mirror of
https://github.com/LostRuins/koboldcpp.git
synced 2025-09-11 01:24:36 +00:00
Merge branch 'master' into concedo_experimental
# Conflicts: # .devops/nix/nixpkgs-instances.nix # .devops/nix/package.nix # .devops/nix/scope.nix # .github/workflows/build.yml # .github/workflows/nix-ci.yml # CMakeLists.txt # flake.nix # ggml.c
This commit is contained in:
Concedo
commit
f96f29be7b
20 changed files with 1807 additions and 162 deletions
|
|
@ -113,6 +113,43 @@ static results_log_softmax log_softmax(int n_vocab, const float * logits, int to
|
|||
return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp};
|
||||
}
|
||||
|
||||
static inline int nearest_int(float fval) {
|
||||
//assert(fval <= 4194303.f);
|
||||
float val = fval + 12582912.f;
|
||||
int i; memcpy(&i, &val, sizeof(int));
|
||||
return (i & 0x007fffff) - 0x00400000;
|
||||
}
|
||||
|
||||
static double log_softmax(int n_vocab, const float * logits, uint16_t * log_prob, int tok) {
|
||||
float max_logit = logits[0];
|
||||
float min_logit = logits[0];
|
||||
for (int i = 1; i < n_vocab; ++i) {
|
||||
max_logit = std::max(max_logit, logits[i]);
|
||||
min_logit = std::min(min_logit, logits[i]);
|
||||
}
|
||||
min_logit = std::max(min_logit, max_logit - 16);
|
||||
double sum_exp = 0.0;
|
||||
for (int i = 0; i < n_vocab; ++i) {
|
||||
sum_exp += expf(logits[i] - max_logit);
|
||||
}
|
||||
const float log_sum_exp = log(sum_exp);
|
||||
const float min_log_prob = min_logit - max_logit - log_sum_exp;
|
||||
const float scale = (max_logit - min_logit)/65535.f;
|
||||
float * d = (float *)log_prob;
|
||||
d[0] = scale;
|
||||
d[1] = min_log_prob;
|
||||
log_prob += 4;
|
||||
if (scale) {
|
||||
const float inv_scale = 1/scale;
|
||||
for (int i = 0; i < n_vocab; ++i) {
|
||||
log_prob[i] = logits[i] > min_logit ? nearest_int(inv_scale*(logits[i] - min_logit)) : 0;
|
||||
}
|
||||
} else {
|
||||
std::memset(log_prob, 0, n_vocab*sizeof(uint16_t));
|
||||
}
|
||||
return max_logit + log_sum_exp - logits[tok];
|
||||
}
|
||||
|
||||
static void process_logits(
|
||||
int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread> & workers,
|
||||
double & nll, double & nll2, float * logit_history, float * prob_history
|
||||
|
|
@ -148,6 +185,114 @@ static void process_logits(
|
|||
}
|
||||
}
|
||||
|
||||
static void process_logits(std::ostream& out, int n_vocab, const float * logits, const int * tokens, int n_token,
|
||||
std::vector<std::thread> & workers, std::vector<uint16_t> & log_probs, double & nll, double & nll2) {
|
||||
std::mutex mutex;
|
||||
const int nv = 2*((n_vocab + 1)/2) + 4;
|
||||
int counter = 0;
|
||||
auto compute = [&mutex, &counter, &log_probs, &nll, &nll2, n_vocab, logits, tokens, n_token, nv] () {
|
||||
double local_nll = 0;
|
||||
double local_nll2 = 0;
|
||||
while (true) {
|
||||
std::unique_lock<std::mutex> lock(mutex);
|
||||
int i = counter++;
|
||||
if (i >= n_token) {
|
||||
nll += local_nll; nll2 += local_nll2;
|
||||
break;
|
||||
}
|
||||
lock.unlock();
|
||||
const double v = log_softmax(n_vocab, logits + i*n_vocab, log_probs.data() + i*nv, tokens[i+1]);
|
||||
local_nll += v;
|
||||
local_nll2 += v*v;
|
||||
}
|
||||
};
|
||||
for (auto & w : workers) {
|
||||
w = std::thread(compute);
|
||||
}
|
||||
compute();
|
||||
for (auto & w : workers) {
|
||||
w.join();
|
||||
}
|
||||
out.write((const char *)log_probs.data(), n_token*nv*sizeof(uint16_t));
|
||||
}
|
||||
|
||||
struct kl_divergence_result {
|
||||
double sum_nll = 0;
|
||||
double sum_nll2 = 0;
|
||||
double sum_kld = 0;
|
||||
double sum_kld2 = 0;
|
||||
double sum_nll_diff = 0;
|
||||
double sum_nll_diff2 = 0;
|
||||
size_t count = 0;
|
||||
};
|
||||
|
||||
static void log_softmax(int n_vocab, const float * logits, const uint16_t * base_log_prob, int tok, kl_divergence_result & kld) {
|
||||
float max_logit = logits[0];
|
||||
for (int i = 1; i < n_vocab; ++i) {
|
||||
max_logit = std::max(max_logit, logits[i]);
|
||||
}
|
||||
double sum_exp = 0.0;
|
||||
for (int i = 0; i < n_vocab; ++i) {
|
||||
sum_exp += expf(logits[i] - max_logit);
|
||||
}
|
||||
const float log_sum_exp = log(sum_exp);
|
||||
const float * d = (const float *)base_log_prob;
|
||||
const float scale = d[0];
|
||||
const float min_log_prob = d[1];
|
||||
base_log_prob += 4;
|
||||
float nll = max_logit + log_sum_exp - logits[tok];
|
||||
kld.sum_nll += nll;
|
||||
kld.sum_nll2 += nll*nll;
|
||||
nll += (scale*base_log_prob[tok] + min_log_prob);
|
||||
kld.sum_nll_diff += nll;
|
||||
kld.sum_nll_diff2 += nll*nll;
|
||||
max_logit += log_sum_exp;
|
||||
double sum = 0;
|
||||
for (int i = 0; i < n_vocab; ++i) {
|
||||
const float p_log_base = scale*base_log_prob[i] + min_log_prob;
|
||||
if (p_log_base > -16.f) {
|
||||
const float p_base = expf(p_log_base);
|
||||
sum += p_base * (p_log_base - logits[i] + max_logit);
|
||||
}
|
||||
}
|
||||
kld.sum_kld += sum;
|
||||
kld.sum_kld2 += sum*sum;
|
||||
++kld.count;
|
||||
}
|
||||
|
||||
static void process_logits(int n_vocab, const float * logits, const int * tokens, int n_token,
|
||||
std::vector<std::thread> & workers, const std::vector<uint16_t> & base_log_probs, kl_divergence_result & kld) {
|
||||
std::mutex mutex;
|
||||
const int nv = 2*((n_vocab + 1)/2) + 4;
|
||||
int counter = 0;
|
||||
auto compute = [&mutex, &counter, &base_log_probs, &kld, n_vocab, logits, tokens, n_token, nv] () {
|
||||
kl_divergence_result local_kld;
|
||||
while (true) {
|
||||
std::unique_lock<std::mutex> lock(mutex);
|
||||
int i = counter++;
|
||||
if (i >= n_token) {
|
||||
kld.sum_nll += local_kld.sum_nll;
|
||||
kld.sum_nll2 += local_kld.sum_nll2;
|
||||
kld.sum_kld += local_kld.sum_kld;
|
||||
kld.sum_kld2 += local_kld.sum_kld2;
|
||||
kld.sum_nll_diff += local_kld.sum_nll_diff;
|
||||
kld.sum_nll_diff2 += local_kld.sum_nll_diff2;
|
||||
kld.count += local_kld.count;
|
||||
break;
|
||||
}
|
||||
lock.unlock();
|
||||
log_softmax(n_vocab, logits + i*n_vocab, base_log_probs.data() + i*nv, tokens[i+1], local_kld);
|
||||
}
|
||||
};
|
||||
for (auto & w : workers) {
|
||||
w = std::thread(compute);
|
||||
}
|
||||
compute();
|
||||
for (auto & w : workers) {
|
||||
w.join();
|
||||
}
|
||||
}
|
||||
|
||||
static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) {
|
||||
// Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
|
||||
// Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
|
||||
|
|
@ -295,6 +440,18 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
|
|||
const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
|
||||
const int n_ctx = llama_n_ctx(ctx);
|
||||
|
||||
std::ofstream logits_stream;
|
||||
if (!params.logits_file.empty()) {
|
||||
logits_stream.open(params.logits_file.c_str());
|
||||
if (!logits_stream.is_open()) {
|
||||
fprintf(stderr, "%s: failed to open %s for writing\n", __func__, params.logits_file.c_str());
|
||||
return {};
|
||||
}
|
||||
fprintf(stderr, "%s: saving all logits to %s\n", __func__, params.logits_file.c_str());
|
||||
logits_stream.write("_logits_", 8);
|
||||
logits_stream.write((const char *)&n_ctx, sizeof(n_ctx));
|
||||
}
|
||||
|
||||
auto tim1 = std::chrono::high_resolution_clock::now();
|
||||
fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
|
||||
|
||||
|
|
@ -337,6 +494,15 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
|
|||
|
||||
std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
|
||||
|
||||
std::vector<uint16_t> log_probs;
|
||||
if (!params.logits_file.empty()) {
|
||||
logits_stream.write((const char *)&n_vocab, sizeof(n_vocab));
|
||||
logits_stream.write((const char *)&n_chunk, sizeof(n_chunk));
|
||||
logits_stream.write((const char *)tokens.data(), n_chunk*n_ctx*sizeof(tokens[0]));
|
||||
const int nv = 2*((n_vocab + 1)/2) + 4;
|
||||
log_probs.resize(n_ctx * nv);
|
||||
}
|
||||
|
||||
for (int i = 0; i < n_chunk; ++i) {
|
||||
const int start = i * n_ctx;
|
||||
const int end = start + n_ctx;
|
||||
|
|
@ -399,8 +565,13 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
|
|||
// process the entire prompt.
|
||||
const int first = n_ctx/2;
|
||||
const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
|
||||
process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
|
||||
workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
|
||||
if (!params.logits_file.empty()) {
|
||||
process_logits(logits_stream, n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
|
||||
workers, log_probs, nll, nll2);
|
||||
} else {
|
||||
process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
|
||||
workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
|
||||
}
|
||||
count += n_ctx - first - 1;
|
||||
|
||||
// perplexity is e^(average negative log-likelihood)
|
||||
|
|
@ -541,14 +712,14 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
|
|||
// This is needed as usual for LLaMA models
|
||||
const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
|
||||
|
||||
// The tasks should be randomized so the score stabilizes quickly.
|
||||
bool randomize_tasks = true;
|
||||
|
||||
// Number of tasks to use when computing the score
|
||||
if (params.hellaswag_tasks < hs_task_count) {
|
||||
hs_task_count = params.hellaswag_tasks;
|
||||
}
|
||||
|
||||
// The tasks should be randomized so the score stabilizes quickly.
|
||||
bool randomize_tasks = true;
|
||||
|
||||
// The random seed should not impact the final result if the computation is done over enough tasks, so kept hardcoded for now
|
||||
std::mt19937 rng(1);
|
||||
|
||||
|
|
@ -1032,6 +1203,531 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
|
|||
printf("Final Winogrande score(%d tasks): %.4lf +/- %.4lf\n", n_done, 100*p, sigma);
|
||||
}
|
||||
|
||||
static bool deserialize_string(std::istream& in, std::string& str) {
|
||||
uint32_t size;
|
||||
if (!in.read((char *)&size, sizeof(size)).fail()) {
|
||||
str.resize(size);
|
||||
if (!in.read((char *)str.data(), size).fail()) return true;
|
||||
}
|
||||
return false;
|
||||
}
|
||||
|
||||
struct multiple_choice_answers {
|
||||
std::vector<std::string> answers;
|
||||
std::vector<int> labels;
|
||||
bool deserialize(std::istream& in) {
|
||||
uint32_t n;
|
||||
in.read((char *)&n, sizeof(n));
|
||||
if (in.fail() || n > 100) return false; // 100 as max. number of answers should be good enough for any practical purpose
|
||||
answers.resize(n);
|
||||
labels.resize(n);
|
||||
for (auto& a : answers) {
|
||||
if (!deserialize_string(in, a)) return false;
|
||||
}
|
||||
in.read((char *)labels.data(), n*sizeof(int));
|
||||
return !in.fail();
|
||||
}
|
||||
};
|
||||
|
||||
struct multiple_choice_task {
|
||||
std::string question; // the question (or context that needs to be continued)
|
||||
multiple_choice_answers mc1; // possible answers (continuations) with a single correct answer
|
||||
multiple_choice_answers mc2; // possible answers (continuations) with multiple correct answers - not handled yet
|
||||
bool deserialize(std::istream& in) {
|
||||
if (!deserialize_string(in, question)) return false;
|
||||
return mc1.deserialize(in) && mc2.deserialize(in);
|
||||
}
|
||||
|
||||
// For evaluation
|
||||
size_t i_batch; // starting index in the llama_batch
|
||||
size_t common_prefix; // max number of initial tokens that are the same in all sentences
|
||||
size_t required_tokens; // needed number of tokens to evaluate all answers
|
||||
std::vector<std::vector<llama_token>> seq_tokens;
|
||||
std::vector<float> log_probs;
|
||||
};
|
||||
|
||||
static bool multiple_choice_prepare_one_task(llama_context * ctx, bool add_bos, multiple_choice_task& task, bool log_error) {
|
||||
if (task.question.empty() || task.mc1.answers.empty()) {
|
||||
if (log_error) {
|
||||
printf("%s: found bad task with empty question and/or answers\n", __func__);
|
||||
}
|
||||
return false;
|
||||
}
|
||||
task.seq_tokens.reserve(task.mc1.answers.size());
|
||||
for (auto& answer : task.mc1.answers) {
|
||||
if (answer.empty()) {
|
||||
if (log_error) {
|
||||
printf("%s: found empty answer\n", __func__);
|
||||
}
|
||||
return false;
|
||||
}
|
||||
task.seq_tokens.emplace_back(::llama_tokenize(ctx, task.question + " " + answer, add_bos));
|
||||
}
|
||||
auto min_len = task.seq_tokens.front().size();
|
||||
for (auto& seq : task.seq_tokens) {
|
||||
min_len = std::min(min_len, seq.size());
|
||||
}
|
||||
task.common_prefix = 0;
|
||||
for (size_t k = 0; k < min_len; ++k) {
|
||||
auto token = task.seq_tokens[0][k];
|
||||
bool all_same = true;
|
||||
for (size_t i = 1; i < task.seq_tokens.size(); ++i) {
|
||||
if (task.seq_tokens[i][k] != token) {
|
||||
all_same = false;
|
||||
break;
|
||||
}
|
||||
}
|
||||
if (!all_same) {
|
||||
break;
|
||||
}
|
||||
++task.common_prefix;
|
||||
}
|
||||
task.required_tokens = task.common_prefix;
|
||||
for (auto& seq : task.seq_tokens) {
|
||||
task.required_tokens += seq.size() - task.common_prefix;
|
||||
}
|
||||
return true;
|
||||
}
|
||||
|
||||
//
|
||||
// Calculates score for multiple choice tasks with single correct answer from prompt.
|
||||
// Commonly used LLM evaluation metrics of this type are
|
||||
// * ARC
|
||||
// * HellaSwag
|
||||
// * MMLU
|
||||
// * TruthfulQA
|
||||
//
|
||||
// Validation datasets for these 4 tests can be found at
|
||||
// https://huggingface.co/datasets/ikawrakow/validation-datasets-for-llama.cpp
|
||||
// The data for these datasets was extracted from
|
||||
// git@hf.co:datasets/allenai/ai2_arc
|
||||
// https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl
|
||||
// git@hf.co:datasets/Stevross/mmlu
|
||||
// https://huggingface.co/datasets/truthful_qa
|
||||
//
|
||||
static void multiple_choice_score(llama_context * ctx, const gpt_params & params) {
|
||||
|
||||
std::istringstream strstream(params.prompt);
|
||||
uint32_t n_task;
|
||||
strstream.read((char *)&n_task, sizeof(n_task));
|
||||
if (strstream.fail() || n_task == 0) {
|
||||
printf("%s: no tasks\n", __func__);
|
||||
return;
|
||||
}
|
||||
printf("%s: there are %u tasks in prompt\n", __func__, n_task);
|
||||
std::vector<uint32_t> task_pos(n_task);
|
||||
strstream.read((char *)task_pos.data(), task_pos.size()*sizeof(uint32_t));
|
||||
if (strstream.fail()) {
|
||||
printf("%s: failed to raad task positions from prompt\n", __func__);
|
||||
return;
|
||||
}
|
||||
|
||||
std::vector<multiple_choice_task> tasks;
|
||||
if (params.multiple_choice_tasks == 0 || params.multiple_choice_tasks >= (size_t)n_task) {
|
||||
// Use all tasks
|
||||
tasks.resize(n_task);
|
||||
printf("%s: reading tasks", __func__);
|
||||
int n_dot = n_task/100;
|
||||
int i = 0;
|
||||
for (auto& task : tasks) {
|
||||
++i;
|
||||
if (!task.deserialize(strstream)) {
|
||||
printf("%s: failed to read task %d of %u\n", __func__, i, n_task);
|
||||
return;
|
||||
}
|
||||
if (i%n_dot == 0) printf(".");
|
||||
}
|
||||
printf("done\n");
|
||||
}
|
||||
else {
|
||||
printf("%s: selecting %zu random tasks from %u tasks available\n", __func__, params.multiple_choice_tasks, n_task);
|
||||
std::mt19937 rng(1);
|
||||
std::vector<int> aux(n_task);
|
||||
for (uint32_t i = 0; i < n_task; ++i) aux[i] = i;
|
||||
float scale = 1.f/(1.f + (float)std::mt19937::max());
|
||||
tasks.resize(params.multiple_choice_tasks);
|
||||
for (auto& task : tasks) {
|
||||
int j = (int)(scale * rng() * aux.size());
|
||||
int idx = aux[j];
|
||||
aux[j] = aux.back();
|
||||
aux.pop_back();
|
||||
strstream.seekg(task_pos[idx], std::ios::beg);
|
||||
if (!task.deserialize(strstream)) {
|
||||
printf("%s: failed to read task %d at position %u\n", __func__, idx, task_pos[idx]);
|
||||
return;
|
||||
}
|
||||
}
|
||||
n_task = params.multiple_choice_tasks;
|
||||
}
|
||||
|
||||
// This is needed as usual for LLaMA models
|
||||
const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
|
||||
|
||||
printf("%s: preparing task data", __func__);
|
||||
fflush(stdout);
|
||||
if (n_task > 500) {
|
||||
printf("...");
|
||||
fflush(stdout);
|
||||
std::atomic<int> counter(0);
|
||||
std::atomic<int> n_bad(0);
|
||||
auto prepare = [&counter, &n_bad, &tasks, ctx, add_bos] () {
|
||||
int num_tasks = tasks.size();
|
||||
int n_bad_local = 0;
|
||||
while (true) {
|
||||
int first = counter.fetch_add(K_TOKEN_CHUNK);
|
||||
if (first >= num_tasks) {
|
||||
if (n_bad_local > 0) n_bad += n_bad_local;
|
||||
break;
|
||||
}
|
||||
int last = std::min(first + K_TOKEN_CHUNK, num_tasks);
|
||||
for (int i = first; i < last; ++i) {
|
||||
if (!multiple_choice_prepare_one_task(ctx, add_bos, tasks[i], false)) ++n_bad_local;
|
||||
}
|
||||
}
|
||||
};
|
||||
size_t max_thread = std::thread::hardware_concurrency();
|
||||
max_thread = std::min(max_thread, (tasks.size() + K_TOKEN_CHUNK - 1)/K_TOKEN_CHUNK);
|
||||
std::vector<std::thread> workers(max_thread-1);
|
||||
for (auto& w : workers) w = std::thread(prepare);
|
||||
prepare();
|
||||
for (auto& w : workers) w.join();
|
||||
printf("done\n");
|
||||
fflush(stdout);
|
||||
int nbad = n_bad;
|
||||
if (nbad > 0) {
|
||||
printf("%s: found %d malformed tasks\n", __func__, nbad);
|
||||
return;
|
||||
}
|
||||
} else {
|
||||
int n_dot = n_task/100;
|
||||
int i_task = 0;
|
||||
for (auto& task : tasks) {
|
||||
++i_task;
|
||||
if (!multiple_choice_prepare_one_task(ctx, add_bos, task, true)) {
|
||||
return;
|
||||
}
|
||||
if (i_task%n_dot == 0) {
|
||||
printf(".");
|
||||
fflush(stdout);
|
||||
}
|
||||
}
|
||||
printf("done\n");
|
||||
}
|
||||
|
||||
printf("%s : calculating TruthfulQA score over %zu tasks.\n", __func__, tasks.size());
|
||||
|
||||
printf("\ntask\tacc_norm\n");
|
||||
|
||||
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
|
||||
const int n_ctx = llama_n_ctx(ctx);
|
||||
const int n_batch = params.n_batch;
|
||||
|
||||
const int max_tasks_per_batch = 32;
|
||||
const int max_seq = 4*max_tasks_per_batch;
|
||||
|
||||
llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
|
||||
|
||||
std::vector<float> tok_logits(n_vocab);
|
||||
std::vector<float> batch_logits(n_vocab*n_ctx);
|
||||
|
||||
std::vector<std::pair<size_t, llama_token>> eval_pairs;
|
||||
std::vector<float> eval_results;
|
||||
std::vector<std::thread> workers(std::thread::hardware_concurrency());
|
||||
std::vector<int> batch_indeces;
|
||||
|
||||
int n_done = 0;
|
||||
int n_correct = 0;
|
||||
int n_tot_answers = 0;
|
||||
|
||||
for (size_t i0 = 0; i0 < tasks.size(); i0++) {
|
||||
int n_cur = 0;
|
||||
|
||||
size_t i1 = i0;
|
||||
size_t i_batch = 0; // this tells us where in `llama_batch` we are currently
|
||||
|
||||
llama_batch_clear(batch);
|
||||
|
||||
// batch as much tasks as possible into the available context
|
||||
// each task has 4 unique seuqnce ids - one for each ending
|
||||
// the common prefix is shared among the 4 sequences to save tokens
|
||||
// we extract logits only from the last common token and from all ending tokens of each sequence
|
||||
int s0 = 0;
|
||||
while (n_cur + (int) tasks[i1].required_tokens <= n_ctx) {
|
||||
auto& cur_task = tasks[i1];
|
||||
|
||||
int num_answers = cur_task.seq_tokens.size();
|
||||
if (s0 + num_answers > max_seq) {
|
||||
break;
|
||||
}
|
||||
|
||||
if (int(batch_indeces.size()) != num_answers) {
|
||||
batch_indeces.resize(num_answers);
|
||||
}
|
||||
for (int s = 0; s < num_answers; ++s) batch_indeces[s] = s0 + s;
|
||||
|
||||
for (size_t i = 0; i < cur_task.common_prefix; ++i) {
|
||||
//llama_batch_add(batch, cur_task.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3}, false);
|
||||
llama_batch_add(batch, cur_task.seq_tokens[0][i], i, batch_indeces, false);
|
||||
}
|
||||
batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix
|
||||
|
||||
for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) {
|
||||
for (size_t i = cur_task.common_prefix; i < cur_task.seq_tokens[s].size(); ++i) {
|
||||
llama_batch_add(batch, cur_task.seq_tokens[s][i], i, { s0 + s }, true);
|
||||
}
|
||||
}
|
||||
|
||||
s0 += num_answers;
|
||||
|
||||
cur_task.i_batch = i_batch;
|
||||
i_batch += cur_task.required_tokens;
|
||||
|
||||
n_cur += cur_task.required_tokens;
|
||||
if (++i1 == tasks.size()) {
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
if (i0 == i1) {
|
||||
fprintf(stderr, "%s : task %zu does not fit in the context window\n", __func__, i0);
|
||||
return;
|
||||
}
|
||||
|
||||
llama_kv_cache_clear(ctx);
|
||||
|
||||
// decode all tasks [i0, i1)
|
||||
if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
|
||||
fprintf(stderr, "%s: llama_decode() failed\n", __func__);
|
||||
return;
|
||||
}
|
||||
|
||||
// Compute log-probs in parallel
|
||||
// First we collect all tasks
|
||||
eval_pairs.clear();
|
||||
for (size_t i = i0; i < i1; ++i) {
|
||||
auto& cur_task = tasks[i];
|
||||
size_t li = cur_task.common_prefix;
|
||||
for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) {
|
||||
for (size_t j = cur_task.common_prefix; j < cur_task.seq_tokens[s].size() - 1; j++) {
|
||||
eval_pairs.push_back(std::make_pair(cur_task.i_batch + li++, cur_task.seq_tokens[s][j + 1]));
|
||||
}
|
||||
++li;
|
||||
}
|
||||
}
|
||||
// Then we do the actual calculation
|
||||
compute_logprobs(batch_logits.data(), n_vocab, workers, eval_pairs, eval_results);
|
||||
|
||||
size_t ir = 0;
|
||||
|
||||
// compute the logprobs for each ending of the decoded tasks
|
||||
for (size_t i = i0; i < i1; ++i) {
|
||||
auto & cur_task = tasks[i];
|
||||
//printf("==== Evaluating <%s> with correct answer ", cur_task.question.c_str());
|
||||
//for (int j = 0; j < int(cur_task.mc1.labels.size()); ++j) {
|
||||
// if (cur_task.mc1.labels[j] == 1) {
|
||||
// printf("%d", j+1);
|
||||
// }
|
||||
//}
|
||||
//printf("\n common_prefix: %zu\n", cur_task.common_prefix);
|
||||
|
||||
std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*(cur_task.i_batch + cur_task.common_prefix - 1), n_vocab*sizeof(float));
|
||||
|
||||
const auto first_probs = softmax(tok_logits);
|
||||
|
||||
cur_task.log_probs.resize(cur_task.seq_tokens.size());
|
||||
for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) {
|
||||
size_t count = 1;
|
||||
float log_prob = std::log(first_probs[cur_task.seq_tokens[s][cur_task.common_prefix]]);
|
||||
for (size_t j = cur_task.common_prefix; j < cur_task.seq_tokens[s].size() - 1; j++) {
|
||||
//printf(" %zu %g\n", ir, eval_results[ir]);
|
||||
++count;
|
||||
log_prob += eval_results[ir++];
|
||||
}
|
||||
cur_task.log_probs[s] = log_prob / count;
|
||||
//printf(" Final: %g\n", log_prob / count);
|
||||
//printf(" <%s> : %g\n", cur_task.mc1.answers[s].c_str(), log_prob/count);
|
||||
}
|
||||
|
||||
// Find the ending with maximum logprob
|
||||
size_t logprob_max_idx = 0;
|
||||
float logprob_max_val = cur_task.log_probs[0];
|
||||
for (size_t s = 1; s < cur_task.log_probs.size(); s++) {
|
||||
if (cur_task.log_probs[s] > logprob_max_val) {
|
||||
logprob_max_val = cur_task.log_probs[s];
|
||||
logprob_max_idx = s;
|
||||
}
|
||||
}
|
||||
|
||||
n_tot_answers += cur_task.log_probs.size();
|
||||
if (cur_task.mc1.labels[logprob_max_idx] == 1) {
|
||||
++n_correct;
|
||||
}
|
||||
++n_done;
|
||||
|
||||
// Print the accumulated accuracy mean x 100
|
||||
printf("%d\t%.8lf\n", n_done, 100.*n_correct/n_done);
|
||||
fflush(stdout);
|
||||
}
|
||||
|
||||
i0 = i1 - 1;
|
||||
}
|
||||
|
||||
llama_batch_free(batch);
|
||||
|
||||
if (n_done < 100) return;
|
||||
|
||||
float p = 1.f*n_correct/n_done;
|
||||
float sigma = sqrt(p*(1-p)/(n_done-1));
|
||||
printf("\n Final result: %.4f +/- %.4f\n", 100.f*p, 100.f*sigma);
|
||||
p = 1.f*n_done/n_tot_answers;
|
||||
sigma = sqrt(p*(1-p)/(n_done-1));
|
||||
printf("Random chance: %.4f +/- %.4f\n", 100.f*p, 100.f*sigma);
|
||||
|
||||
printf("\n");
|
||||
}
|
||||
|
||||
static void kl_divergence(llama_context * ctx, const gpt_params & params) {
|
||||
if (params.logits_file.empty()) {
|
||||
fprintf(stderr, "%s: you must provide a name of a file containing the log probabilities of the base model\n", __func__);
|
||||
return;
|
||||
}
|
||||
std::ifstream in(params.logits_file.c_str(), std::ios::binary);
|
||||
if (!in) {
|
||||
fprintf(stderr, "%s: failed to open %s\n", __func__, params.logits_file.c_str());
|
||||
return;
|
||||
}
|
||||
{
|
||||
char check[9]; check[8] = 0;
|
||||
in.read(check, 8);
|
||||
if (in.fail() || strncmp("_logits_", check, 8) != 0) {
|
||||
fprintf(stderr, "%s: %s does not look like a file containing log-probabilities\n", __func__, params.logits_file.c_str());
|
||||
return;
|
||||
}
|
||||
}
|
||||
|
||||
uint32_t n_ctx;
|
||||
in.read((char *)&n_ctx, sizeof(n_ctx));
|
||||
if (n_ctx > llama_n_ctx(ctx)) {
|
||||
fprintf(stderr, "%s: %s has been computed with %d, while the current context is %d. Increase it with -c and retry\n",
|
||||
__func__, params.logits_file.c_str(), n_ctx, params.n_ctx);
|
||||
}
|
||||
|
||||
int n_vocab, n_chunk;
|
||||
in.read((char *)&n_vocab, sizeof(n_vocab));
|
||||
in.read((char *)&n_chunk, sizeof(n_chunk));
|
||||
if (in.fail()) {
|
||||
fprintf(stderr, "%s: failed rwading n_vocab, n_chunk from %s\n", __func__, params.logits_file.c_str());
|
||||
return;
|
||||
}
|
||||
if (n_vocab != llama_n_vocab(llama_get_model(ctx))) {
|
||||
fprintf(stderr, "%s: inconsistent vocabulary (%d vs %d)\n", __func__, n_vocab, llama_n_vocab(llama_get_model(ctx)));
|
||||
}
|
||||
|
||||
std::vector<llama_token> tokens(n_ctx * n_chunk);
|
||||
if (in.read((char *)tokens.data(), tokens.size()*sizeof(tokens[0])).fail()) {
|
||||
fprintf(stderr, "%s: failed reading evaluation tokens from %s\n", __func__, params.logits_file.c_str());
|
||||
return;
|
||||
}
|
||||
|
||||
const int n_batch = params.n_batch;
|
||||
const int num_batches = (n_ctx + n_batch - 1)/n_batch;
|
||||
const int nv = 2*((n_vocab + 1)/2) + 4;
|
||||
const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
|
||||
|
||||
std::vector<uint16_t> log_probs_uint16(size_t(n_ctx - 1 - n_ctx/2) * nv);
|
||||
std::vector<float> logits;
|
||||
if (num_batches > 1) {
|
||||
logits.reserve(n_ctx * n_vocab);
|
||||
}
|
||||
|
||||
std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
|
||||
|
||||
auto mean_and_uncertainty = [] (double sum, double sum2, size_t count) {
|
||||
if (count < 1) {
|
||||
return std::make_pair(0., 0.);
|
||||
}
|
||||
double f = sum/count;
|
||||
double df = sum2/count - f*f;
|
||||
df = df > 0 && count > 10 ? sqrt(df/(count-1)) : 0.;
|
||||
return std::make_pair(f, df);
|
||||
};
|
||||
|
||||
kl_divergence_result kld;
|
||||
|
||||
for (int i = 0; i < n_chunk; ++i) {
|
||||
const int start = i * n_ctx;
|
||||
const int end = start + n_ctx;
|
||||
|
||||
const auto t_start = std::chrono::high_resolution_clock::now();
|
||||
|
||||
if (in.read((char *)log_probs_uint16.data(), log_probs_uint16.size()*sizeof(uint16_t)).fail()) {
|
||||
fprintf(stderr, "%s: failed reading log-probs for chunk %d\n", __func__, i);
|
||||
return;
|
||||
}
|
||||
|
||||
// clear the KV cache
|
||||
llama_kv_cache_clear(ctx);
|
||||
|
||||
for (int j = 0; j < num_batches; ++j) {
|
||||
const int batch_start = start + j * n_batch;
|
||||
const int batch_size = std::min(end - batch_start, n_batch);
|
||||
|
||||
// save original token and restore it after eval
|
||||
const auto token_org = tokens[batch_start];
|
||||
|
||||
// add BOS token for the first batch of each chunk
|
||||
if (add_bos && j == 0) {
|
||||
tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
|
||||
}
|
||||
|
||||
if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
|
||||
fprintf(stderr, "%s : failed to eval\n", __func__);
|
||||
return;
|
||||
}
|
||||
|
||||
// restore the original token in case it was set to BOS
|
||||
tokens[batch_start] = token_org;
|
||||
|
||||
if (num_batches > 1) {
|
||||
const auto * batch_logits = llama_get_logits(ctx);
|
||||
logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
|
||||
}
|
||||
}
|
||||
|
||||
const auto t_end = std::chrono::high_resolution_clock::now();
|
||||
|
||||
if (i == 0) {
|
||||
const float t_total = std::chrono::duration<float>(t_end - t_start).count();
|
||||
fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
|
||||
int total_seconds = (int)(t_total * n_chunk);
|
||||
if (total_seconds >= 60*60) {
|
||||
fprintf(stderr, "%d hours ", total_seconds / (60*60));
|
||||
total_seconds = total_seconds % (60*60);
|
||||
}
|
||||
fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
|
||||
|
||||
printf("\nchunk PPL ln(PPL(Q)/PPL(base)) KL-Divergence\n");
|
||||
}
|
||||
|
||||
const int first = n_ctx/2;
|
||||
const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
|
||||
process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
|
||||
workers, log_probs_uint16, kld);
|
||||
|
||||
auto ppl = mean_and_uncertainty(kld.sum_nll, kld.sum_nll2, kld.count);
|
||||
auto log_ppl_ratio = mean_and_uncertainty(kld.sum_nll_diff, kld.sum_nll_diff2, kld.count);
|
||||
auto kl_div = mean_and_uncertainty(kld.sum_kld, kld.sum_kld2, kld.count);
|
||||
|
||||
printf("%4d %10.4lf %10.5lf ± %10.5f %10.5f ± %10.5lf\n", i+1, exp(ppl.first),
|
||||
log_ppl_ratio.first, log_ppl_ratio.second, kl_div.first, kl_div.second);
|
||||
|
||||
fflush(stdout);
|
||||
|
||||
logits.clear();
|
||||
}
|
||||
printf("\n");
|
||||
|
||||
}
|
||||
|
||||
int main(int argc, char ** argv) {
|
||||
gpt_params params;
|
||||
|
|
@ -1092,6 +1788,10 @@ int main(int argc, char ** argv) {
|
|||
hellaswag_score(ctx, params);
|
||||
} else if (params.winogrande) {
|
||||
winogrande_score(ctx, params);
|
||||
} else if (params.multiple_choice) {
|
||||
multiple_choice_score(ctx, params);
|
||||
} else if (params.kl_divergence) {
|
||||
kl_divergence(ctx, params);
|
||||
} else {
|
||||
results = perplexity(ctx, params);
|
||||
}
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue