tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-29 04:35:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

perplexity : avoid common_batch

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2025-03-20 12:21:40 +02:00
parent 8b80d68338
commit 76fd7d6f5b
1 changed files with 41 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files

85
examples/perplexity/perplexity.cpp
View File

@ -363,15 +363,16 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
// clear the KV cache
llama_kv_self_clear(ctx);
common_batch batch(n_batch, 1);
llama_batch_ext_ptr batch(llama_batch_ext_init(n_batch, 1));
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);
batch.clear();
llama_batch_ext_clear(batch.get());
for (int i = 0; i < batch_size; i++) {
batch.add_text(tokens[batch_start + i], j*n_batch + i, 0, true);
llama_seq_id seq_id = 0;
llama_batch_ext_add_text(batch.get(), tokens[batch_start + i], j*n_batch + i, &seq_id, 1, true);
}
//LOG_DBG(" Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
@ -501,7 +502,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
GGML_ASSERT(n_batch < n_ctx || n_batch % n_ctx == 0);
GGML_ASSERT(params.n_ctx == n_seq * n_ctx);
common_batch batch(std::min(n_batch, n_ctx*n_seq), 1);
llama_batch_ext_ptr batch(llama_batch_ext_init(std::min(n_batch, n_ctx*n_seq), 1));
std::vector<float> logits;
if (num_batches > 1) {
@ -552,7 +553,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
int n_outputs = 0;
batch.clear();
llama_batch_ext_clear(batch.get());
for (int seq = 0; seq < n_seq_batch; seq++) {
int seq_start = batch_start + seq*n_ctx;
@ -567,7 +568,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
for (int k = 0; k < batch_size; ++k) {
const llama_pos pos = j*n_batch + k;
bool output = pos >= first;
batch.add_text(tokens[seq_start + k], pos, seq, output);
llama_batch_ext_add_text(batch.get(), tokens[seq_start + k], pos, &seq, 1, output);
n_outputs += output ? 1 : 0;
}
@ -649,25 +650,14 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
return {tokens, ppl, logit_history, prob_history};
}
static bool decode_helper(llama_context * ctx, common_batch & batch, std::vector<float> & batch_logits, int n_batch, int n_vocab) {
int prev_outputs = 0;
for (int i = 0; i < (int) batch.get_n_tokens(); i += n_batch) {
const int n_tokens = std::min<int>(n_batch, batch.get_n_tokens() - i);
common_batch batch_view = batch.get_view(i, n_tokens);
const int ret = llama_decode_ext(ctx, batch_view.get());
static bool decode_helper(llama_context * ctx, llama_batch_ext_ptr & batch, std::vector<float> & batch_logits, size_t n_outputs, int n_vocab) {
const int ret = llama_decode_ext(ctx, batch.get());
if (ret != 0) {
LOG_ERR("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
LOG_ERR("failed to decode the batch, ret = %d\n", ret);
return false;
}
int n_outputs = batch_view.n_outputs;
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;
}
memcpy(batch_logits.data(), llama_get_logits(ctx), n_outputs*n_vocab*sizeof(float));
return true;
}
@ -836,14 +826,12 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
double acc = 0.0f;
const int n_ctx = llama_n_ctx(ctx);
const int n_batch = params.n_batch;
const int n_vocab = llama_vocab_n_tokens(vocab);
const int max_tasks_per_batch = 32;
const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
common_batch batch(n_ctx, 4);
llama_batch_ext_ptr batch(llama_batch_ext_init(n_ctx, 4));
std::vector<float> tok_logits(n_vocab);
// TODO: this could be made smaller; it's currently the worst-case size
@ -859,7 +847,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
size_t i1 = i0;
size_t i_logits = 0; // this tells us how many logits were needed before this point in the batch
batch.clear();
llama_batch_ext_clear(batch.get());
// batch as much tasks as possible into the available context
// each task has 4 unique sequence ids - one for each ending
@ -875,7 +863,8 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
}
for (size_t i = 0; i < hs_cur.common_prefix; ++i) {
batch.add_text_multi_seq(hs_cur.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3 }, false);
std::vector<llama_seq_id> seq_ids = { s0 + 0, s0 + 1, s0 + 2, s0 + 3 };
llama_batch_ext_add_text(batch.get(), hs_cur.seq_tokens[0][i], i, seq_ids.data(), seq_ids.size(), false);
}
llama_batch_ext_set_output_last(batch.get());
n_logits += 1;
@ -885,7 +874,8 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
// TODO: don't evaluate the last token of each sequence
for (size_t i = hs_cur.common_prefix; i < seq_tokens_size; ++i) {
const bool needs_logits = i < seq_tokens_size - 1;
batch.add_text_multi_seq(hs_cur.seq_tokens[s][i], i, { s0 + s }, needs_logits);
llama_seq_id seq_id = s0 + s;
llama_batch_ext_add_text(batch.get(), hs_cur.seq_tokens[s][i], i, &seq_id, 1, needs_logits);
n_logits += needs_logits;
}
}
@ -907,7 +897,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
llama_kv_self_clear(ctx);
// decode all tasks [i0, i1)
if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
if (!decode_helper(ctx, batch, batch_logits, i_logits, n_vocab)) {
LOG_ERR("%s: llama_decode() failed\n", __func__);
return;
}
@ -1118,14 +1108,12 @@ static void winogrande_score(llama_context * ctx, const common_params & params)
LOG_INF("%s : calculating winogrande score over selected tasks.\n", __func__);
const int n_ctx = llama_n_ctx(ctx);
const int n_batch = params.n_batch;
const int n_vocab = llama_vocab_n_tokens(vocab);
const int max_tasks_per_batch = 128;
const int max_seq = std::min(2*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
common_batch batch(n_ctx, 2);
llama_batch_ext_ptr batch(llama_batch_ext_init(n_ctx, 2));
std::vector<float> tok_logits(n_vocab);
// TODO: this could be made smaller; it's currently the worst-case size
@ -1144,7 +1132,7 @@ static void winogrande_score(llama_context * ctx, const common_params & params)
size_t i1 = i0;
size_t i_logits = 0;
batch.clear();
llama_batch_ext_clear(batch.get());
while (n_cur + (int) data[i1].required_tokens <= n_ctx) {
int n_logits = 0;
@ -1154,7 +1142,8 @@ static void winogrande_score(llama_context * ctx, const common_params & params)
}
for (size_t i = 0; i < data[i1].common_prefix; ++i) {
batch.add_text_multi_seq(data[i1].seq_tokens[0][i], i, { s0 + 0, s0 + 1 }, false);
std::vector<llama_seq_id> seq_ids{ s0 + 0, s0 + 1 };
llama_batch_ext_add_text(batch.get(), data[i1].seq_tokens[0][i], i, seq_ids.data(), seq_ids.size(), false);
}
llama_batch_ext_set_output_last(batch.get());
n_logits += 1;
@ -1162,7 +1151,8 @@ static void winogrande_score(llama_context * ctx, const common_params & params)
for (int s = 0; s < 2; ++s) {
// TODO: end before the last token, no need to predict past the end of the sequences
for (size_t i = data[i1].common_prefix; i < data[i1].seq_tokens[s].size(); ++i) {
batch.add_text_multi_seq(data[i1].seq_tokens[s][i], i, { s0 + s }, true);
llama_seq_id seq_id = s0 + s;
llama_batch_ext_add_text(batch.get(), data[i1].seq_tokens[s][i], i, &seq_id, 1, true);
n_logits += 1;
}
}
@ -1184,7 +1174,7 @@ static void winogrande_score(llama_context * ctx, const common_params & params)
llama_kv_self_clear(ctx);
// decode all tasks [i0, i1)
if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
if (!decode_helper(ctx, batch, batch_logits, i_logits, n_vocab)) {
LOG_ERR("%s: llama_decode() failed\n", __func__);
return;
}
@ -1472,14 +1462,12 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
LOG("\ntask\tacc_norm\n");
const int n_ctx = llama_n_ctx(ctx);
const int n_batch = params.n_batch;
const int n_vocab = llama_vocab_n_tokens(vocab);
const int max_tasks_per_batch = 32;
const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
common_batch batch(n_ctx, max_seq);
llama_batch_ext_ptr batch(llama_batch_ext_init(n_ctx, max_seq));
std::vector<float> tok_logits(n_vocab);
std::vector<float> batch_logits(size_t(n_ctx)*n_vocab);
@ -1499,7 +1487,7 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
size_t i1 = i0;
size_t i_logits = 0; // this tells us how many logits were needed before this point in the batch
batch.clear();
llama_batch_ext_clear(batch.get());
// batch as much tasks as possible into the available context
// each task has 4 unique sequence ids - one for each ending
@ -1518,11 +1506,12 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
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 (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);
batch.add_text_multi_seq(cur_task.seq_tokens[0][i], i, batch_indeces, false);
llama_batch_ext_add_text(batch.get(), cur_task.seq_tokens[0][i], i, batch_indeces.data(), batch_indeces.size(), false);
}
llama_batch_ext_set_output_last(batch.get()); // we need logits for the last token of the common prefix
n_logits += 1;
@ -1532,7 +1521,8 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
// TODO: don't evaluate the last token of each sequence
for (size_t i = cur_task.common_prefix; i < seq_tokens_size; ++i) {
const bool needs_logits = i < seq_tokens_size - 1;
batch.add_text_multi_seq(cur_task.seq_tokens[s][i], i, { s0 + s }, needs_logits);
llama_seq_id seq_id = { s0 + s };
llama_batch_ext_add_text(batch.get(), cur_task.seq_tokens[s][i], i, &seq_id, 1, needs_logits);
n_logits += needs_logits;
}
}
@ -1556,7 +1546,7 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
llama_kv_self_clear(ctx);
// decode all tasks [i0, i1)
if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
if (!decode_helper(ctx, batch, batch_logits, i_logits, n_vocab)) {
LOG_ERR("%s: llama_decode() failed\n", __func__);
return;
}
@ -1743,7 +1733,7 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
// clear the KV cache
llama_kv_self_clear(ctx);
common_batch batch(n_batch, 1);
llama_batch_ext_ptr batch(llama_batch_ext_init(n_batch, 1));
for (int j = 0; j < num_batches; ++j) {
const int batch_start = start + j * n_batch;
@ -1757,9 +1747,10 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
tokens[batch_start] = llama_vocab_bos(vocab);
}
batch.clear();
llama_batch_ext_clear(batch.get());
for (int i = 0; i < batch_size; i++) {
batch.add_text_multi_seq(tokens[batch_start + i], j*n_batch + i, {0}, true);
llama_seq_id seq_id = 0;
llama_batch_ext_add_text(batch.get(), tokens[batch_start + i], j*n_batch + i, &seq_id, 1, true);
}
if (llama_decode_ext(ctx, batch.get())) {