for (int i = 1; i < (int) train_tokens.size() - n_tokens; ++i) {
if (!params.samples_start_after_nl || (train_tokens[i-1] == llama_token_nl())) {
if (!params.samples_start_after_nl || (train_tokens[i-1] == llama_token_nl(lctx))) {
train_samples.push_back(i);
}
}
@ -3341,7 +3341,7 @@ int main(int argc, char ** argv) {
struct ggml_tensor * target_logits = ggml_new_tensor_2d(model.ctx, GGML_TYPE_F32, n_vocab, n_tokens);
struct ggml_tensor * target_probs = ggml_new_tensor_2d(model.ctx, GGML_TYPE_F32, n_vocab, n_tokens);
get_example_targets(train_samples.data(), train_samples.size(), train_tokens.data(), train_tokens.size(), rand()%train_samples.size(), tokens_input, target_logits, target_probs);
get_example_targets(lctx, train_samples.data(), train_samples.size(), train_tokens.data(), train_tokens.size(), rand()%train_samples.size(), tokens_input, target_logits, target_probs);
for (int i=sample_ctx; i<n_tokens; ++i) {
ggml_set_i32_1d(tokens_input, i, n_vocab/2);
}