llama.cpp/examples/passkey/passkey.cpp

274 lines
8.3 KiB
C++
Raw Normal View History

#include "arg.h"
#include "common.h"
#include "llama.h"
#include <cmath>
#include <cstdio>
#include <string>
#include <vector>
static void print_usage(int, char ** argv) {
LOG_TEE("\nexample usage:\n");
LOG_TEE("\n %s -m model.gguf --junk 250 --pos 90 --keep 32 --grp-attn-n 2 [--seed 1234]\n", argv[0]);
LOG_TEE("\n");
}
int main(int argc, char ** argv) {
gpt_params params;
params.n_junk = 250;
params.n_keep = 32;
params.i_pos = -1;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PASSKEY, print_usage)) {
return 1;
}
int n_junk = params.n_junk;
int n_keep = params.n_keep;
int n_grp = params.grp_attn_n;
int i_pos = params.i_pos;
if (i_pos == -1) {
i_pos = rand() % n_junk;
}
const std::string prompt_prefix = "There is an important info hidden inside a lot of irrelevant text. Find it and memorize them. I will quiz you about the important information there.";
const std::string prompt_suffix = " What is the pass key? The pass key is";
// generate junk text
params.prompt = prompt_prefix;
const int passkey = rand() % 50000 + 1;
for (int i = 0; i < n_junk; i++) {
if (i % n_junk == i_pos) {
params.prompt += " The pass key is " + std::to_string(passkey) + ". Remember it. " + std::to_string(passkey) + " is the pass key.";
}
params.prompt += " The grass is green. The sky is blue. The sun is yellow. Here we go. There and back again.";
}
params.prompt += prompt_suffix;
// init LLM
ggml : add numa options (#5377) * Added numa options to allow finer grained control as well as plumbing for a new mirror mode that will require numa.h * Reverted Makefile * Fixed include * Removed sched.h from ggml.h, moved ggml_get_numa_affinity into ggml.c, removed trailing whitespace and fixed up a few inconsistent variables * removed trailing whitespace * Added numa options to allow finer grained control as well as plumbing for a new mirror mode that will require numa.h * Reverting Makefile * Fixed a number of issues with the move from BOOL to ggml_numa_strategies. Added a note about mirror mode note being implemented yet * Removing MIRROR_MODE code for this PR * Removing last bit of MIRROR_MODE code for this PR * Removing unneeded branch in server.cpp example and moving get_numa_affinity and making it static * Fixed lingering init_llama_backend() bool calls in tests and examples * Remote enum llama_numa_strategies * Revert bad merge with dynatemp flags * add missing enum ggml_numa_strategies declaration and revert sync problem with master * add missing enum ggml_numa_strategies declaration * fixed ggml_init_numa variable * Update ggml.h Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com> * Update READMEs with info about numa flags, change INTERLEAVE strategy name to DISTRIBUTE everywhere, implement the improved distribution strategy from @rankaiyx, fix a spelling mistake and un-merge some bad merges * split numa init out from llama_backend_init and created llama_numa_init. Updated all code paths and samples * Fix up some boolean vs enum comparisons * Added #ifdefs for non-Linux OS that don't have cpu_set_t datatype * Update ggml.h Align enum values Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Update ggml.c Remove whitespace Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Update ggml.c align paremeters Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Update examples/server/server.cpp remove whitespace and align brace Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Update common/common.cpp Remove whitespace and align brace Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * unified ggml_numa_strategy enum and fixed text alignment in server.cpp example * Update ggml.c simplified return for platforms without NUMA support Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com> * removed redundant else from cli argument processing of --numa * whitespace --------- Co-authored-by: root <root@nenya.lothlorien.ca> Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Jared Van Bortel <jared@nomic.ai>
2024-02-16 09:31:07 +00:00
llama_backend_init();
llama_numa_init(params.numa);
// initialize the model
llama_model_params model_params = llama_model_params_from_gpt_params(params);
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
if (model == NULL) {
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
return 1;
}
// initialize the context
llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
ctx_params.n_ctx = llama_n_ctx_train(model)*n_grp + n_keep;
GGML_ASSERT(ctx_params.n_batch % n_grp == 0 && "n_batch must be divisible by n_grp");
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;
}
auto sparams = llama_sampler_chain_default_params();
llama_sampler * smpl = llama_sampler_chain_init(sparams);
llama_sampler_chain_add(smpl, llama_sampler_init_greedy());
// tokenize the prompt
std::vector<llama_token> tokens_list;
tokens_list = ::llama_tokenize(ctx, params.prompt, true);
// tokenize the prefix and use it as a sink
const int n_tokens_prefix = ::llama_tokenize(ctx, prompt_prefix, true).size();
const int n_tokens_all = tokens_list.size();
// we leave a margin of 16 tokens for the generated text - it should contain just the passkey
const int n_predict = 16;
// total length of the sequences including the prompt
const int n_len = n_tokens_all + n_predict;
const int n_ctx = llama_n_ctx(ctx) - n_keep;
const int n_kv_req = llama_n_ctx(ctx);
const int n_batch = ctx_params.n_batch;
const int n_batch_grp = ctx_params.n_batch/n_grp;
LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d, n_junk = %d, i_pos = %d\n", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch, n_junk, i_pos);
// print the prompt token-by-token
LOG_TEE("\n");
LOG_TEE("prefix tokens: %d\n", n_tokens_prefix);
LOG_TEE("prompt tokens: %d\n", n_tokens_all);
//LOG_TEE("prompt: %s\n", params.prompt.c_str());
llama_batch batch = llama_batch_init(params.n_batch, 0, 1);
int n_past = 0;
// fill the KV cache
for (int i = 0; i < n_ctx; i += n_batch) {
if (i > 0 && n_grp > 1) {
// if SelfExtend is enabled, we compress the position from the last batch by a factor of n_grp
const int ib = i/n_batch - 1;
const int bd = n_batch_grp*(n_grp - 1);
llama_kv_cache_seq_add (ctx, 0, n_past - n_batch, n_past, ib*bd);
llama_kv_cache_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
llama_kv_cache_update (ctx);
n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
}
llama_batch_clear(batch);
for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) {
llama_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
}
if (i + n_batch >= n_tokens_all) {
batch.logits[batch.n_tokens - 1] = true;
}
if (llama_decode(ctx, batch) != 0) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
return 1;
}
LOG_TEE("%s: processed: [%6d, %6d)\n", __func__, i, std::min(i + n_batch, n_tokens_all));
if (i + n_batch >= n_tokens_all) {
break;
}
}
for (int i = n_ctx; i < n_tokens_all; i += n_batch) {
const int n_discard = n_batch;
LOG_TEE("%s: shifting KV cache with %d\n", __func__, n_discard);
llama_kv_cache_seq_rm (ctx, 0, n_keep , n_keep + n_discard);
llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard);
//llama_kv_cache_defrag (ctx);
llama_kv_cache_update (ctx);
n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
llama_batch_clear(batch);
for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) {
llama_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
}
if (i + n_batch >= n_tokens_all) {
batch.logits[batch.n_tokens - 1] = true;
}
if (llama_decode(ctx, batch) != 0) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
return 1;
}
LOG_TEE("%s: processed: [%6d, %6d)\n", __func__, i, std::min(i + n_batch, n_tokens_all));
}
{
const int n_discard = n_past - n_ctx + n_predict;
if (n_discard > 0) {
LOG_TEE("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
llama_kv_cache_seq_rm (ctx, 0, n_keep , n_keep + n_discard);
llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard);
//llama_kv_cache_defrag (ctx);
llama_kv_cache_update (ctx);
n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
}
}
LOG_TEE("\n");
LOG_TEE("%s: passkey = %d, inserted at position %d / %d (token pos: ~%d)\n", __func__, passkey, i_pos, n_junk, (i_pos * n_tokens_all) / n_junk);
LOG_TEE("\n");
// main loop
int n_cur = n_tokens_all;
int n_decode = 0;
LOG_TEE("%s", prompt_suffix.c_str());
fflush(stdout);
const auto t_main_start = ggml_time_us();
while (n_cur <= n_len) {
// sample the next token
{
const llama_token new_token_id = llama_sampler_sample(smpl, ctx, batch.n_tokens - 1);
// is it an end of generation?
if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
LOG_TEE("\n");
break;
}
LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
fflush(stdout);
n_decode += 1;
// prepare the next batch
llama_batch_clear(batch);
// push this new token for next evaluation
llama_batch_add(batch, new_token_id, n_past++, { 0 }, true);
}
n_cur += 1;
// evaluate the current batch with the transformer model
if (llama_decode(ctx, batch)) {
fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
return 1;
}
}
LOG_TEE("\n");
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));
LOG_TEE("\n");
llama_perf_context_print(ctx);
fprintf(stderr, "\n");
llama_sampler_free(smpl);
llama_batch_free(batch);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
return 0;
}