@@ -1646,6 +1646,7 @@ struct llama_cparams {
16461646 float defrag_thold;
16471647
16481648 bool mul_mat_q;
1649+ bool embeddings;
16491650 bool offload_kqv;
16501651 bool do_pooling;
16511652
@@ -1936,16 +1937,16 @@ struct llama_context {
19361937 int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1)
19371938 int32_t n_eval = 0; // number of eval calls
19381939
1939- // decode output (2-dimensional array: [n_tokens][n_vocab])
1940+ // logits output (2-dimensional array: [n_tokens][n_vocab])
19401941 std::vector<float> logits;
19411942#ifndef NDEBUG
19421943 // guard against access to unset logits
19431944 std::vector<bool> logits_valid;
19441945#endif
19451946 bool logits_all = false;
19461947
1947- // input embedding (1 -dimensional array: [n_embd])
1948- std::vector<float> embedding ;
1948+ // embeddings output (2 -dimensional array: [n_tokens] [n_embd])
1949+ std::vector<float> embeddings ;
19491950
19501951 // memory buffers used to evaluate the model
19511952 std::vector<uint8_t> buf_compute_meta;
@@ -5987,9 +5988,10 @@ struct llm_build_context {
59875988
59885989 // get input vectors with right size
59895990 const size_t stride1 = n_tokens * ggml_type_size(lctx.inp_tokens->type);
5990- struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
5991+
5992+ struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
59915993 struct ggml_tensor * inp_mean = ggml_view_2d(ctx0, lctx.inp_mean, n_tokens, n_tokens, stride1, 0);
5992- struct ggml_tensor * inp_cls = ggml_view_1d(ctx0, lctx.inp_cls, n_tokens, 0);
5994+ struct ggml_tensor * inp_cls = ggml_view_1d(ctx0, lctx.inp_cls, n_tokens, 0);
59935995
59945996 // construct input embeddings (token, type, position)
59955997 inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
@@ -7971,17 +7973,17 @@ static int llama_decode_internal(
79717973 ggml_cgraph * gf = llama_build_graph(lctx, batch, false);
79727974
79737975 // the output is always the last tensor in the graph
7974- struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1];
7975- struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2];
7976+ struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1];
7977+ struct ggml_tensor * embd = gf->nodes[gf->n_nodes - 2];
79767978
79777979 if (strcmp(res->name, "result_output") == 0) {
79787980 // the embeddings could be the second to last tensor, or the third to last tensor
7979- if (strcmp(embeddings ->name, "result_norm") != 0) {
7980- embeddings = gf->nodes[gf->n_nodes - 3];
7981- GGML_ASSERT(strcmp(embeddings ->name, "result_norm") == 0);
7981+ if (strcmp(embd ->name, "result_norm") != 0) {
7982+ embd = gf->nodes[gf->n_nodes - 3];
7983+ GGML_ASSERT(strcmp(embd ->name, "result_norm") == 0);
79827984 }
79837985 } else if (strcmp(res->name, "result_embd") == 0) {
7984- embeddings = res;
7986+ embd = res;
79857987 res = nullptr;
79867988 } else {
79877989 GGML_ASSERT(false);
@@ -8051,46 +8053,53 @@ static int llama_decode_internal(
80518053 logits_out.clear();
80528054#endif
80538055
8054- ggml_backend_t res_backend = ggml_backend_sched_get_node_backend(lctx.sched, res);
8055- GGML_ASSERT(res_backend != nullptr);
8056+ ggml_backend_t backend_res = ggml_backend_sched_get_node_backend(lctx.sched, res);
8057+ GGML_ASSERT(backend_res != nullptr);
8058+
80568059 if (batch.logits) {
80578060 logits_out.resize(n_vocab * n_tokens);
80588061 for (uint32_t i = 0; i < n_tokens; i++) {
80598062 if (batch.logits[i] == 0) {
80608063 continue;
80618064 }
8062- ggml_backend_tensor_get_async(res_backend , res, logits_out.data() + (n_vocab*i), (n_vocab*i)*sizeof(float), n_vocab*sizeof(float));
8065+ ggml_backend_tensor_get_async(backend_res , res, logits_out.data() + (n_vocab*i), (n_vocab*i)*sizeof(float), n_vocab*sizeof(float));
80638066#ifndef NDEBUG
80648067 logits_valid[i] = true;
80658068#endif
80668069 }
80678070 } else if (lctx.logits_all) {
80688071 logits_out.resize(n_vocab * n_tokens);
8069- ggml_backend_tensor_get_async(res_backend , res, logits_out.data(), 0, n_vocab*n_tokens*sizeof(float));
8072+ ggml_backend_tensor_get_async(backend_res , res, logits_out.data(), 0, n_vocab*n_tokens*sizeof(float));
80708073#ifndef NDEBUG
80718074 std::fill(logits_valid.begin(), logits_valid.end(), true);
80728075#endif
80738076 } else {
80748077 logits_out.resize(n_vocab);
8075- ggml_backend_tensor_get_async(res_backend , res, logits_out.data(), (n_vocab*(n_tokens - 1))*sizeof(float), n_vocab*sizeof(float));
8078+ ggml_backend_tensor_get_async(backend_res , res, logits_out.data(), (n_vocab*(n_tokens - 1))*sizeof(float), n_vocab*sizeof(float));
80768079#ifndef NDEBUG
80778080 logits_valid[0] = true;
80788081#endif
80798082 }
8080- ggml_backend_synchronize(res_backend );
8083+ ggml_backend_synchronize(backend_res );
80818084 }
80828085
80838086 // extract embeddings
8084- if (!lctx.embedding.empty() ) {
8085- auto & embedding_out = lctx.embedding ;
8087+ if (cparams.embeddings && embd ) {
8088+ auto & embeddings_out = lctx.embeddings ;
80868089
8087- const int64_t embd_pos = res ? n_embd * (n_tokens-1) : 0 ;
8088- const int64_t embd_size = res ? n_embd : n_embd * n_tokens ;
8090+ ggml_backend_t backend_embd = ggml_backend_sched_get_node_backend(lctx.sched, embd) ;
8091+ GGML_ASSERT(backend_embd != nullptr) ;
80898092
8090- embedding_out.resize(embd_size);
8091- ggml_backend_t embeddings_backend = ggml_backend_sched_get_node_backend(lctx.sched, embeddings);
8092- ggml_backend_tensor_get_async(embeddings_backend, embeddings, embedding_out.data(), embd_pos*sizeof(float), embd_size*sizeof(float));
8093- ggml_backend_synchronize(embeddings_backend);
8093+ if (batch.logits) {
8094+ embeddings_out.resize(n_embd * n_tokens);
8095+ for (uint32_t i = 0; i < n_tokens; i++) {
8096+ if (batch.logits[i] == 0) {
8097+ continue;
8098+ }
8099+ ggml_backend_tensor_get_async(backend_embd, embd, embeddings_out.data() + (n_embd*i), (n_embd*i)*sizeof(float), n_embd*sizeof(float));
8100+ }
8101+ }
8102+ ggml_backend_synchronize(backend_embd);
80948103 }
80958104
80968105 // measure the performance only for the single-token evals
@@ -11634,7 +11643,7 @@ struct llama_context_params llama_context_default_params() {
1163411643 /*.type_v =*/ GGML_TYPE_F16,
1163511644 /*.mul_mat_q =*/ true,
1163611645 /*.logits_all =*/ false,
11637- /*.embedding =*/ false,
11646+ /*.embeddings =*/ false,
1163811647 /*.offload_kqv =*/ true,
1163911648 /*.do_pooling =*/ true,
1164011649 };
@@ -11785,6 +11794,7 @@ struct llama_context * llama_new_context_with_model(
1178511794 cparams.yarn_beta_slow = params.yarn_beta_slow;
1178611795 cparams.defrag_thold = params.defrag_thold;
1178711796 cparams.mul_mat_q = params.mul_mat_q;
11797+ cparams.embeddings = params.embeddings;
1178811798 cparams.offload_kqv = params.offload_kqv;
1178911799 cparams.do_pooling = params.do_pooling;
1179011800
@@ -11933,8 +11943,8 @@ struct llama_context * llama_new_context_with_model(
1193311943 // resized during inference, reserve maximum
1193411944 ctx->logits.reserve(hparams.n_vocab*cparams.n_batch);
1193511945
11936- if (params.embedding ) {
11937- ctx->embedding.resize (hparams.n_embd);
11946+ if (params.embeddings ) {
11947+ ctx->embeddings.reserve (hparams.n_embd*cparams.n_batch );
1193811948 }
1193911949
1194011950 // graph inputs
@@ -12369,7 +12379,7 @@ size_t llama_get_state_size(const struct llama_context * ctx) {
1236912379 // assume worst case for logits although only currently set ones are serialized
1237012380 const size_t s_logits = ctx->logits.capacity() * sizeof(float);
1237112381 const size_t s_embedding_size = sizeof(size_t);
12372- const size_t s_embedding = ctx->embedding.size () * sizeof(float);
12382+ const size_t s_embedding = ctx->embeddings.capacity () * sizeof(float);
1237312383 const size_t s_kv_size = sizeof(size_t);
1237412384 const size_t s_kv_ntok = sizeof(int);
1237512385 const size_t s_kv = ctx->kv_self.total_size();
@@ -12470,12 +12480,12 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
1247012480
1247112481 // copy embeddings
1247212482 {
12473- const size_t embedding_size = ctx->embedding .size();
12483+ const size_t embeddings_size = ctx->embeddings .size();
1247412484
12475- data_ctx->write(&embedding_size , sizeof(embedding_size ));
12485+ data_ctx->write(&embeddings_size , sizeof(embeddings_size ));
1247612486
12477- if (embedding_size ) {
12478- data_ctx->write(ctx->embedding .data(), embedding_size * sizeof(float));
12487+ if (embeddings_size ) {
12488+ data_ctx->write(ctx->embeddings .data(), embeddings_size * sizeof(float));
1247912489 }
1248012490 }
1248112491
@@ -12581,15 +12591,17 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
1258112591
1258212592 // set embeddings
1258312593 {
12584- size_t embedding_size;
12594+ size_t embeddings_size;
12595+
12596+ memcpy(&embeddings_size, inp, sizeof(embeddings_size)); inp += sizeof(embeddings_size);
1258512597
12586- memcpy(&embedding_size, inp, sizeof(embedding_size)); inp += sizeof(embedding_size );
12598+ GGML_ASSERT(ctx->embeddings.capacity() == embeddings_size );
1258712599
12588- GGML_ASSERT(ctx->embedding.capacity() == embedding_size);
12600+ if (embeddings_size) {
12601+ ctx->embeddings.resize(embeddings_size);
1258912602
12590- if (embedding_size) {
12591- memcpy(ctx->embedding.data(), inp, embedding_size * sizeof(float));
12592- inp += embedding_size * sizeof(float);
12603+ memcpy(ctx->embeddings.data(), inp, embeddings_size * sizeof(float));
12604+ inp += embeddings_size * sizeof(float);
1259312605 }
1259412606 }
1259512607
@@ -12829,11 +12841,11 @@ float * llama_get_logits_ith(struct llama_context * ctx, int32_t i) {
1282912841}
1283012842
1283112843float * llama_get_embeddings(struct llama_context * ctx) {
12832- return ctx->embedding .data();
12844+ return ctx->embeddings .data();
1283312845}
1283412846
1283512847float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i) {
12836- return ctx->embedding .data() + i*ctx->model.hparams.n_embd;
12848+ return ctx->embeddings .data() + i*ctx->model.hparams.n_embd;
1283712849}
1283812850
1283912851const char * llama_token_get_text(const struct llama_model * model, llama_token token) {
0 commit comments