Feature: decode system and user prompt in batches

This commit is contained in:
Han Yin 2025-04-07 14:20:51 -07:00
parent 02465137ca
commit c14c11dcbd
1 changed files with 73 additions and 49 deletions

View File

@ -111,6 +111,8 @@ static int init_context(llama_model *model) {
// Context parameters setup // Context parameters setup
llama_context_params ctx_params = llama_context_default_params(); llama_context_params ctx_params = llama_context_default_params();
ctx_params.n_ctx = CONTEXT_SIZE; ctx_params.n_ctx = CONTEXT_SIZE;
ctx_params.n_batch = BATCH_SIZE;
ctx_params.n_ubatch = BATCH_SIZE;
ctx_params.n_threads = n_threads; ctx_params.n_threads = n_threads;
ctx_params.n_threads_batch = n_threads; ctx_params.n_threads_batch = n_threads;
auto *context = llama_init_from_model(g_model, ctx_params); auto *context = llama_init_from_model(g_model, ctx_params);
@ -171,9 +173,21 @@ JNIEXPORT void JNICALL
Java_android_llama_cpp_LLamaAndroid_cleanUp(JNIEnv * /*unused*/, jobject /*unused*/) { Java_android_llama_cpp_LLamaAndroid_cleanUp(JNIEnv * /*unused*/, jobject /*unused*/) {
llama_model_free(g_model); llama_model_free(g_model);
llama_free(g_context); llama_free(g_context);
llama_backend_free();
delete g_batch; delete g_batch;
common_sampler_free(g_sampler); common_sampler_free(g_sampler);
llama_backend_free();
}
static std::string get_backend() {
std::vector<std::string> backends;
for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
auto * reg = ggml_backend_reg_get(i);
std::string name = ggml_backend_reg_name(reg);
if (name != "CPU") {
backends.push_back(ggml_backend_reg_name(reg));
}
}
return backends.empty() ? "CPU" : join(backends, ",");
} }
extern "C" extern "C"
@ -205,7 +219,7 @@ Java_android_llama_cpp_LLamaAndroid_benchModel(JNIEnv *env, jobject /*unused*/,
const auto t_pp_start = ggml_time_us(); const auto t_pp_start = ggml_time_us();
if (llama_decode(g_context, *g_batch) != 0) { if (llama_decode(g_context, *g_batch) != 0) {
LOGw("llama_decode() failed during prompt processing"); LOGe("llama_decode() failed during prompt processing");
} }
const auto t_pp_end = ggml_time_us(); const auto t_pp_end = ggml_time_us();
@ -216,18 +230,15 @@ Java_android_llama_cpp_LLamaAndroid_benchModel(JNIEnv *env, jobject /*unused*/,
llama_memory_clear(llama_get_memory(g_context), false); llama_memory_clear(llama_get_memory(g_context), false);
const auto t_tg_start = ggml_time_us(); const auto t_tg_start = ggml_time_us();
for (i = 0; i < tg; i++) { for (i = 0; i < tg; i++) {
common_batch_clear(*g_batch); common_batch_clear(*g_batch);
for (j = 0; j < pl; j++) { for (j = 0; j < pl; j++) {
common_batch_add(*g_batch, 0, i, {j}, true); common_batch_add(*g_batch, 0, i, {j}, true);
} }
LOGi("llama_decode() text generation: %d", i);
if (llama_decode(g_context, *g_batch) != 0) { if (llama_decode(g_context, *g_batch) != 0) {
LOGw("llama_decode() failed during text generation"); LOGe("llama_decode() failed during text generation");
} }
} }
const auto t_tg_end = ggml_time_us(); const auto t_tg_end = ggml_time_us();
llama_memory_clear(llama_get_memory(g_context), false); llama_memory_clear(llama_get_memory(g_context), false);
@ -282,13 +293,42 @@ Java_android_llama_cpp_LLamaAndroid_benchModel(JNIEnv *env, jobject /*unused*/,
/** /**
* Prediction loop's long-term and short-term states * Prediction loop's long-term and short-term states
*/ */
static int current_position; static llama_pos current_position;
static int token_predict_budget; static llama_pos token_predict_budget;
static std::string cached_token_chars; static std::string cached_token_chars;
int token_predict_budget; static int decode_tokens_in_batches(
std::string cached_token_chars; llama_context *context,
const llama_tokens& tokens,
const llama_pos start_pos,
bool compute_last_logit = false,
llama_batch *batch = g_batch) {
// Process tokens in batches using the global batch
LOGd("Decode %d tokens starting at position %d", tokens.size(), start_pos);
for (int i = 0; i < (int) tokens.size(); i += BATCH_SIZE) {
int cur_batch_size = std::min((int) tokens.size() - i, BATCH_SIZE);
common_batch_clear(*batch);
LOGv("Preparing a batch size of %d starting at: %d", cur_batch_size, i);
// Add tokens to the batch with proper positions
for (int j = 0; j < cur_batch_size; j++) {
llama_token token_id = tokens[i + j];
llama_pos position = start_pos + i + j;
bool want_logit = compute_last_logit && (i + j == tokens.size() - 1);
common_batch_add(*batch, token_id, position, {0}, want_logit);
}
// Decode this batch
int decode_result = llama_decode(context, *batch);
if (decode_result) {
LOGe("llama_decode failed w/ %d", decode_result);
return -1;
}
}
return 0;
}
extern "C" extern "C"
JNIEXPORT jint JNICALL JNIEXPORT jint JNICALL
@ -316,22 +356,14 @@ Java_android_llama_cpp_LLamaAndroid_processSystemPrompt(
LOGv("token: `%s`\t -> `%d`", common_token_to_piece(g_context, id).c_str(), id); LOGv("token: `%s`\t -> `%d`", common_token_to_piece(g_context, id).c_str(), id);
} }
// Add system prompt tokens to batch // Decode system tokens in batches
common_batch_clear(*g_batch); if (decode_tokens_in_batches(g_context, system_tokens, current_position)) {
// TODO-hyin: support batch processing! LOGe("llama_decode() failed!");
for (int i = 0; i < system_tokens.size(); i++) {
common_batch_add(*g_batch, system_tokens[i], i, {0}, false);
}
// Decode batch
int decode_result = llama_decode(g_context, *g_batch);
if (decode_result != 0) {
LOGe("llama_decode() failed: %d", decode_result);
return -1; return -1;
} }
// Update position // Update position
current_position = system_tokens.size(); current_position = (int) system_tokens.size();
return 0; return 0;
} }
@ -360,29 +392,21 @@ Java_android_llama_cpp_LLamaAndroid_processUserPrompt(
} }
// Check if context space is enough for desired tokens // Check if context space is enough for desired tokens
int desired_budget = current_position + user_tokens.size() + n_predict; int desired_budget = current_position + (int) user_tokens.size() + n_predict;
if (desired_budget > llama_n_ctx(g_context)) { if (desired_budget > llama_n_ctx(g_context)) {
LOGe("error: total tokens exceed context size"); LOGe("error: total tokens exceed context size");
return -1; return -1;
} }
token_predict_budget = desired_budget; token_predict_budget = desired_budget;
// Add user prompt tokens to batch // Decode user tokens in batches
common_batch_clear(*g_batch); if (decode_tokens_in_batches(g_context, user_tokens, current_position, true)) {
for (int i = 0; i < user_tokens.size(); i++) { LOGe("llama_decode() failed!");
common_batch_add(*g_batch, user_tokens[i], current_position + i, {0}, false);
}
g_batch->logits[g_batch->n_tokens - 1] = true; // Set logits true only for last token
// Decode batch
int decode_result = llama_decode(g_context, *g_batch);
if (decode_result != 0) {
LOGe("llama_decode() failed: %d", decode_result);
return -2; return -2;
} }
// Update position // Update position
current_position += user_tokens.size(); // Update position current_position += (int) user_tokens.size(); // Update position
return 0; return 0;
} }
@ -436,13 +460,7 @@ Java_android_llama_cpp_LLamaAndroid_predictLoop(
const auto new_token_id = common_sampler_sample(g_sampler, g_context, -1); const auto new_token_id = common_sampler_sample(g_sampler, g_context, -1);
common_sampler_accept(g_sampler, new_token_id, true); common_sampler_accept(g_sampler, new_token_id, true);
// Stop if next token is EOG // Populate the batch with new token, then decode
if (llama_vocab_is_eog(llama_model_get_vocab(g_model), new_token_id)) {
LOGd("id: %d,\tIS EOG!\nSTOP.", new_token_id);
return nullptr;
}
// Update the context with the new token
common_batch_clear(*g_batch); common_batch_clear(*g_batch);
common_batch_add(*g_batch, new_token_id, current_position, {0}, true); common_batch_add(*g_batch, new_token_id, current_position, {0}, true);
if (llama_decode(g_context, *g_batch) != 0) { if (llama_decode(g_context, *g_batch) != 0) {
@ -450,22 +468,28 @@ Java_android_llama_cpp_LLamaAndroid_predictLoop(
return nullptr; return nullptr;
} }
// Convert to text // Update position
current_position++;
// Stop if next token is EOG
if (llama_vocab_is_eog(llama_model_get_vocab(g_model), new_token_id)) {
LOGd("id: %d,\tIS EOG!\nSTOP.", new_token_id);
return nullptr;
}
// If not EOG, convert to text
auto new_token_chars = common_token_to_piece(g_context, new_token_id); auto new_token_chars = common_token_to_piece(g_context, new_token_id);
cached_token_chars += new_token_chars; cached_token_chars += new_token_chars;
// Create Java string // Create and return Java string
jstring result = nullptr; jstring result = nullptr;
if (is_valid_utf8(cached_token_chars.c_str())) { if (is_valid_utf8(cached_token_chars.c_str())) {
result = env->NewStringUTF(cached_token_chars.c_str()); result = env->NewStringUTF(cached_token_chars.c_str());
LOGd("id: %d,\tcached: `%s`,\tnew: `%s`", new_token_id, cached_token_chars.c_str(), new_token_chars.c_str()); LOGv("id: %d,\tcached: `%s`,\tnew: `%s`", new_token_id, cached_token_chars.c_str(), new_token_chars.c_str());
cached_token_chars.clear(); cached_token_chars.clear();
} else { } else {
LOGd("id: %d,\tappend to cache", new_token_id); LOGv("id: %d,\tappend to cache", new_token_id);
result = env->NewStringUTF(""); result = env->NewStringUTF("");
} }
// Update position
current_position++;
return result; return result;
} }