Skip to content

perplexity : fix integer overflow #9783

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Oct 9, 2024
Merged

perplexity : fix integer overflow #9783

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
22cc760
perplexity : fix integer overflow
ggerganov Oct 8, 2024
fbefe17
perplexity : keep n_vocab as int and make appropriate casts
ggerganov Oct 8, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
85 changes: 49 additions & 36 deletions examples/perplexity/perplexity.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -169,7 +169,7 @@ static void process_logits(
break;
}
lock.unlock();
const results_log_softmax results = log_softmax(n_vocab, logits + i*n_vocab, tokens[i+1]);
const results_log_softmax results = log_softmax(n_vocab, logits + size_t(i)*n_vocab, tokens[i+1]);
const double v = -results.log_softmax;
local_nll += v;
local_nll2 += v*v;
Expand Down Expand Up @@ -203,7 +203,7 @@ static void process_logits(std::ostream& out, int n_vocab, const float * logits,
break;
}
lock.unlock();
const double v = log_softmax(n_vocab, logits + i*n_vocab, log_probs.data() + i*nv, tokens[i+1]);
const double v = log_softmax(n_vocab, logits + size_t(i)*n_vocab, log_probs.data() + i*nv, tokens[i+1]);
local_nll += v;
local_nll2 += v*v;
}
Expand Down Expand Up @@ -281,7 +281,9 @@ static std::pair<double, float> log_softmax(int n_vocab, const float * logits, c
kld.sum_kld += sum;
kld.sum_kld2 += sum*sum;
++kld.count;
if (imax == imax_base) ++kld.n_same_top;
if (imax == imax_base) {
++kld.n_same_top;
}

const float p_base = expf(-nll_base);
const float p = expf(-nll);
Expand Down Expand Up @@ -323,7 +325,7 @@ static void process_logits(int n_vocab, const float * logits, const int * tokens
break;
}
lock.unlock();
std::pair<double, float> v = log_softmax(n_vocab, logits + i*n_vocab, base_log_probs.data() + i*nv, tokens[i+1], local_kld);
std::pair<double, float> v = log_softmax(n_vocab, logits + size_t(i)*n_vocab, base_log_probs.data() + i*nv, tokens[i+1], local_kld);
kld_values[i] = (float)v.first;
p_diff_values[i] = v.second;
}
Expand Down Expand Up @@ -383,9 +385,10 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
const int n_chunk_max = (tokens.size() - calc_chunk + params.ppl_stride - 1) / params.ppl_stride;

const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
const int n_batch = params.n_batch;

const int n_vocab = llama_n_vocab(llama_get_model(ctx));

int count = 0;
double nll = 0.0;

Expand Down Expand Up @@ -424,8 +427,8 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
}

const auto batch_logits = llama_get_logits(ctx);
logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
const auto * batch_logits = llama_get_logits(ctx);
logits.insert(logits.end(), batch_logits, batch_logits + size_t(batch_size) * n_vocab);

if (j == 0) {
tokens[batch_start] = token_org;
Expand All @@ -447,11 +450,10 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &

//LOG_DBG("%s: using tokens %d...%d\n",__func__,params.n_ctx - params.ppl_stride + start, params.n_ctx + start);
for (int j = n_ctx - params.ppl_stride - 1; j < n_ctx - 1; ++j) {

// Calculate probability of next token, given the previous ones.
const std::vector<float> tok_logits(
logits.begin() + (j + 0) * n_vocab,
logits.begin() + (j + 1) * n_vocab);
logits.begin() + size_t(j + 0) * n_vocab,
logits.begin() + size_t(j + 1) * n_vocab);

const float prob = softmax(tok_logits)[tokens[start + j + 1]];
logit_history[start + j + 1] = tok_logits[tokens[start + j + 1]];
Expand Down Expand Up @@ -521,9 +523,10 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
const int n_chunk_max = tokens.size() / n_ctx;

const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
const int n_batch = params.n_batch;

const int n_vocab = llama_n_vocab(llama_get_model(ctx));

int count = 0;
double nll = 0.0;
double nll2 = 0.0;
Expand All @@ -538,7 +541,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par

std::vector<float> logits;
if (num_batches > 1) {
logits.reserve((size_t)n_ctx * n_vocab);
logits.reserve(size_t(n_ctx) * n_vocab);
}

LOG_INF("%s: calculating perplexity over %d chunks, n_ctx=%d, batch_size=%d, n_seq=%d\n", __func__, n_chunk, n_ctx, n_batch, n_seq);
Expand Down Expand Up @@ -620,7 +623,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par

if (num_batches > 1 && n_outputs > 0) {
const auto * batch_logits = llama_get_logits(ctx);
logits.insert(logits.end(), batch_logits, batch_logits + n_outputs * n_vocab);
logits.insert(logits.end(), batch_logits, batch_logits + size_t(n_outputs) * n_vocab);
}
}

Expand Down Expand Up @@ -661,7 +664,9 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
} else {
double av = nll/count;
double av2 = nll2/count - av*av;
if (av2 > 0) av2 = sqrt(av2/(count-1));
if (av2 > 0) {
av2 = sqrt(av2/(count-1));
}
LOG("%8d %.4lf %4lf %4lf\n", i*n_ctx, std::exp(nll / count), av, av2);
}
}
Expand All @@ -686,10 +691,10 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
return {tokens, ppl, logit_history, prob_history};
}

static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<float> & batch_logits, int32_t n_batch, int32_t n_vocab) {
static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<float> & batch_logits, int n_batch, int n_vocab) {
int prev_outputs = 0;
for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
for (int i = 0; i < (int) batch.n_tokens; i += n_batch) {
const int n_tokens = std::min<int>(n_batch, batch.n_tokens - i);
Comment on lines +696 to +697
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

IIRC the C standard only guarantees at least 16 bit for int.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

According to Wikpedia:

The standard integer size is platform-dependent. In C, it is denoted by int and required to be at least 16 bits. Windows and Unix systems have 32-bit ints on both 32-bit and 64-bit architectures.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

According to cppreference:

int — basic integer type. The keyword int may be omitted if any of the modifiers listed below are used. If no length modifiers are present, it's guaranteed to have a width of at least 16 bits. However, on 32/64 bit systems it is almost exclusively guaranteed to have width of at least 32 bits (see below).

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I doubt this is a concern. We can safely assume that int will always be 32-bit


llama_batch batch_view = {
n_tokens,
Expand All @@ -713,7 +718,7 @@ static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<
n_outputs += batch_view.logits[i] != 0;
}

memcpy(batch_logits.data() + prev_outputs*n_vocab, llama_get_logits(ctx), n_outputs*n_vocab*sizeof(float));
memcpy(batch_logits.data() + size_t(prev_outputs)*n_vocab, llama_get_logits(ctx), size_t(n_outputs)*n_vocab*sizeof(float));

prev_outputs += n_outputs;
}
Expand All @@ -728,19 +733,23 @@ static void compute_logprobs(const float * batch_logits, int n_vocab, std::vecto
if (eval_results.size() != eval_pairs.size()) {
eval_results.resize(eval_pairs.size());
}
if (eval_pairs.empty()) return;
if (eval_pairs.empty()) {
return;
}

size_t max_threads = std::min((eval_pairs.size() + K_TOKEN_CHUNK - 1)/K_TOKEN_CHUNK, workers.size());

std::atomic<int> counter(0);
auto compute = [&counter, &eval_pairs, &eval_results, batch_logits, n_vocab] () {
float local_logprobs[K_TOKEN_CHUNK];
while (true) {
size_t first = counter.fetch_add(K_TOKEN_CHUNK, std::memory_order_relaxed);
if (first >= eval_results.size()) break;
size_t last = std::min(first + K_TOKEN_CHUNK, eval_results.size());
const size_t first = counter.fetch_add(K_TOKEN_CHUNK, std::memory_order_relaxed);
if (first >= eval_results.size()) {
break;
}
const size_t last = std::min(first + K_TOKEN_CHUNK, eval_results.size());
for (size_t i = first; i < last; ++i) {
auto logits = batch_logits + eval_pairs[i].first * n_vocab;
const auto * logits = batch_logits + eval_pairs[i].first * n_vocab;
float max_logit = logits[0];
for (int j = 1; j < n_vocab; ++j) {
max_logit = std::max(max_logit, logits[j]);
Expand Down Expand Up @@ -877,18 +886,19 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {

double acc = 0.0f;

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 n_vocab = llama_n_vocab(llama_get_model(ctx));

const int max_tasks_per_batch = 32;
const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));

llama_batch batch = llama_batch_init(n_ctx, 0, 4);

std::vector<float> tok_logits(n_vocab);
// TODO: this could be made smaller; it's currently the worst-case size
std::vector<float> batch_logits(n_vocab*n_ctx);
std::vector<float> batch_logits(size_t(n_ctx)*n_vocab);

std::vector<std::pair<size_t, llama_token>> eval_pairs;
std::vector<float> eval_results;
Expand Down Expand Up @@ -975,7 +985,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
auto & hs_cur = hs_data[i];

// get the logits of the last token of the common prefix
std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*hs_cur.i_logits, n_vocab*sizeof(float));
std::memcpy(tok_logits.data(), batch_logits.data() + hs_cur.i_logits*n_vocab, n_vocab*sizeof(float));

const auto first_probs = softmax(tok_logits);

Expand Down Expand Up @@ -1158,18 +1168,19 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {

LOG_INF("%s : calculating winogrande score over selected tasks.\n", __func__);

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 n_vocab = llama_n_vocab(llama_get_model(ctx));

const int max_tasks_per_batch = 128;
const int max_seq = std::min(2*max_tasks_per_batch, (int) llama_n_seq_max(ctx));

llama_batch batch = llama_batch_init(n_ctx, 0, 2);

std::vector<float> tok_logits(n_vocab);
// TODO: this could be made smaller; it's currently the worst-case size
std::vector<float> batch_logits(n_vocab*n_ctx);
std::vector<float> batch_logits(size_t(n_ctx)*n_vocab);

std::vector<std::pair<size_t, llama_token>> eval_pairs;
std::vector<float> eval_results;
Expand Down Expand Up @@ -1509,17 +1520,18 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params

LOG("\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 n_vocab = llama_n_vocab(llama_get_model(ctx));

const int max_tasks_per_batch = 32;
const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));

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<float> batch_logits(size_t(n_ctx)*n_vocab);

std::vector<std::pair<size_t, llama_token>> eval_pairs;
std::vector<float> eval_results;
Expand Down Expand Up @@ -1627,7 +1639,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
//LOG("\n common_prefix: %zu\n", cur_task.common_prefix);

// get the logits of the last token of the common prefix
std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*cur_task.i_logits, n_vocab*sizeof(float));
std::memcpy(tok_logits.data(), batch_logits.data() + cur_task.i_logits*n_vocab, n_vocab*sizeof(float));

const auto first_probs = softmax(tok_logits);

Expand Down Expand Up @@ -1709,7 +1721,8 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
__func__, params.logits_file.c_str(), n_ctx, params.n_ctx);
}

int n_vocab, n_chunk;
int n_vocab;
int n_chunk;
in.read((char *)&n_vocab, sizeof(n_vocab));
in.read((char *)&n_chunk, sizeof(n_chunk));
if (in.fail()) {
Expand All @@ -1720,7 +1733,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
LOG_ERR("%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);
std::vector<llama_token> tokens(size_t(n_ctx) * n_chunk);
if (in.read((char *)tokens.data(), tokens.size()*sizeof(tokens[0])).fail()) {
LOG_ERR("%s: failed reading evaluation tokens from %s\n", __func__, params.logits_file.c_str());
return;
Expand All @@ -1737,7 +1750,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
std::vector<float> p_diff_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk);
std::vector<float> logits;
if (num_batches > 1) {
logits.reserve(n_ctx * n_vocab);
logits.reserve(size_t(n_ctx) * n_vocab);
}

std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
Expand Down Expand Up @@ -1801,7 +1814,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {

if (num_batches > 1) {
const auto * batch_logits = llama_get_logits(ctx);
logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
logits.insert(logits.end(), batch_logits, batch_logits + size_t(batch_size) * n_vocab);
}
}

Expand All @@ -1822,7 +1835,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {

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,
process_logits(n_vocab, all_logits + size_t(first)*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
workers, log_probs_uint16, kld, kld_ptr, p_diff_ptr);
p_diff_ptr += n_ctx - 1 - first;
kld_ptr += n_ctx - 1 - first;
Expand Down
Loading