Skip to content

Commit

Permalink
simple : add parallel decoding support
Browse files Browse the repository at this point in the history
  • Loading branch information
@ggerganov
ggerganov committed Sep 20, 2023
1 parent addae65 commit b377bf2
Show file tree
Hide file tree
Showing 7 changed files with 185 additions and 74 deletions.
6 changes: 3 additions & 3 deletions common/common.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -956,19 +956,19 @@ llama_token llama_sample_token(
if (mirostat == 1) {
static float mirostat_mu = 2.0f * mirostat_tau;
const int mirostat_m = 100;
llama_sample_temperature(ctx, &cur_p, temp);
llama_sample_temp(ctx, &cur_p, temp);
id = llama_sample_token_mirostat(ctx, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
} else if (mirostat == 2) {
static float mirostat_mu = 2.0f * mirostat_tau;
llama_sample_temperature(ctx, &cur_p, temp);
llama_sample_temp(ctx, &cur_p, temp);
id = llama_sample_token_mirostat_v2(ctx, &cur_p, mirostat_tau, mirostat_eta, &mirostat_mu);
} else {
// Temperature sampling
llama_sample_top_k (ctx, &cur_p, top_k, 1);
llama_sample_tail_free (ctx, &cur_p, tfs_z, 1);
llama_sample_typical (ctx, &cur_p, typical_p, 1);
llama_sample_top_p (ctx, &cur_p, top_p, 1);
llama_sample_temperature(ctx, &cur_p, temp);
llama_sample_temp(ctx, &cur_p, temp);

{
const int n_top = 10;
Expand Down
8 changes: 4 additions & 4 deletions examples/embd-input/embd-input-lib.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -79,7 +79,7 @@ bool eval_float(void * model, float * input, int N){
if (n_eval > n_batch) {
n_eval = n_batch;
}
llama_batch batch = { uint32_t(n_eval), nullptr, (input+i*n_emb), nullptr, nullptr, nullptr, n_past, 1, 0, };
llama_batch batch = { int32_t(n_eval), nullptr, (input+i*n_emb), nullptr, nullptr, nullptr, n_past, 1, 0, };
if (llama_decode(ctx, batch, params.n_threads)) {
fprintf(stderr, "%s : failed to eval\n", __func__);
return false;
Expand Down Expand Up @@ -183,19 +183,19 @@ llama_token sampling_id(struct MyModel* mymodel) {
if (mirostat == 1) {
static float mirostat_mu = 2.0f * mirostat_tau;
const int mirostat_m = 100;
llama_sample_temperature(ctx, &candidates_p, temp);
llama_sample_temp(ctx, &candidates_p, temp);
id = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
} else if (mirostat == 2) {
static float mirostat_mu = 2.0f * mirostat_tau;
llama_sample_temperature(ctx, &candidates_p, temp);
llama_sample_temp(ctx, &candidates_p, temp);
id = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
} else {
// Temperature sampling
llama_sample_top_k(ctx, &candidates_p, top_k, 1);
llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1);
llama_sample_typical(ctx, &candidates_p, typical_p, 1);
llama_sample_top_p(ctx, &candidates_p, top_p, 1);
llama_sample_temperature(ctx, &candidates_p, temp);
llama_sample_temp(ctx, &candidates_p, temp);
id = llama_sample_token(ctx, &candidates_p);
}
}
Expand Down
8 changes: 4 additions & 4 deletions examples/parallel/parallel.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -123,7 +123,7 @@ int main(int argc, char ** argv) {

std::vector<llama_token> tokens_system;
tokens_system = ::llama_tokenize(ctx, k_system, true);
const uint32_t n_tokens_system = tokens_system.size();
const int32_t n_tokens_system = tokens_system.size();

llama_seq_id g_seq_id = 0;

Expand All @@ -144,7 +144,7 @@ int main(int argc, char ** argv) {

batch.n_tokens = n_tokens_system;

for (uint32_t i = 0; i < batch.n_tokens; ++i) {
for (int32_t i = 0; i < batch.n_tokens; ++i) {
batch.token[i] = tokens_system[i];
batch.pos[i] = i;
batch.seq_id[i] = 0;
Expand All @@ -156,7 +156,7 @@ int main(int argc, char ** argv) {
return 1;
}

// assign the system KV cachce to all parallel sequences
// assign the system KV cache to all parallel sequences
for (int32_t i = 1; i < n_clients; ++i) {
llama_kv_cache_seq_cp(ctx, 0, i, 0, n_tokens_system);
}
Expand Down Expand Up @@ -248,7 +248,7 @@ int main(int argc, char ** argv) {
int32_t n_batch = params.n_batch;

for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
const uint32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));

llama_batch batch_view = {
n_tokens,
Expand Down
6 changes: 3 additions & 3 deletions examples/server/server.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -523,13 +523,13 @@ struct llama_server_context
{
static float mirostat_mu = 2.0f * mirostat_tau;
const int mirostat_m = 100;
llama_sample_temperature(ctx, &candidates_p, temp);
llama_sample_temp(ctx, &candidates_p, temp);
result.tok = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
}
else if (mirostat == 2)
{
static float mirostat_mu = 2.0f * mirostat_tau;
llama_sample_temperature(ctx, &candidates_p, temp);
llama_sample_temp(ctx, &candidates_p, temp);
result.tok = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
}
else
Expand All @@ -540,7 +540,7 @@ struct llama_server_context
llama_sample_tail_free(ctx, &candidates_p, tfs_z, min_keep);
llama_sample_typical(ctx, &candidates_p, typical_p, min_keep);
llama_sample_top_p(ctx, &candidates_p, top_p, min_keep);
llama_sample_temperature(ctx, &candidates_p, temp);
llama_sample_temp(ctx, &candidates_p, temp);
result.tok = llama_sample_token(ctx, &candidates_p);
}
}
Expand Down
182 changes: 144 additions & 38 deletions examples/simple/simple.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -32,12 +32,18 @@ int main(int argc, char ** argv) {
params.prompt = "Hello my name is";
}

// total length of the sequences including the prompt
const int n_len = 32;

// init LLM

llama_backend_init(params.numa);

llama_context_params ctx_params = llama_context_default_params();

ctx_params.seed = 1234;
ctx_params.n_ctx = 2048;

llama_model * model = llama_load_model_from_file(params.model.c_str(), ctx_params);

if (model == NULL) {
Expand All @@ -47,87 +53,187 @@ int main(int argc, char ** argv) {

llama_context * ctx = llama_new_context_with_model(model, ctx_params);

if (ctx == NULL) {
fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
return 1;
}

// tokenize the prompt

std::vector<llama_token> tokens_list;
tokens_list = ::llama_tokenize(ctx, params.prompt, true);

const int max_context_size = llama_n_ctx(ctx);
const int max_tokens_list_size = max_context_size - 4;
const int n_ctx = llama_n_ctx(ctx);
const int n_kv_req = tokens_list.size() + (n_len - tokens_list.size())*n_parallel;

if ((int) tokens_list.size() > max_tokens_list_size) {
fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) tokens_list.size(), max_tokens_list_size);
LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, n_parallel, n_kv_req);

// make sure wi
if (n_kv_req > n_ctx) {
LOG_TEE("%s: error: n_kv_req > n_ctx, the required KV cache size is not big enough\n", __func__);
LOG_TEE("%s: either reduce n_parallel or increase n_ctx\n", __func__);
return 1;
}

fprintf(stderr, "\n\n");
fprintf(stderr, "\n");

for (auto id : tokens_list) {
fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
}

fflush(stderr);

// create a llama_batch with size 512
// we use this object to submit token data for decoding

llama_batch batch = llama_batch_init(512, 0);

// evaluate the initial prompt
batch.n_tokens = tokens_list.size();

for (int32_t i = 0; i < batch.n_tokens; i++) {
batch.token[i] = tokens_list[i];
batch.pos[i] = i;
batch.seq_id[i] = 0;
batch.logits[i] = false;
}

// llama_decode will output logits only for the last token of the prompt
batch.logits[batch.n_tokens - 1] = true;

if (llama_decode(ctx, batch, params.n_threads) != 0) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
return 1;
}

// assign the system KV cache to all parallel sequences
for (int32_t i = 1; i < n_parallel; ++i) {
llama_kv_cache_seq_cp(ctx, 0, i, 0, batch.n_tokens);
}

if (n_parallel > 1) {
LOG_TEE("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
}

// main loop

// The LLM keeps a contextual cache memory of previous token evaluation.
// Usually, once this cache is full, it is required to recompute a compressed context based on previous
// tokens (see "infinite text generation via context swapping" in the main example), but in this minimalist
// example, we will just stop the loop once this cache is full or once an end of stream is detected.
// we will store the parallel decoded sequences in this vector
std::vector<std::string> streams(n_parallel);

const int n_gen = std::min(32, max_context_size);
// remember the batch index of the last tokenn for each parallel sequence
// we will use this to know which logits to sample from
std::vector<int32_t> i_batch(n_parallel, batch.n_tokens - 1);

int n_cur = 0;
int n_cur = batch.n_tokens;
int n_decode = 0;

while (n_cur < n_gen) {
// evaluate the transformer
const auto t_main_start = ggml_time_us();

if (llama_decode(ctx, llama_batch_get_one(tokens_list.data(), int(tokens_list.size()), n_cur, 0), params.n_threads)) {
fprintf(stderr, "%s : failed to eval\n", __func__);
while (n_cur <= n_len) {
// evaluate the current batch with the transformer model
if (llama_decode(ctx, batch, params.n_threads)) {
fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
return 1;
}

n_cur += tokens_list.size();
tokens_list.clear();
// prepare the next batch
batch.n_tokens = 0;

// sample the next token
// sample the next token for each parallel sequence / stream
for (int32_t i = 0; i < n_parallel; ++i) {
if (i_batch[i] < 0) {
// the stream has already finished
continue;
}

llama_token new_token_id = 0;
auto n_vocab = llama_n_vocab(ctx);
auto logits = llama_get_logits(ctx) + i_batch[i] * n_vocab;

auto logits = llama_get_logits(ctx);
auto n_vocab = llama_n_vocab(ctx);
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);

std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
}

for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
}
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };

llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
const int top_k = 40;
const float top_p = 0.9f;
const float temp = 0.4f;

new_token_id = llama_sample_token_greedy(ctx , &candidates_p);
llama_sample_top_k(ctx, &candidates_p, top_k, 1);
llama_sample_top_p(ctx, &candidates_p, top_p, 1);
llama_sample_temp (ctx, &candidates_p, temp);

// is it an end of stream ?
if (new_token_id == llama_token_eos(ctx)) {
fprintf(stderr, " [end of text]\n");
const llama_token new_token_id = llama_sample_token(ctx, &candidates_p);

//const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);

// is it an end of stream ?
// mark this stream as finished
if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) {
i_batch[i] = -1;
LOG_TEE("\n");
if (n_parallel > 1) {
LOG_TEE("%s: stream %d finished", __func__, i);
}

continue;
}

if (n_parallel == 1) {
// print the new token :
LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
fflush(stdout);
}

streams[i] += llama_token_to_piece(ctx, new_token_id);

// push this new token for next evaluation
batch.token [batch.n_tokens] = new_token_id;
batch.pos [batch.n_tokens] = n_cur;
batch.seq_id[batch.n_tokens] = i;
batch.logits[batch.n_tokens] = true;

i_batch[i] = batch.n_tokens;

batch.n_tokens += 1;

n_decode += 1;
}

if (batch.n_tokens == 0) {
// all streams are finished
break;
}

// print the new token :
printf("%s", llama_token_to_piece(ctx, new_token_id).c_str());
fflush(stdout);
n_cur += 1;
}

LOG_TEE("\n");

// push this new token for next evaluation
tokens_list.push_back(new_token_id);
if (n_parallel > 1) {
LOG_TEE("\n");

for (int32_t i = 0; i < n_parallel; ++i) {
LOG_TEE("sequence %d:\n\n%s%s\n\n", i, params.prompt.c_str(), streams[i].c_str());
}
}

const auto t_main_end = ggml_time_us();

LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
__func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));

llama_print_timings(ctx);

fprintf(stderr, "\n");

llama_free(ctx);
llama_free_model(model);

llama_backend_free();

fprintf(stderr, "\n\n");

return 0;
}
Loading

0 comments on commit b377bf2

Please sign in to comment.