Skip to content

Commit 1c48616

Browse files
ngxsonngxson
committed
adapt all examples
1 parent b226c5b commit 1c48616

File tree

20 files changed

+92
-37
lines changed

20 files changed

+92
-37
lines changed

‎common/common.cpp‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -912,7 +912,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
912912
}
913913

914914
if (llama_model_has_encoder(model)) {
915-
llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size(), 0, 0));
915+
llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size()));
916916
llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
917917
if (decoder_start_token_id == -1) {
918918
decoder_start_token_id = bos;
@@ -921,7 +921,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
921921
tmp.push_back(decoder_start_token_id);
922922
}
923923
if (llama_model_has_decoder(model)) {
924-
llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
924+
llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
925925
}
926926
llama_kv_cache_clear(lctx);
927927
llama_synchronize(lctx);

‎examples/batched-bench/batched-bench.cpp‎

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -74,7 +74,6 @@ int main(int argc, char ** argv) {
7474
batch.n_seq_id + i,
7575
batch.seq_id + i,
7676
batch.logits + i,
77-
0, 0, 0, // unused
7877
};
7978

8079
const int ret = llama_decode(ctx, batch_view);

‎examples/cvector-generator/cvector-generator.cpp‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -339,7 +339,7 @@ static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
339339

340340
static bool get_hidden_layers(llama_context * ctx, std::vector<llama_token> & tokens) {
341341
llama_kv_cache_clear(ctx);
342-
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
342+
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
343343
fprintf(stderr, "%s : failed to eval\n", __func__);
344344
return false;
345345
}

‎examples/eval-callback/eval-callback.cpp‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -131,7 +131,7 @@ static bool run(llama_context * ctx, const gpt_params & params) {
131131

132132
std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
133133

134-
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
134+
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
135135
LOG_ERR("%s : failed to eval\n", __func__);
136136
return false;
137137
}

‎examples/imatrix/imatrix.cpp‎

Lines changed: 11 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -508,12 +508,21 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
508508
tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
509509
}
510510

511-
// TODO: use batch.logits to save computations instead of relying on logits_all == true
512-
if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
511+
llama_batch batch = llama_batch_init(batch_size, 0, 1);
512+
for (int i = 0; i < batch_size; i++) {
513+
batch. token[i] = tokens[batch_start + i];
514+
batch. pos[i] = j*n_batch + i;
515+
batch.logits[i] = true;
516+
batch.seq_id[i][0] = 0;
517+
}
518+
519+
if (llama_decode(ctx, batch)) {
513520
LOG_ERR("%s : failed to eval\n", __func__);
514521
return false;
515522
}
516523

524+
llama_batch_free(batch);
525+
517526
// restore the original token in case it was set to BOS
518527
tokens[batch_start] = token_org;
519528

‎examples/infill/infill.cpp‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -396,7 +396,7 @@ int main(int argc, char ** argv) {
396396

397397
LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
398398

399-
if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
399+
if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
400400
LOG_ERR("%s : failed to eval\n", __func__);
401401
return 1;
402402
}

‎examples/llama-bench/llama-bench.cpp‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1446,7 +1446,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat
14461446
for (int i = 1; i < n_tokens; i++) {
14471447
tokens[i] = std::rand() % n_vocab;
14481448
}
1449-
llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens, n_past + n_processed, 0));
1449+
llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens));
14501450
n_processed += n_tokens;
14511451
}
14521452

@@ -1462,7 +1462,7 @@ static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads)
14621462
llama_token token = llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
14631463

14641464
for (int i = 0; i < n_gen; i++) {
1465-
llama_decode(ctx, llama_batch_get_one(&token, 1, n_past + i, 0));
1465+
llama_decode(ctx, llama_batch_get_one(&token, 1));
14661466
llama_synchronize(ctx);
14671467
token = std::rand() % n_vocab;
14681468
}

‎examples/llama.android/llama/src/main/cpp/llama-android.cpp‎

Lines changed: 0 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -283,9 +283,6 @@ Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens,
283283
nullptr,
284284
nullptr,
285285
nullptr,
286-
0,
287-
0,
288-
0,
289286
};
290287

291288
if (embd) {

‎examples/llava/llava-cli.cpp‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -20,7 +20,7 @@ static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_toke
2020
if (n_eval > n_batch) {
2121
n_eval = n_batch;
2222
}
23-
if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
23+
if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval))) {
2424
LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
2525
return false;
2626
}

‎examples/llava/llava.cpp‎

Lines changed: 36 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -401,6 +401,39 @@ bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, co
401401
return true;
402402
}
403403

404+
struct llava_embd_batch {
405+
std::vector<llama_pos> pos;
406+
std::vector<int32_t> n_seq_id;
407+
std::vector<llama_seq_id> seq_id_0;
408+
std::vector<llama_seq_id *> seq_ids;
409+
std::vector<int8_t> logits;
410+
llama_batch batch;
411+
llava_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
412+
pos .resize(n_tokens);
413+
n_seq_id.resize(n_tokens);
414+
seq_ids .resize(n_tokens + 1);
415+
logits .resize(n_tokens);
416+
seq_id_0.resize(1);
417+
seq_id_0[0] = seq_id;
418+
seq_ids [n_tokens] = nullptr;
419+
batch = {
420+
/*n_tokens =*/ n_tokens,
421+
/*tokens =*/ nullptr,
422+
/*embd =*/ embd,
423+
/*pos =*/ pos.data(),
424+
/*n_seq_id =*/ n_seq_id.data(),
425+
/*seq_id =*/ seq_ids.data(),
426+
/*logits =*/ logits.data(),
427+
};
428+
for (int i = 0; i < n_tokens; i++) {
429+
batch.pos [i] = pos_0 + i;
430+
batch.n_seq_id[i] = 1;
431+
batch.seq_id [i] = seq_id_0.data();
432+
batch.logits [i] = false;
433+
}
434+
}
435+
};
436+
404437
bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past) {
405438
int n_embd = llama_n_embd(llama_get_model(ctx_llama));
406439

@@ -409,8 +442,9 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
409442
if (n_eval > n_batch) {
410443
n_eval = n_batch;
411444
}
412-
llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
413-
if (llama_decode(ctx_llama, batch)) {
445+
float * embd = image_embed->embed+i*n_embd;
446+
llava_embd_batch llava_batch = llava_embd_batch(embd, n_eval, *n_past, 0);
447+
if (llama_decode(ctx_llama, llava_batch.batch)) {
414448
LOG_ERR("%s : failed to eval\n", __func__);
415449
return false;
416450
}

0 commit comments

Comments
 (0)