llama.cpp/examples/speculative/speculative.cpp

639 lines
24 KiB
C++
Raw Normal View History

#include "arg.h"
#include "common.h"
#include "sampling.h"
#include "log.h"
#include "llama.h"
#include <algorithm>
#include <cstdio>
#include <cstring>
#include <random>
#include <set>
#include <string>
#include <vector>
#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE 100
#define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
struct seq_draft {
bool active = false;
bool drafting = false;
bool skip = false;
int i_batch_dft = 0;
std::vector<int> i_batch_tgt;
std::vector<llama_token> tokens;
std::vector<std::vector<llama_token_data>> dists;
struct gpt_sampler * smpl = nullptr;
};
int main(int argc, char ** argv) {
gpt_params params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) {
return 1;
}
gpt_init();
if (params.model_draft.empty()) {
LOG_ERR("%s: --model-draft is required\n", __func__);
return 1;
}
// max number of parallel drafting sequences (i.e. tree branches)
const int n_seq_dft = params.n_parallel;
// probability threshold for splitting a draft branch (only for n_seq_dft > 1)
const float p_split = params.p_split;
std::default_random_engine rng(params.sparams.seed);
std::uniform_real_distribution<> u_dist;
// init llama.cpp
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);
llama_model * model_tgt = NULL;
llama_model * model_dft = NULL;
llama_context * ctx_tgt = NULL;
llama_context * ctx_dft = NULL;
// load the target model
llama_init_result llama_init_tgt = llama_init_from_gpt_params(params);
model_tgt = llama_init_tgt.model;
ctx_tgt = llama_init_tgt.context;
// load the draft model
params.model = params.model_draft;
params.n_gpu_layers = params.n_gpu_layers_draft;
Threadpool: take 2 (#8672) * Introduce ggml_compute_threadpool - OpenMP functional: check - Vanilla ggml functional: Check - ggml w/threadpool functional: Check - OpenMP no regression: No glaring problems - Vanilla ggml no regression: No glaring problems - ggml w/threadpool no regression: No glaring problems * Minor fixes * fixed use after release bug * fixed a harmless race condition * Fix Android bulid issue * fix more race conditions * fix deadlock for cases where cgraph.n_nodes == 1 and fix --poll case * threadpool: use cpu_get_num_math to set the default number of threadpool threads This way we avoid using E-Cores and Hyperthreaded siblings. * bench: create fresh threadpool for each test For benchmarking it's better to start a fresh pool for each test with the exact number of threads needed for that test. Having larger pools is suboptimal (causes more load, etc). * atomics: always use stdatomics with clang and use relaxed memory order when polling in ggml_barrier This also removes sched_yield() calls from ggml_barrier() to match OpenMP behavior. * threadpool: make polling the default to match openmp behavior All command line args now allow for setting poll to 0 (false). * threadpool: do not wakeup threads in already paused threadpool * fix potential race condition in check_for_work * threadpool: do not create two threadpools if their params are identical * threadpool: reduce pause/resume/wakeup overhead in common cases We now start threadpool in paused state only if we have two. The resume is now implicit (ie new work) which allows for reduced locking and context-switch overhead. * threadpool: add support for hybrid polling poll params (--poll, ...) now specify "polling level", i.e. how aggresively we poll before waiting on cond.var. poll=0 means no polling, 1 means poll for 128K rounds then wait, 2 for 256K rounds, ... The default value of 50 (ie 50x128K rounds) seems like a decent default across modern platforms. We can tune this further as things evolve. * threadpool: reduce the number of barrier required New work is now indicated with an atomic counter that is incremented for each new graph that needs to be computed. This removes the need for extra barrier for clearing the "new_work" and removes the special case for trivial graphs. * threadpool: remove special-casing for disposable threadpools With the efficient hybrid polling there is no need to make disposable pools any different. This simplifies the overall logic and reduces branching. Include n_threads in debug print for disposable threadpool. Declare pause and stop flags as atomic_bool This doesn't actually generate any memory barriers and simply informs the thread sanitizer that these flags can be written & read by different threads without locking. * threadpool: do not clear barrier counters between graphs computes (fixes race with small graphs) This fixes the race condition with very small graphs where the main thread happens to start a new graph while the workers are just about to exit from barriers. * threadpool: use relaxed order for chunk sync Full memory barrier is an overkill for this since each thread works on different chunk * threadpool: remove abort_callback from threadpool state * threadpool: better naming for thread/cpumask releated functions * threadpool: consistent use of int type for n_threads params * threadpool: add support for ggml_threadpool_params_default/init Also removes the need for explicit mask_specified param. all-zero cpumask means use default (usually inherited) cpu affinity mask. * threadpool: move typedef into ggml.h * threadpool: fix apply_priority() function name * threadpool: fix swift wrapper errors due to n_threads int type cleanup * threadpool: enable --cpu-mask and other threadpool related options only if threadpool is enabled * threadpool: replace checks for compute_thread ret code with proper status check * threadpool: simplify threadpool init logic and fix main thread affinity application Most of the init code is now exactly the same between threadpool and openmp. * threadpool: update threadpool resume/pause function names * threadpool: enable openmp by default for now * threadpool: don't forget to free workers state when omp is enabled * threadpool: avoid updating process priority on the platforms that do not require it On Windows we need to change overall process priority class in order to set thread priorities, but on Linux, Mac, etc we do not need to touch the overall process settings. * threadpool: update calling thread prio and affinity only at start/resume This avoids extra syscalls for each graph_compute() * llama-bench: turn threadpool params into vectors, add output headers, etc * llama-bench: add support for cool off between tests --delay This helps for long running tests on platforms that are thermally limited (phones, laptops, etc). --delay (disabled by default) introduces the sleep for N seconds before starting each test. * threadpool: move process priority setting into the apps (bench and cli) This avoids changing the overall process priority on Windows for the apps that use ggml/llama.cpp directy. * threadpool: move all pause/resume logic into ggml * threadpool: futher api cleanup and prep for future refactoring All threadpool related functions and structs use ggml_threadpool prefix. * threadpool: minor indent fixes * threadpool: improve setprioty error message * Update examples/llama-bench/llama-bench.cpp Co-authored-by: slaren <slarengh@gmail.com> * threadpool: fix indent in set_threadpool call * use int32_t for n_thread type in public llama.cpp API * threadpool: use _new and _free instead of _create and _release * fix two more public APIs to use int32_t for n_threads * build: set _GNU_SOURCE for Adroid --------- Co-authored-by: Max Krasnyansky <quic_maxk@quicinc.com> Co-authored-by: fmz <quic_fzaghlou@quic.com> Co-authored-by: Max Krasnyansky <max.krasnyansky@gmail.com> Co-authored-by: slaren <slarengh@gmail.com>
2024-08-29 23:20:53 +00:00
if (params.draft_cpuparams.n_threads > 0) {
params.cpuparams.n_threads = params.draft_cpuparams.n_threads;
}
Threadpool: take 2 (#8672) * Introduce ggml_compute_threadpool - OpenMP functional: check - Vanilla ggml functional: Check - ggml w/threadpool functional: Check - OpenMP no regression: No glaring problems - Vanilla ggml no regression: No glaring problems - ggml w/threadpool no regression: No glaring problems * Minor fixes * fixed use after release bug * fixed a harmless race condition * Fix Android bulid issue * fix more race conditions * fix deadlock for cases where cgraph.n_nodes == 1 and fix --poll case * threadpool: use cpu_get_num_math to set the default number of threadpool threads This way we avoid using E-Cores and Hyperthreaded siblings. * bench: create fresh threadpool for each test For benchmarking it's better to start a fresh pool for each test with the exact number of threads needed for that test. Having larger pools is suboptimal (causes more load, etc). * atomics: always use stdatomics with clang and use relaxed memory order when polling in ggml_barrier This also removes sched_yield() calls from ggml_barrier() to match OpenMP behavior. * threadpool: make polling the default to match openmp behavior All command line args now allow for setting poll to 0 (false). * threadpool: do not wakeup threads in already paused threadpool * fix potential race condition in check_for_work * threadpool: do not create two threadpools if their params are identical * threadpool: reduce pause/resume/wakeup overhead in common cases We now start threadpool in paused state only if we have two. The resume is now implicit (ie new work) which allows for reduced locking and context-switch overhead. * threadpool: add support for hybrid polling poll params (--poll, ...) now specify "polling level", i.e. how aggresively we poll before waiting on cond.var. poll=0 means no polling, 1 means poll for 128K rounds then wait, 2 for 256K rounds, ... The default value of 50 (ie 50x128K rounds) seems like a decent default across modern platforms. We can tune this further as things evolve. * threadpool: reduce the number of barrier required New work is now indicated with an atomic counter that is incremented for each new graph that needs to be computed. This removes the need for extra barrier for clearing the "new_work" and removes the special case for trivial graphs. * threadpool: remove special-casing for disposable threadpools With the efficient hybrid polling there is no need to make disposable pools any different. This simplifies the overall logic and reduces branching. Include n_threads in debug print for disposable threadpool. Declare pause and stop flags as atomic_bool This doesn't actually generate any memory barriers and simply informs the thread sanitizer that these flags can be written & read by different threads without locking. * threadpool: do not clear barrier counters between graphs computes (fixes race with small graphs) This fixes the race condition with very small graphs where the main thread happens to start a new graph while the workers are just about to exit from barriers. * threadpool: use relaxed order for chunk sync Full memory barrier is an overkill for this since each thread works on different chunk * threadpool: remove abort_callback from threadpool state * threadpool: better naming for thread/cpumask releated functions * threadpool: consistent use of int type for n_threads params * threadpool: add support for ggml_threadpool_params_default/init Also removes the need for explicit mask_specified param. all-zero cpumask means use default (usually inherited) cpu affinity mask. * threadpool: move typedef into ggml.h * threadpool: fix apply_priority() function name * threadpool: fix swift wrapper errors due to n_threads int type cleanup * threadpool: enable --cpu-mask and other threadpool related options only if threadpool is enabled * threadpool: replace checks for compute_thread ret code with proper status check * threadpool: simplify threadpool init logic and fix main thread affinity application Most of the init code is now exactly the same between threadpool and openmp. * threadpool: update threadpool resume/pause function names * threadpool: enable openmp by default for now * threadpool: don't forget to free workers state when omp is enabled * threadpool: avoid updating process priority on the platforms that do not require it On Windows we need to change overall process priority class in order to set thread priorities, but on Linux, Mac, etc we do not need to touch the overall process settings. * threadpool: update calling thread prio and affinity only at start/resume This avoids extra syscalls for each graph_compute() * llama-bench: turn threadpool params into vectors, add output headers, etc * llama-bench: add support for cool off between tests --delay This helps for long running tests on platforms that are thermally limited (phones, laptops, etc). --delay (disabled by default) introduces the sleep for N seconds before starting each test. * threadpool: move process priority setting into the apps (bench and cli) This avoids changing the overall process priority on Windows for the apps that use ggml/llama.cpp directy. * threadpool: move all pause/resume logic into ggml * threadpool: futher api cleanup and prep for future refactoring All threadpool related functions and structs use ggml_threadpool prefix. * threadpool: minor indent fixes * threadpool: improve setprioty error message * Update examples/llama-bench/llama-bench.cpp Co-authored-by: slaren <slarengh@gmail.com> * threadpool: fix indent in set_threadpool call * use int32_t for n_thread type in public llama.cpp API * threadpool: use _new and _free instead of _create and _release * fix two more public APIs to use int32_t for n_threads * build: set _GNU_SOURCE for Adroid --------- Co-authored-by: Max Krasnyansky <quic_maxk@quicinc.com> Co-authored-by: fmz <quic_fzaghlou@quic.com> Co-authored-by: Max Krasnyansky <max.krasnyansky@gmail.com> Co-authored-by: slaren <slarengh@gmail.com>
2024-08-29 23:20:53 +00:00
params.cpuparams_batch.n_threads = params.draft_cpuparams_batch.n_threads;
llama_init_result llama_init_dft = llama_init_from_gpt_params(params);
model_dft = llama_init_dft.model;
ctx_dft = llama_init_dft.context;
const bool vocab_type_tgt = llama_vocab_type(model_tgt);
LOG_DBG("vocab_type tgt: %d\n", vocab_type_tgt);
const bool vocab_type_dft = llama_vocab_type(model_dft);
LOG_DBG("vocab_type dft: %d\n", vocab_type_dft);
if (vocab_type_tgt != vocab_type_dft) {
LOG_ERR("%s: draft model vocab type must match target model to use speculation but ", __func__);
LOG_ERR("vocab_type_dft = %d while vocab_type_tgt = %d\n", vocab_type_dft, vocab_type_tgt);
return 1;
}
if (
llama_add_bos_token(model_tgt) != llama_add_bos_token(model_dft) ||
llama_add_eos_token(model_tgt) != llama_add_eos_token(model_dft) ||
llama_token_bos(model_tgt) != llama_token_bos(model_dft) ||
llama_token_eos(model_tgt) != llama_token_eos(model_dft)
) {
LOG_ERR("%s: draft model special tokens must match target model to use speculation\n", __func__);
return 1;
}
{
const int n_vocab_tgt = llama_n_vocab(model_tgt);
const int n_vocab_dft = llama_n_vocab(model_dft);
const int vocab_diff = n_vocab_tgt > n_vocab_dft
? n_vocab_tgt - n_vocab_dft
: n_vocab_dft - n_vocab_tgt;
if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
LOG_ERR("%s: draft model vocab must closely match target model to use speculation but ", __func__);
LOG_ERR("target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
return 1;
}
for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
const char * token_text_tgt = llama_token_get_text(model_tgt, i);
const char * token_text_dft = llama_token_get_text(model_dft, i);
if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
LOG_ERR("%s: draft model vocab must match target model to use speculation but ", __func__);
LOG_ERR("token %d content differs - target '%s', draft '%s'\n", i,
llama_token_to_piece(ctx_tgt, i).c_str(),
llama_token_to_piece(ctx_dft, i).c_str());
return 1;
}
}
}
// Tokenize the prompt
std::vector<llama_token> inp;
inp = ::llama_tokenize(ctx_tgt, params.prompt, true, true);
const int max_context_size = llama_n_ctx(ctx_tgt);
const int max_tokens_list_size = max_context_size - 4;
if ((int) inp.size() > max_tokens_list_size) {
LOG_ERR("%s: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
return 1;
}
LOG("\n\n");
for (auto id : inp) {
LOG("%s", llama_token_to_piece(ctx_tgt, id).c_str());
}
const int n_input = inp.size();
const auto t_enc_start = ggml_time_us();
// eval the prompt with both models
llama_decode(ctx_tgt, llama_batch_get_one( inp.data(), n_input - 1, 0, 0));
llama_decode(ctx_tgt, llama_batch_get_one(&inp.back(), 1, n_input - 1, 0));
llama_decode(ctx_dft, llama_batch_get_one( inp.data(), n_input, 0, 0));
const auto t_enc_end = ggml_time_us();
// the 2 models should have the same vocab
//GGML_ASSERT(n_vocab == llama_n_vocab(model_dft));
// how many tokens to draft each time
int n_draft = params.n_draft;
int n_predict = 0;
int n_drafted = 0;
int n_accept = 0;
int n_past_tgt = inp.size();
int n_past_dft = inp.size();
// used to determine end of generation
bool has_eos = false;
// target model sampling context (reuse the llama_context's sampling instance)
struct gpt_sampler * smpl = gpt_sampler_init(model_tgt, params.sparams);
struct llama_sampler * softmax = llama_sampler_init_softmax();
// draft sequence data
std::vector<seq_draft> drafts(n_seq_dft);
for (int s = 0; s < n_seq_dft; ++s) {
// allocate gpt_sampler for each draft sequence
drafts[s].smpl = gpt_sampler_init(model_dft, params.sparams);
}
llama_batch batch_dft = llama_batch_init(params.n_ctx, 0, 1);
llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, n_seq_dft);
const auto t_dec_start = ggml_time_us();
// sample from the last token of the prompt
drafts[0].i_batch_tgt.resize(1);
drafts[0].i_batch_tgt[0] = 0;
while (true) {
std::set<int> active_seqs = {};
// print current draft sequences
for (int s = 0; s < n_seq_dft; ++s) {
if (!drafts[s].active) {
continue;
}
active_seqs.insert(s);
const auto & tokens = drafts[s].tokens;
LOG_DBG("draft %d: %s\n", s, string_from(ctx_dft, tokens).c_str());
}
int i_dft = 0;
int s_keep = 0;
llama_token token_id;
std::string token_str;
// loop until we fail to accept a drafted token or we run out of drafted tokens
while (true) {
// check if the target token matches any of the drafts
// for stochastic sampling, attempt to match the token with the drafted tokens
{
bool accept = false;
if (params.sparams.temp > 0) {
// stochastic verification
gpt_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft], true);
auto & dist_tgt = *gpt_sampler_get_candidates(smpl);
float p_tgt = 0.0f;
float p_dft = 0.0f;
while (active_seqs.size() > 0) {
// randomly select a sequence to verify from active sequences
std::uniform_int_distribution<unsigned int> u_int_dist(0, active_seqs.size() - 1);
int s = *std::next(active_seqs.begin(), u_int_dist(rng));
if (i_dft >= (int) drafts[s].tokens.size()) {
drafts[s].active = false;
active_seqs.erase(s);
continue;
}
if (accept) {
// if we already accepted a token, we can skip the rest
if (drafts[s].tokens[i_dft] != drafts[s_keep].tokens[i_dft]) {
drafts[s].active = false;
active_seqs.erase(s);
}
continue;
}
LOG_DBG("verifying sequence #%d at pos #%d from %d active sequence(s)\n", s, i_dft, (int) active_seqs.size());
float r = u_dist(rng);
llama_token_data_array dist_dft = { drafts[s].dists[i_dft].data() , drafts[s].dists[i_dft].size(), LLAMA_TOKEN_NULL, true };
//GGML_ASSERT(dist_tgt.size <= dist_dft.size);
// acquire the token probabilities assigned by the draft and target models
for (size_t i = 0; i < dist_tgt.size; i++) {
if (dist_tgt.data[i].id == drafts[s].tokens[i_dft]) {
p_tgt = dist_tgt.data[i].p;
}
if (dist_dft.data[i].id == drafts[s].tokens[i_dft]) {
p_dft = dist_dft.data[i].p;
}
if (p_tgt && p_dft) {
break;
}
}
LOG_DBG("r = %f, p_dft = %f, p_tgt = %f\n", r, p_dft, p_tgt);
if (r <= p_tgt / p_dft) {
s_keep = s;
accept = true;
token_id = drafts[s].tokens[i_dft];
token_str = llama_token_to_piece(ctx_tgt, token_id);
gpt_sampler_accept(smpl, token_id, true);
LOG_DBG("draft token %d of sequence %d (%d, '%s') accepted\n", i_dft, s, token_id, token_str.c_str());
break;
} else {
LOG_DBG("draft token %d of sequence %d (%d, '%s') rejected\n", i_dft, s, drafts[s].tokens[i_dft], llama_token_to_piece(ctx_tgt, drafts[s].tokens[i_dft]).c_str());
drafts[s].active = false;
// calculate residual probability
GGML_ASSERT(dist_tgt.sorted);
GGML_ASSERT(dist_dft.sorted);
// sort dist by id
std::sort(dist_tgt.data, dist_tgt.data + dist_tgt.size, [](const llama_token_data &a, const llama_token_data &b) {
return a.id < b.id;
});
std::sort(dist_dft.data, dist_dft.data + dist_dft.size, [](const llama_token_data &a, const llama_token_data &b) {
return a.id < b.id;
});
float sum_probs = 0.0f;
for (size_t i = 0; i < dist_tgt.size; i++) {
if (i < dist_dft.size) {
dist_tgt.data[i].p = std::max(0.0f, dist_tgt.data[i].p - dist_dft.data[i].p);
} else {
dist_tgt.data[i].p = std::max(0.0f, dist_tgt.data[i].p);
}
sum_probs += dist_tgt.data[i].p;
}
for (size_t i = 0; i < dist_tgt.size; i++) {
dist_tgt.data[i].p /= sum_probs;
}
// sort dist_tgt by p desc
std::sort(dist_tgt.data, dist_tgt.data + dist_tgt.size, [](const llama_token_data &a, const llama_token_data &b) {
return a.p > b.p;
});
}
active_seqs.erase(s);
for(int i = 0; i < n_seq_dft; i++) {
if (i == s) {
continue;
}
if (drafts[i].tokens[i_dft] == drafts[s].tokens[i_dft]) {
// synchronize active status for sequences with the same drafted token
drafts[i].active = drafts[i].active && accept;
if (!drafts[i].active) {
active_seqs.erase(s);
}
}
}
}
if (!accept) {
// all drafted tokens were rejected
// sample from the target model
LOG_DBG("all drafted tokens were rejected, sampling from residual distribution\n");
std::vector<float> probs(dist_tgt.size);
for (size_t i = 0; i < dist_tgt.size; ++i) {
probs[i] = dist_tgt.data[i].p;
}
std::discrete_distribution<> dist(probs.begin(), probs.end());
const int idx = dist(rng);
token_id = dist_tgt.data[idx].id;
gpt_sampler_accept(smpl, token_id, true);
token_str = llama_token_to_piece(ctx_tgt, token_id);
}
} else {
// greedy verification
// sample from the target model
LOG_DBG("sampling target: s_keep = %3d, i_dft = %3d, i_batch_tgt = %3d\n", s_keep, i_dft, drafts[s_keep].i_batch_tgt[i_dft]);
token_id = gpt_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft]);
gpt_sampler_accept(smpl, token_id, true);
token_str = llama_token_to_piece(ctx_tgt, token_id);
for (int s = 0; s < n_seq_dft; ++s) {
if (!drafts[s].active) {
continue;
}
if (i_dft < (int) drafts[s].tokens.size() && token_id == drafts[s].tokens[i_dft]) {
LOG_DBG("the sampled target token matches the %dth drafted token of sequence %d (%d, '%s') - accepted\n", i_dft, s, token_id, token_str.c_str());
s_keep = s;
accept = true;
} else {
drafts[s].active = false;
}
}
}
if (llama_token_is_eog(model_tgt, token_id)) {
has_eos = true;
}
++n_predict;
if (accept) {
++n_accept;
++n_past_tgt;
++n_past_dft;
++i_dft;
if (params.use_color) {
// Color token according to its origin sequence
LOG("\u001b[%dm%s\u001b[37m", (36 - s_keep % 6), token_str.c_str());
} else {
LOG("%s", token_str.c_str());
}
continue;
} else {
LOG("%s", token_str.c_str());
break;
}
}
}
{
LOG_DBG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", token_id, token_str.c_str());
// TODO: simplify
{
LOG_DBG("keeping sequence %d, n_past_tgt = %d, n_past_dft = %d\n", s_keep, n_past_tgt, n_past_dft);
llama_kv_cache_seq_keep(ctx_dft, s_keep);
llama_kv_cache_seq_cp (ctx_dft, s_keep, 0, -1, -1);
llama_kv_cache_seq_keep(ctx_dft, 0);
llama_kv_cache_seq_rm (ctx_tgt, s_keep, n_past_tgt, -1);
llama_kv_cache_seq_keep(ctx_tgt, s_keep);
llama_kv_cache_seq_cp (ctx_tgt, s_keep, 0, -1, -1);
llama_kv_cache_seq_keep(ctx_tgt, 0);
}
for (int s = 0; s < n_seq_dft; ++s) {
drafts[s].active = false;
drafts[s].tokens.clear();
drafts[s].i_batch_tgt.clear();
drafts[s].dists.clear();
}
// note: will be erased after the speculation phase
drafts[0].tokens.push_back(token_id);
drafts[0].dists.push_back(std::vector<llama_token_data>());
drafts[0].i_batch_tgt.push_back(0);
llama_batch_clear(batch_dft);
llama_batch_add (batch_dft, token_id, n_past_dft, { 0 }, true);
llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
// LOG_DBG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
llama_decode(ctx_dft, batch_dft);
++n_past_dft;
}
if (n_predict > params.n_predict || has_eos) {
break;
}
if (drafts[0].smpl) {
gpt_sampler_free(drafts[0].smpl);
}
drafts[0].smpl = gpt_sampler_clone(smpl);
int n_seq_cur = 1;
int n_past_cur = n_past_dft;
for (int s = 0; s < n_seq_dft; ++s) {
drafts[s].active = false;
drafts[s].drafting = false;
}
drafts[0].active = true;
drafts[0].drafting = true;
drafts[0].i_batch_dft = 0;
llama_batch_clear(batch_tgt);
llama_batch_add (batch_tgt, drafts[0].tokens[0], n_past_tgt, { 0 }, true);
// sample n_draft tokens from the draft model using tree-based sampling
for (int i = 0; i < n_draft; ++i) {
batch_dft.n_tokens = 0;
for (int s = 0; s < n_seq_dft; ++s) {
drafts[s].skip = false;
}
for (int s = 0; s < n_seq_dft; ++s) {
if (!drafts[s].drafting || drafts[s].skip) {
continue;
}
gpt_sampler_sample(drafts[s].smpl, ctx_dft, drafts[s].i_batch_dft, true);
const auto * cur_p = gpt_sampler_get_candidates(drafts[s].smpl);
for (int k = 0; k < std::min(n_seq_dft + 3, (int) cur_p->size); ++k) {
LOG_DBG(" - draft candidate %3d for seq %3d, pos %3d: %6d (%8.3f) '%s'\n",
k, s, i, cur_p->data[k].id, cur_p->data[k].p, llama_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
}
std::vector<int> sa(1, s);
// attempt to split the branch if the probability is high enough
for (int f = 1; f < 8; ++f) {
if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_split) {
LOG_DBG("splitting seq %3d into %3d\n", s, n_seq_cur);
llama_kv_cache_seq_rm(ctx_dft, n_seq_cur, -1, -1);
llama_kv_cache_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
// all previous tokens from this branch are now also part of the new branch
for (int t = 0; t < batch_tgt.n_tokens; ++t) {
for (int p = 0; p < batch_tgt.n_seq_id[t]; ++p) {
if (batch_tgt.seq_id[t][p] == s) {
batch_tgt.seq_id[t][batch_tgt.n_seq_id[t]] = n_seq_cur;
batch_tgt.n_seq_id[t]++;
break;
}
}
}
// copy the draft state
drafts[n_seq_cur].active = true;
drafts[n_seq_cur].drafting = true;
drafts[n_seq_cur].skip = true;
drafts[n_seq_cur].tokens = drafts[s].tokens;
drafts[n_seq_cur].dists = drafts[s].dists;
drafts[n_seq_cur].i_batch_dft = drafts[s].i_batch_dft;
drafts[n_seq_cur].i_batch_tgt = drafts[s].i_batch_tgt;
if (drafts[n_seq_cur].smpl) {
gpt_sampler_free(drafts[n_seq_cur].smpl);
}
drafts[n_seq_cur].smpl = gpt_sampler_clone(drafts[s].smpl);
sa.push_back(n_seq_cur);
n_seq_cur++;
} else {
break;
}
}
// add drafted token for each sequence
for (int is = 0; is < (int) sa.size(); ++is) {
const llama_token id = cur_p->data[is].id;
const int s = sa[is];
gpt_sampler_accept(drafts[s].smpl, id, true);
drafts[s].tokens.push_back(id);
// save cur_p.data into drafts[s].dists
drafts[s].dists.push_back({cur_p->data, cur_p->data + cur_p->size});
// add unique drafted tokens to the target batch
drafts[s].i_batch_tgt.push_back(batch_tgt.n_tokens);
llama_batch_add(batch_tgt, id, n_past_tgt + i + 1, { s }, true);
// add the token to the batch for batched decoding with the draft model
drafts[s].i_batch_dft = batch_dft.n_tokens;
llama_batch_add(batch_dft, id, n_past_cur, { s }, true);
2023-10-18 15:49:40 +00:00
if (batch_tgt.n_tokens > n_draft) {
drafts[s].drafting = false;
}
}
}
// no sequence is drafting anymore
if (batch_dft.n_tokens == 0) {
break;
}
// evaluate the drafted tokens on the draft model
llama_decode(ctx_dft, batch_dft);
++n_past_cur;
++n_drafted;
if (batch_tgt.n_tokens > n_draft) {
break;
}
}
// evaluate the target model on the drafted tokens
{
llama_kv_cache_seq_keep(ctx_tgt, 0);
for (int s = 1; s < n_seq_dft; ++s) {
llama_kv_cache_seq_cp(ctx_tgt, 0, s, -1, -1);
}
// LOG_DBG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str());
llama_decode(ctx_tgt, batch_tgt);
++n_past_tgt;
}
// the first token is always proposed by the target model before the speculation loop so we erase it here
for (int s = 0; s < n_seq_dft; ++s) {
if (!drafts[s].active) {
continue;
}
drafts[s].tokens.erase(drafts[s].tokens.begin());
drafts[s].dists.erase(drafts[s].dists.begin());
}
}
auto t_dec_end = ggml_time_us();
LOG("\n\n");
LOG_INF("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
LOG_INF("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
LOG_INF("\n");
LOG_INF("n_draft = %d\n", n_draft);
LOG_INF("n_predict = %d\n", n_predict);
LOG_INF("n_drafted = %d\n", n_drafted);
LOG_INF("n_accept = %d\n", n_accept);
LOG_INF("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
LOG_INF("\n");
LOG_INF("draft:\n\n");
// TODO: print sampling/grammar timings for all drafts
llama_perf_context_print(ctx_dft);
LOG_INF("\n");
LOG_INF("target:\n\n");
gpt_perf_print(ctx_tgt, smpl);
gpt_sampler_free(smpl);
for (int s = 0; s < n_seq_dft; ++s) {
gpt_sampler_free(drafts[s].smpl);
}
llama_sampler_free(softmax);
llama_batch_free(batch_dft);
llama_free(ctx_tgt);
llama_free_model(model_tgt);
llama_free(ctx_dft);
llama_free_model(model_dft);
llama_backend_free();
LOG("\n\n");
return 0;
}