server : add /v1/completions endpoint (#5914)

* add-`/v1/completions`-endpoint

* add legacy comment to `/completion` endpoint

.devops/nix/package.nix

@@ -1,5 +1,6 @@
 {
   lib,
+  glibc,
   config,
   stdenv,
   mkShell,
@@ -30,6 +31,11 @@
   useRocm ? config.rocmSupport,
   useVulkan ? false,
   llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake
+
+  # It's necessary to consistently use backendStdenv when building with CUDA support,
+  # otherwise we get libstdc++ errors downstream.
+  effectiveStdenv ? if useCuda then cudaPackages.backendStdenv else stdenv,
+  enableStatic ? effectiveStdenv.hostPlatform.isStatic
 }@inputs:
 
 let
@@ -41,10 +47,7 @@ let
     versionOlder
     ;
 
-  # It's necessary to consistently use backendStdenv when building with CUDA support,
-  # otherwise we get libstdc++ errors downstream.
   stdenv = throw "Use effectiveStdenv instead";
-  effectiveStdenv = if useCuda then cudaPackages.backendStdenv else inputs.stdenv;
 
   suffices =
     lib.optionals useBlas [ "BLAS" ]
@@ -167,6 +170,9 @@ effectiveStdenv.mkDerivation (
         # TODO: Replace with autoAddDriverRunpath
         # once https://github.com/NixOS/nixpkgs/pull/275241 has been merged
         cudaPackages.autoAddOpenGLRunpathHook
+      ]
+      ++ optionals (effectiveStdenv.hostPlatform.isGnu && enableStatic) [
+        glibc.static
       ];
 
     buildInputs =
@@ -181,7 +187,7 @@ effectiveStdenv.mkDerivation (
       [
         (cmakeBool "LLAMA_NATIVE" false)
         (cmakeBool "LLAMA_BUILD_SERVER" true)
-        (cmakeBool "BUILD_SHARED_LIBS" true)
+        (cmakeBool "BUILD_SHARED_LIBS" (!enableStatic))
         (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
         (cmakeBool "LLAMA_BLAS" useBlas)
         (cmakeBool "LLAMA_CLBLAST" useOpenCL)
@@ -190,6 +196,7 @@ effectiveStdenv.mkDerivation (
         (cmakeBool "LLAMA_METAL" useMetalKit)
         (cmakeBool "LLAMA_MPI" useMpi)
         (cmakeBool "LLAMA_VULKAN" useVulkan)
+        (cmakeBool "LLAMA_STATIC" enableStatic)
       ]
       ++ optionals useCuda [
         (

.github/workflows/server.yml

@@ -58,7 +58,8 @@ jobs:
             cmake \
             python3-pip \
             wget \
-            psmisc
+            psmisc \
+            language-pack-en
 
       - name: Build
         id: cmake_build

Makefile

@@ -724,10 +724,9 @@ save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(C
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/oai.hpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h examples/llava/llava.h examples/llava/llava.cpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
-	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h %.hpp $< examples/llava/clip.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) -o $@ $(LDFLAGS) $(LWINSOCK2)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
 gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)

README.md

@@ -10,6 +10,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Recent API changes
 
+- [2024 Mar 4] Embeddings API updated https://github.com/ggerganov/llama.cpp/pull/5796
 - [2024 Mar 3] `struct llama_context_params` https://github.com/ggerganov/llama.cpp/pull/5849
 
 ### Hot topics

ci/run.sh

@@ -45,7 +45,8 @@ fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then
     if [ -z ${ONEAPI_ROOT} ]; then
-        echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:\n source /opt/intel/oneapi/setvars.sh"
+        echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:"
+        echo "source /opt/intel/oneapi/setvars.sh"
         exit 1
     fi
 

common/CMakeLists.txt

@@ -19,7 +19,12 @@ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
         endif()
     endif()
 
-    set(GIT_INDEX "${GIT_DIR}/index")
+    if(EXISTS "${GIT_DIR}/index")
+        set(GIT_INDEX "${GIT_DIR}/index")
+    else()
+        message(WARNING "Git index not found in git repository.")
+        set(GIT_INDEX "")
+    endif()
 else()
     message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
     set(GIT_INDEX "")

common/common.cpp

@@ -513,12 +513,6 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_sequences = std::stoi(argv[i]);
-        } else if (arg == "--p-accept" || arg == "-pa") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.p_accept = std::stof(argv[i]);
         } else if (arg == "--p-split" || arg == "-ps") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -1044,7 +1038,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
     printf("  -np N, --parallel N   number of parallel sequences to decode (default: %d)\n", params.n_parallel);
     printf("  -ns N, --sequences N  number of sequences to decode (default: %d)\n", params.n_sequences);
-    printf("  -pa N, --p-accept N   speculative decoding accept probability (default: %.1f)\n", (double)params.p_accept);
     printf("  -ps N, --p-split N    speculative decoding split probability (default: %.1f)\n", (double)params.p_split);
     printf("  -cb, --cont-batching  enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
     printf("  --mmproj MMPROJ_FILE  path to a multimodal projector file for LLaVA. see examples/llava/README.md\n");
@@ -1299,7 +1292,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.n_threads_batch   = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
     cparams.seed              = params.seed;
     cparams.logits_all        = params.logits_all;
-    cparams.embedding         = params.embedding;
+    cparams.embeddings        = params.embedding;
     cparams.rope_scaling_type = params.rope_scaling_type;
     cparams.rope_freq_base    = params.rope_freq_base;
     cparams.rope_freq_scale   = params.rope_freq_scale;

common/common.h

@@ -53,11 +53,10 @@ struct gpt_params {
     int32_t n_ctx                 = 512;   // context size
     int32_t n_batch               = 512;   // batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                = 0;     // number of tokens to keep from initial prompt
-    int32_t n_draft               = 8;     // number of tokens to draft during speculative decoding
+    int32_t n_draft               = 5;     // number of tokens to draft during speculative decoding
     int32_t n_chunks              = -1;    // max number of chunks to process (-1 = unlimited)
     int32_t n_parallel            = 1;     // number of parallel sequences to decode
     int32_t n_sequences           = 1;     // number of sequences to decode
-    float   p_accept              = 0.5f;  // speculative decoding accept probability
     float   p_split               = 0.1f;  // speculative decoding split probability
     int32_t n_gpu_layers          = -1;    // number of layers to store in VRAM (-1 - use default)
     int32_t n_gpu_layers_draft    = -1;    // number of layers to store in VRAM for the draft model (-1 - use default)

common/sampling.cpp

@@ -295,6 +295,77 @@ static llama_token llama_sampling_sample_impl(
     return id;
 }
 
+static llama_token_data_array llama_sample_probability_distribution_impl(
+                  struct llama_sampling_context * ctx_sampling,
+                  struct llama_context * ctx_main,
+                  struct llama_context * ctx_cfg,
+                  const int idx) {
+    const llama_sampling_params & params = ctx_sampling->params;
+
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
+
+    const int32_t penalty_last_n  = params.penalty_last_n < 0 ? params.n_prev : params.penalty_last_n;
+    const float   penalty_repeat  = params.penalty_repeat;
+    const float   penalty_freq    = params.penalty_freq;
+    const float   penalty_present = params.penalty_present;
+    const bool    penalize_nl     = params.penalize_nl;
+
+    auto & prev = ctx_sampling->prev;
+    auto & cur  = ctx_sampling->cur;
+
+    // Get a pointer to the logits
+    float * logits = llama_get_logits_ith(ctx_main, idx);
+
+    // Declare original_logits at the beginning of the function scope
+    std::vector<float> original_logits;
+
+    // apply params.logit_bias map
+    for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
+        logits[it->first] += it->second;
+    }
+
+    if (ctx_cfg) {
+        float * logits_guidance = llama_get_logits_ith(ctx_cfg, idx);
+        llama_sample_apply_guidance(ctx_main, logits, logits_guidance, params.cfg_scale);
+    }
+
+    cur.clear();
+
+    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+        cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
+    }
+
+    llama_token_data_array cur_p = { cur.data(), cur.size(), false };
+
+    // apply penalties
+    const auto& penalty_tokens = params.use_penalty_prompt_tokens ? params.penalty_prompt_tokens : prev;
+    const int penalty_tokens_used_size = std::min((int)penalty_tokens.size(), penalty_last_n);
+    if (penalty_tokens_used_size) {
+        const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))];
+
+        llama_sample_repetition_penalties(ctx_main, &cur_p,
+                penalty_tokens.data() + penalty_tokens.size() - penalty_tokens_used_size,
+                penalty_tokens_used_size, penalty_repeat, penalty_freq, penalty_present);
+
+        if (!penalize_nl) {
+            for (size_t idx = 0; idx < cur_p.size; idx++) {
+                if (cur_p.data[idx].id == llama_token_nl(llama_get_model(ctx_main))) {
+                    cur_p.data[idx].logit = nl_logit;
+                    break;
+                }
+            }
+        }
+    }
+
+    // apply grammar checks
+    if (ctx_sampling->grammar != NULL) {
+        llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
+    }
+
+    llama_sample_softmax(ctx_main, &cur_p);
+    return cur_p;
+}
+
 llama_token llama_sampling_sample(
                   struct llama_sampling_context * ctx_sampling,
                   struct llama_context * ctx_main,
@@ -304,6 +375,14 @@ llama_token llama_sampling_sample(
     return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, false);
 }
 
+llama_token_data_array llama_sampling_probability_distribution(
+                  struct llama_sampling_context * ctx_sampling,
+                  struct llama_context * ctx_main,
+                  struct llama_context * ctx_cfg,
+                  const int idx) {
+    return llama_sample_probability_distribution_impl(ctx_sampling,ctx_main, ctx_cfg, idx);
+}
+
 void llama_sampling_accept(
         struct llama_sampling_context * ctx_sampling,
         struct llama_context * ctx_main,

common/sampling.h

@@ -131,6 +131,13 @@ llama_token llama_sampling_sample(
         struct llama_context * ctx_cfg,
         int idx = 0);
 
+// returns the probability that token of given id will be sampled
+llama_token_data_array llama_sampling_probability_distribution(
+        struct llama_sampling_context * ctx_sampling,
+        struct llama_context * ctx_main,
+        struct llama_context * ctx_cfg,
+        int idx = 0);
+
 void llama_sampling_accept(
         struct llama_sampling_context * ctx_sampling,
         struct llama_context * ctx_main,

convert-hf-to-gguf.py

@@ -36,8 +36,10 @@ class SentencePieceTokenTypes(IntEnum):
     UNUSED = 5
     BYTE = 6
 
+
 AnyModel = TypeVar("AnyModel", bound="type[Model]")
 
+
 class Model(ABC):
     _model_classes: dict[str, type[Model]] = {}
 
@@ -187,6 +189,7 @@ class Model(ABC):
     @classmethod
     def register(cls, *names: str) -> Callable[[AnyModel], AnyModel]:
         assert names
+
         def func(modelcls: type[Model]):
             for name in names:
                 cls._model_classes[name] = modelcls

convert.py

@@ -1377,7 +1377,6 @@ def main(args_in: list[str] | None = None) -> None:
         # We currently only support Q8_0 output on little endian systems.
         output_choices.append("q8_0")
     parser = argparse.ArgumentParser(description="Convert a LLaMA model to a GGML compatible file")
-    parser.add_argument("--awq-path",     type=Path,              help="Path to scale awq cache file", default=None)
     parser.add_argument("--dump",         action="store_true",    help="don't convert, just show what's in the model")
     parser.add_argument("--dump-single",  action="store_true",    help="don't convert, just show what's in a single model file")
     parser.add_argument("--vocab-only",   action="store_true",    help="extract only the vocab")
@@ -1393,18 +1392,6 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
 
     args = parser.parse_args(args_in)
-    if args.awq_path:
-        sys.path.insert(1, str(Path(__file__).parent / 'awq-py'))
-        from awq.apply_awq import add_scale_weights  # type: ignore[import-not-found]
-        tmp_model_path = args.model / "weighted_model"
-        if tmp_model_path.is_dir():
-            print(f"{tmp_model_path} exists as a weighted model.")
-        else:
-            tmp_model_path.mkdir(parents=True, exist_ok=True)
-            print("Saving new weighted model ...")
-            add_scale_weights(str(args.model), str(args.awq_path), str(tmp_model_path))
-            print(f"Saved weighted model at {tmp_model_path}.")
-        args.model = tmp_model_path
 
     if args.dump_single:
         model_plus = lazy_load_file(args.model)

examples/embedding/embedding.cpp

@@ -19,11 +19,11 @@ static std::vector<std::string> split_lines(const std::string & s) {
 
 static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, int seq_id) {
     for (size_t i = 0; i < tokens.size(); i++) {
-        llama_batch_add(batch, tokens[i], i, { seq_id }, false);
+        llama_batch_add(batch, tokens[i], i, { seq_id }, i == tokens.size() - 1);
     }
 }
 
-static void normalize(float * vec, float * out, int n) {
+static void normalize(const float * vec, float * out, int n) {
     float norm = 0;
     for (int i = 0; i < n; i++) {
         norm += vec[i] * vec[i];
@@ -45,10 +45,23 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
     }
 
     // normalize on copy
-    for (int k = 0; k < n_seq; k++) {
-        float * emb = llama_get_embeddings_ith(ctx, k);
-        float * out = output + k * n_embd;
-        normalize(emb, out, n_embd);
+    for (int i = 0; i < batch.n_tokens; i++) {
+        if (!batch.logits[i]) {
+            continue;
+        }
+
+        // try to get sequence embeddings - supported only when pooling_type is not NONE
+        const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
+        if (embd == NULL) {
+            embd = llama_get_embeddings_ith(ctx, i);
+            if (embd == NULL) {
+                fprintf(stderr, "%s: failed to get embeddings for token %d\n", __func__, i);
+                continue;
+            }
+        }
+
+        float * out = output + batch.seq_id[i][0] * n_embd;
+        normalize(embd, out, n_embd);
     }
 }
 
@@ -132,7 +145,7 @@ int main(int argc, char ** argv) {
 
     // initialize batch
     const int n_prompts = prompts.size();
-    struct llama_batch batch = llama_batch_init(n_batch, 0, n_prompts);
+    struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
     // allocate output
     const int n_embd = llama_n_embd(model);
@@ -145,6 +158,7 @@ int main(int argc, char ** argv) {
     for (int k = 0; k < n_prompts; k++) {
         // clamp to n_batch tokens
         auto & inp = inputs[k];
+
         const uint64_t n_toks = inp.size();
 
         // encode if at capacity

examples/server-embd.py

@@ -0,0 +1,34 @@
+import asyncio
+import requests
+import numpy as np
+
+n = 8
+
+result = []
+
+async def requests_post_async(*args, **kwargs):
+    return await asyncio.to_thread(requests.post, *args, **kwargs)
+
+async def main():
+    model_url = "http://127.0.0.1:6900"
+    responses: list[requests.Response] = await asyncio.gather(*[requests_post_async(
+        url= f"{model_url}/embedding",
+        json= {"content": str(0)*1024}
+    ) for i in range(n)])
+
+    for response in responses:
+        embedding = response.json()["embedding"]
+        print(embedding[-8:])
+        result.append(embedding)
+
+asyncio.run(main())
+
+# compute cosine similarity
+
+for i in range(n-1):
+    for j in range(i+1, n):
+        embedding1 = np.array(result[i])
+        embedding2 = np.array(result[j])
+        similarity = np.dot(embedding1, embedding2) / (np.linalg.norm(embedding1) * np.linalg.norm(embedding2))
+        print(f"Similarity between {i} and {j}: {similarity:.2f}")
+

examples/server/CMakeLists.txt

@@ -1,12 +1,12 @@
 set(TARGET server)
 option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
-add_executable(${TARGET} server.cpp oai.hpp utils.hpp json.hpp httplib.h)
+add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
 )
-target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT})
 if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()

examples/server/README.md

@@ -436,7 +436,7 @@ Notice that each `probs` is an array of length `n_probs`.
         "next_token": {
             "has_next_token": true,
             "n_remain": -1,
-            "num_tokens_predicted": 0,
+            "n_decoded": 0,
             "stopped_eos": false,
             "stopped_limit": false,
             "stopped_word": false,

examples/server/oai.hpp

@@ -1,225 +0,0 @@
-#pragma once
-
-#include <string>
-#include <vector>
-#include <set>
-#include <mutex>
-#include <condition_variable>
-#include <unordered_map>
-
-#include "json.hpp"
-#include "utils.hpp"
-
-#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
-
-using json = nlohmann::json;
-
-inline static json oaicompat_completion_params_parse(
-    const struct llama_model * model,
-    const json &body, /* openai api json semantics */
-    const std::string &chat_template)
-{
-    json llama_params;
-
-    llama_params["__oaicompat"] = true;
-
-    // Map OpenAI parameters to llama.cpp parameters
-    //
-    // For parameters that are defined by the OpenAI documentation (e.g.
-    // temperature), we explicitly specify OpenAI's intended default; we
-    // need to do that because sometimes OpenAI disagrees with llama.cpp
-    //
-    // https://platform.openai.com/docs/api-reference/chat/create
-    llama_sampling_params default_sparams;
-    llama_params["model"]             = json_value(body, "model", std::string("unknown"));
-    llama_params["prompt"]            = format_chat(model, chat_template, body["messages"]);
-    llama_params["cache_prompt"]      = json_value(body, "cache_prompt", false);
-    llama_params["temperature"]       = json_value(body, "temperature", 0.0);
-    llama_params["top_k"]             = json_value(body, "top_k", default_sparams.top_k);
-    llama_params["top_p"]             = json_value(body, "top_p", 1.0);
-    llama_params["n_predict"]         = json_value(body, "max_tokens", -1);
-    llama_params["logit_bias"]        = json_value(body, "logit_bias",json::object());
-    llama_params["frequency_penalty"] = json_value(body, "frequency_penalty", 0.0);
-    llama_params["presence_penalty"]  = json_value(body, "presence_penalty", 0.0);
-    llama_params["seed"]              = json_value(body, "seed", LLAMA_DEFAULT_SEED);
-    llama_params["stream"]            = json_value(body, "stream", false);
-    llama_params["mirostat"]          = json_value(body, "mirostat", default_sparams.mirostat);
-    llama_params["mirostat_tau"]      = json_value(body, "mirostat_tau", default_sparams.mirostat_tau);
-    llama_params["mirostat_eta"]      = json_value(body, "mirostat_eta", default_sparams.mirostat_eta);
-    llama_params["penalize_nl"]       = json_value(body, "penalize_nl", default_sparams.penalize_nl);
-    llama_params["typical_p"]         = json_value(body, "typical_p", default_sparams.typical_p);
-    llama_params["repeat_last_n"]     = json_value(body, "repeat_last_n", default_sparams.penalty_last_n);
-    llama_params["ignore_eos"]        = json_value(body, "ignore_eos", false);
-    llama_params["tfs_z"]             = json_value(body, "tfs_z", default_sparams.tfs_z);
-
-    if (body.count("grammar") != 0) {
-        llama_params["grammar"] = json_value(body, "grammar", json::object());
-    }
-
-    // Handle 'stop' field
-    if (body.contains("stop") && body["stop"].is_string()) {
-        llama_params["stop"] = json::array({body["stop"].get<std::string>()});
-    } else {
-        llama_params["stop"] = json_value(body, "stop", json::array());
-    }
-
-    // Ensure there is ChatML-specific end sequence among stop words
-    llama_params["stop"].push_back("<|im_end|>");
-
-    return llama_params;
-}
-
-inline static json format_final_response_oaicompat(const json &request, const task_result &response, bool streaming = false)
-{
-    json result = response.result_json;
-
-    bool stopped_word        = result.count("stopped_word") != 0;
-    bool stopped_eos         = json_value(result, "stopped_eos", false);
-    int num_tokens_predicted = json_value(result, "tokens_predicted", 0);
-    int num_prompt_tokens    = json_value(result, "tokens_evaluated", 0);
-    std::string content      = json_value(result, "content", std::string(""));
-
-    std::string finish_reason = "length";
-    if (stopped_word || stopped_eos) {
-        finish_reason = "stop";
-    }
-
-    json choices =
-        streaming ? json::array({json{{"finish_reason", finish_reason},
-                                        {"index", 0},
-                                        {"delta", json::object()}}})
-                  : json::array({json{{"finish_reason", finish_reason},
-                                        {"index", 0},
-                                        {"message", json{{"content", content},
-                                                         {"role", "assistant"}}}}});
-
-    std::time_t t = std::time(0);
-
-    json res =
-        json{{"choices", choices},
-            {"created", t},
-            {"model",
-                json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
-            {"object", streaming ? "chat.completion.chunk" : "chat.completion"},
-            {"usage",
-                json{{"completion_tokens", num_tokens_predicted},
-                     {"prompt_tokens",     num_prompt_tokens},
-                     {"total_tokens",      num_tokens_predicted + num_prompt_tokens}}},
-            {"id", gen_chatcmplid()}};
-
-    if (server_verbose) {
-        res["__verbose"] = result;
-    }
-
-    if (result.contains("completion_probabilities")) {
-        res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array());
-    }
-
-    return res;
-}
-
-// return value is vector as there is one case where we might need to generate two responses
-inline static std::vector<json> format_partial_response_oaicompat(const task_result &response) {
-    json result = response.result_json;
-
-    if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) {
-        return std::vector<json>({response.result_json});
-    }
-
-    bool first = json_value(result, "oaicompat_token_ctr", 0) == 0;
-    std::string modelname = json_value(result, "model", std::string(DEFAULT_OAICOMPAT_MODEL));
-
-    bool stopped_word   = json_value(result, "stopped_word", false);
-    bool stopped_eos    = json_value(result, "stopped_eos", false);
-    bool stopped_limit  = json_value(result, "stopped_limit", false);
-    std::string content = json_value(result, "content", std::string(""));
-
-    std::string finish_reason;
-    if (stopped_word || stopped_eos) {
-        finish_reason = "stop";
-    }
-    if (stopped_limit) {
-        finish_reason = "length";
-    }
-
-    std::time_t t = std::time(0);
-
-    json choices;
-
-    if (!finish_reason.empty()) {
-        choices = json::array({json{{"finish_reason", finish_reason},
-                                    {"index", 0},
-                                    {"delta", json::object()}}});
-    } else {
-        if (first) {
-            if (content.empty()) {
-                choices = json::array({json{{"finish_reason", nullptr},
-                                            {"index", 0},
-                                            {"delta", json{{"role", "assistant"}}}}});
-            } else {
-                // We have to send this as two updates to conform to openai behavior
-                json initial_ret = json{{"choices", json::array({json{
-                                        {"finish_reason", nullptr},
-                                        {"index", 0},
-                                        {"delta", json{
-                                            {"role", "assistant"}
-                                        }}}})},
-                            {"created", t},
-                            {"id", gen_chatcmplid()},
-                            {"model", modelname},
-                            {"object", "chat.completion.chunk"}};
-
-                json second_ret = json{
-                            {"choices", json::array({json{{"finish_reason", nullptr},
-                                                            {"index", 0},
-                                                            {"delta", json{
-                                                            {"content", content}}}
-                                                            }})},
-                            {"created", t},
-                            {"id", gen_chatcmplid()},
-                            {"model", modelname},
-                            {"object", "chat.completion.chunk"}};
-
-                return std::vector<json>({initial_ret, second_ret});
-            }
-        } else {
-            // Some idiosyncrasy in task processing logic makes several trailing calls
-            // with empty content, we ignore these at the calee site.
-            if (content.empty()) {
-                return std::vector<json>({json::object()});
-            }
-
-            choices = json::array({json{
-                {"finish_reason", nullptr},
-                {"index", 0},
-                {"delta",
-                json{
-                    {"content", content},
-                }},
-            }});
-        }
-    }
-
-    json ret = json{{"choices", choices},
-                    {"created", t},
-                    {"id", gen_chatcmplid()},
-                    {"model", modelname},
-                    {"object", "chat.completion.chunk"}};
-
-    return std::vector<json>({ret});
-}
-
-inline static json format_embeddings_response_oaicompat(const json &request, const json &embeddings)
-{
-    json res =
-        json{
-            {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
-            {"object", "list"},
-            {"usage",
-                json{{"prompt_tokens", 0},
-                     {"total_tokens", 0}}},
-            {"data", embeddings}
-        };
-    return res;
-}
-

examples/server/tests/features/embeddings.feature

@@ -0,0 +1,94 @@
+@llama.cpp
+@embeddings
+Feature: llama.cpp server
+
+  Background: Server startup
+    Given a server listening on localhost:8080
+    And   a model file bert-bge-small/ggml-model-f16.gguf from HF repo ggml-org/models
+    And   a model alias bert-bge-small
+    And   42 as server seed
+    And   2 slots
+    And   1024 as batch size
+    And   2048 KV cache size
+    And   embeddings extraction
+    Then  the server is starting
+    Then  the server is healthy
+
+  Scenario: Embedding
+    When embeddings are computed for:
+    """
+    What is the capital of Bulgaria ?
+    """
+    Then embeddings are generated
+
+  Scenario: OAI Embeddings compatibility
+    Given a model bert-bge-small
+    When an OAI compatible embeddings computation request for:
+    """
+    What is the capital of Spain ?
+    """
+    Then embeddings are generated
+
+  Scenario: OAI Embeddings compatibility with multiple inputs
+    Given a model bert-bge-small
+    Given a prompt:
+      """
+      In which country Paris is located ?
+      """
+    And a prompt:
+      """
+      Is Madrid the capital of Spain ?
+      """
+    When an OAI compatible embeddings computation request for multiple inputs
+    Then embeddings are generated
+
+  Scenario: Multi users embeddings
+    Given a prompt:
+      """
+      Write a very long story about AI.
+      """
+    And a prompt:
+      """
+      Write another very long music lyrics.
+      """
+    And a prompt:
+      """
+      Write a very long poem.
+      """
+    And a prompt:
+      """
+      Write a very long joke.
+      """
+    Given concurrent embedding requests
+    Then the server is busy
+    Then the server is idle
+    Then all embeddings are generated
+
+  Scenario: Multi users OAI compatibility embeddings
+    Given a prompt:
+      """
+      In which country Paris is located ?
+      """
+    And a prompt:
+      """
+      Is Madrid the capital of Spain ?
+      """
+    And a prompt:
+      """
+      What is the biggest US city ?
+      """
+    And a prompt:
+      """
+      What is the capital of Bulgaria ?
+      """
+    And   a model bert-bge-small
+    Given concurrent OAI embedding requests
+    Then the server is busy
+    Then the server is idle
+    Then all embeddings are generated
+
+  Scenario: All embeddings should be the same
+    Given 10 fixed prompts
+    And   a model bert-bge-small
+    Given concurrent OAI embedding requests
+    Then all embeddings are the same

examples/server/tests/features/parallel.feature

@@ -9,7 +9,6 @@ Feature: Parallel
     And   512 as batch size
     And   64 KV cache size
     And   2 slots
-    And   embeddings extraction
     And   continuous batching
     Then  the server is starting
     Then  the server is healthy
@@ -99,48 +98,3 @@ Feature: Parallel
     Then the server is busy
     Then the server is idle
     Then all prompts are predicted
-
-  Scenario: Multi users embeddings
-    Given a prompt:
-      """
-      Write a very long story about AI.
-      """
-    And a prompt:
-      """
-      Write another very long music lyrics.
-      """
-    And a prompt:
-      """
-      Write a very long poem.
-      """
-    And a prompt:
-      """
-      Write a very long joke.
-      """
-    Given concurrent embedding requests
-    Then the server is busy
-    Then the server is idle
-    Then all embeddings are generated
-
-  Scenario: Multi users OAI compatibility embeddings
-    Given a prompt:
-      """
-      In which country Paris is located ?
-      """
-    And a prompt:
-      """
-      Is Madrid the capital of Spain ?
-      """
-    And a prompt:
-      """
-      What is the biggest US city ?
-      """
-    And a prompt:
-      """
-      What is the capital of Bulgaria ?
-      """
-    And   a model tinyllama-2
-    Given concurrent OAI embedding requests
-    Then the server is busy
-    Then the server is idle
-    Then all embeddings are generated

examples/server/tests/features/server.feature

@@ -49,34 +49,6 @@ Feature: llama.cpp server
       | llama-2      | Book                        | What is the best book                | 8          | (Mom\|what)+           | 8           | disabled         |
       | codellama70b | You are a coding assistant. | Write the fibonacci function in c++. | 64         | (thanks\|happy\|bird)+ | 32          | enabled          |
 
-  Scenario: Embedding
-    When embeddings are computed for:
-    """
-    What is the capital of Bulgaria ?
-    """
-    Then embeddings are generated
-
-  Scenario: OAI Embeddings compatibility
-    Given a model tinyllama-2
-    When an OAI compatible embeddings computation request for:
-    """
-    What is the capital of Spain ?
-    """
-    Then embeddings are generated
-
-  Scenario: OAI Embeddings compatibility with multiple inputs
-    Given a model tinyllama-2
-    Given a prompt:
-      """
-      In which country Paris is located ?
-      """
-    And a prompt:
-      """
-      Is Madrid the capital of Spain ?
-      """
-    When an OAI compatible embeddings computation request for multiple inputs
-    Then embeddings are generated
-
   Scenario: Tokenize / Detokenize
     When tokenizing:
     """

examples/server/tests/features/steps/steps.py

@@ -10,6 +10,7 @@ from contextlib import closing
 from re import RegexFlag
 
 import aiohttp
+import numpy as np
 import openai
 from behave import step
 from behave.api.async_step import async_run_until_complete
@@ -24,6 +25,9 @@ def step_server_config(context, server_fqdn, server_port):
     if 'PORT' in os.environ:
         context.server_port = int(os.environ['PORT'])
         print(f"$PORT set, overriding server port with to {context.server_port}")
+    if 'FQDN' in os.environ:
+        context.server_fqdn = os.environ['FQDN']
+        print(f"$FQDN set, overriding server fqdn with to {context.server_fqdn}")
 
     context.base_url = f'http://{context.server_fqdn}:{context.server_port}'
 
@@ -34,6 +38,7 @@ def step_server_config(context, server_fqdn, server_port):
     context.n_ga_w = None
     context.n_gpu_layer = None
     context.n_predict = None
+    context.n_prompts = 0
     context.n_server_predict = None
     context.n_slots = None
     context.prompt_prefix = None
@@ -202,6 +207,7 @@ def step_n_tokens_predicted(context, predicted_n):
 @step(u'a user prompt {user_prompt}')
 def step_user_prompt(context, user_prompt):
     context.prompts.append(user_prompt)
+    context.n_prompts = len(context.prompts)
 
 
 @step(u'a system prompt {system_prompt}')
@@ -290,6 +296,12 @@ def step_prompt_passkey(context):
     context.prompt_passkey = context.text
 
 
+@step(u'{n_prompts:d} fixed prompts')
+def step_fixed_prompts(context, n_prompts):
+    context.prompts.extend([str(0)*(context.n_batch if context.n_batch is not None else 512) for i in range(n_prompts)])
+    context.n_prompts = n_prompts
+
+
 @step(u'a "{passkey}" passkey challenge prompt with the passkey inserted every {i_pos:d} junk')
 def step_prompt_passkey(context, passkey, i_pos):
     prompt = ""
@@ -301,6 +313,7 @@ def step_prompt_passkey(context, passkey, i_pos):
         passkey_highlight = "\x1b[33m" + passkey + "\x1b[0m"
         print(f"Passkey challenge:\n```{prompt.replace(passkey, passkey_highlight)}```\n")
     context.prompts.append(context.prompt_prefix + prompt + context.prompt_suffix)
+    context.n_prompts = len(context.prompts)
 
 
 @step(u'an OAI compatible chat completions request with {api_error} api error')
@@ -341,11 +354,13 @@ async def step_oai_chat_completions(context, api_error):
 @step(u'a prompt')
 def step_a_prompt(context):
     context.prompts.append(context.text)
+    context.n_prompts = len(context.prompts)
 
 
 @step(u'a prompt {prompt}')
 def step_a_prompt_prompt(context, prompt):
     context.prompts.append(prompt)
+    context.n_prompts = len(context.prompts)
 
 
 @step(u'concurrent completion requests')
@@ -430,25 +445,47 @@ async def all_prompts_are_predicted(context, expected_predicted_n=None):
 @step(u'embeddings are computed for')
 @async_run_until_complete
 async def step_compute_embedding(context):
+    context.n_prompts = 1
     context.embeddings = await request_embedding(context.text, base_url=context.base_url)
 
 
+@step(u'all embeddings are the same')
+@async_run_until_complete
+async def step_all_embeddings_are_the_same(context):
+    n_embedding_requests = await gather_tasks_results(context)
+    assert n_embedding_requests > 0
+    embeddings = []
+    for i in range(n_embedding_requests):
+        embedding = context.tasks_result.pop().pop()
+        embeddings.append(embedding)
+        assert_embeddings(embedding)
+    n = len(embeddings)
+    for i in range(n-1):
+        for j in range(i+1, n):
+            embedding1 = np.array(embeddings[i])
+            embedding2 = np.array(embeddings[j])
+            if context.debug:
+                print(f"embedding1: {embedding1[-8:]}\n")
+                print(f"embedding2: {embedding2[-8:]}\n")
+            similarity = np.dot(embedding1, embedding2) / (np.linalg.norm(embedding1) * np.linalg.norm(embedding2))
+            msg = f"Similarity between {i} and {j}: {similarity:.10f}"
+            if context.debug:
+                print(f"{msg}\n")
+            assert np.isclose(similarity, 1.0, rtol=1e-05, atol=1e-08, equal_nan=False), msg
+
 @step(u'embeddings are generated')
 def step_assert_embeddings(context):
-    if len(context.prompts) == 0:
-        assert_embeddings(context.embeddings)
-    else:
-        assert len(context.embeddings) == len(context.prompts), (f"unexpected response:\n"
-                                                                 f"context.prompts={context.prompts}\n"
-                                                                 f"context.embeddings={context.embeddings}")
-        for embedding in context.embeddings:
-            context.prompts.pop()
-            assert_embeddings(embedding)
+    assert context.n_prompts == len(context.embeddings), (f"unexpected response:\n"
+                                                             f"context.n_prompts={context.n_prompts}\n"
+                                                             f"context.embeddings={context.embeddings}")
+    for embedding in context.embeddings:
+        assert_embeddings(embedding)
 
 
 @step(u'an OAI compatible embeddings computation request for')
 @async_run_until_complete
 async def step_oai_compute_embeddings(context):
+    context.n_prompts = 1
     context.embeddings = await request_oai_embeddings(context.text,
                                                       base_url=context.base_url,
                                                       user_api_key=context.user_api_key,
@@ -462,6 +499,7 @@ async def step_oai_compute_embeddings_multiple_inputs(context):
                                                       base_url=context.base_url,
                                                       user_api_key=context.user_api_key,
                                                       model=context.model)
+    context.prompts.clear()
 
 
 @step(u'concurrent embedding requests')
@@ -488,9 +526,9 @@ async def step_concurrent_oai_embedding_requests(context):
 @async_run_until_complete()
 async def all_embeddings_are_generated(context):
     n_embedding_requests = await gather_tasks_results(context)
-    assert n_embedding_requests > 0
+    assert n_embedding_requests == context.n_prompts
     for i in range(n_embedding_requests):
-        assert_embeddings(context.tasks_result.pop())
+        assert_embeddings(context.tasks_result.pop().pop())
 
 
 @step(u'tokenizing')
@@ -588,11 +626,11 @@ def step_supported_models(context, i_model, param, preposition, param_value):
 
 
 async def concurrent_requests(context, f_completion, *args, **kwargs):
-    n_prompts = len(context.prompts)
+    context.n_prompts = len(context.prompts)
     if context.debug:
-        print(f"starting {n_prompts} concurrent completion requests...")
-    assert n_prompts > 0
-    for prompt_no in range(n_prompts):
+        print(f"starting {context.n_prompts} concurrent completion requests...")
+    assert context.n_prompts > 0
+    for prompt_no in range(context.n_prompts):
         shifted_args = [context.prompts.pop(), *args]
         context.concurrent_tasks.append(asyncio.create_task(f_completion(*shifted_args, **kwargs)))
     await asyncio.sleep(0.1)
@@ -765,7 +803,7 @@ async def request_embedding(content, base_url=None):
                                 }) as response:
             assert response.status == 200
             response_json = await response.json()
-            return response_json['embedding']
+            return [response_json['embedding']]
 
 
 async def request_oai_embeddings(input,
@@ -775,6 +813,7 @@ async def request_oai_embeddings(input,
     user_api_key = user_api_key if user_api_key is not None else 'nope'
     if async_client:
         origin = 'llama.cpp'
+        headers=[]
         if user_api_key is not None:
             headers = {'Authorization': f'Bearer {user_api_key}', 'Origin': origin}
         async with aiohttp.ClientSession() as session:
@@ -783,14 +822,21 @@ async def request_oai_embeddings(input,
                                         "input": input,
                                         "model": model,
                                     },
-                                    headers=headers) as response:
+                                    headers=headers,
+                                    timeout=3600) as response:
                 assert response.status == 200, f"received status code not expected: {response.status}"
                 assert response.headers['Access-Control-Allow-Origin'] == origin
                 assert response.headers['Content-Type'] == "application/json; charset=utf-8"
                 response_json = await response.json()
                 assert response_json['model'] == model, f"invalid model received: {response_json['model']}"
                 assert response_json['object'] == 'list'
-                return response_json['data']
+                if isinstance(input, collections.abc.Sequence):
+                    embeddings = []
+                    for an_oai_embeddings in response_json['data']:
+                        embeddings.append(an_oai_embeddings['embedding'])
+                else:
+                    embeddings = [response_json['data']['embedding']]
+                return embeddings
     else:
         openai.api_key = user_api_key
         openai.api_base = f'{base_url}/v1'
@@ -804,7 +850,7 @@ async def request_oai_embeddings(input,
             for an_oai_embeddings in oai_embeddings.data:
                 embeddings.append(an_oai_embeddings.embedding)
         else:
-            embeddings = oai_embeddings.data.embedding
+            embeddings = [oai_embeddings.data.embedding]
         return embeddings
 
 
@@ -899,6 +945,8 @@ def assert_embeddings(embeddings):
     assert len(embeddings) > 0
     embeddings_computed = False
     for emb in embeddings:
+        if not isinstance(emb, float):
+            assert False, f"Bad embeddings: {embeddings}"
         if emb != 0:
             embeddings_computed = True
     assert embeddings_computed, f"Embeddings: {embeddings}"

examples/server/tests/requirements.txt

@@ -1,5 +1,6 @@
 aiohttp~=3.9.3
 behave~=1.2.6
 huggingface_hub~=0.20.3
+numpy~=1.24.4
 openai~=0.25.0
 prometheus-client~=0.20.0

examples/server/utils.hpp

@@ -1,15 +1,16 @@
 #pragma once
 
-#include <string>
-#include <vector>
-#include <set>
-#include <mutex>
-#include <condition_variable>
-#include <unordered_map>
+#include "llama.h"
+#include "common.h"
 
 #include "json.hpp"
 
-#include "../llava/clip.h"
+#include <string>
+#include <vector>
+#include <sstream>
+#include <random>
+
+#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
 
 using json = nlohmann::json;
 
@@ -37,83 +38,35 @@ extern bool server_log_json;
 #define LOG_WARNING(MSG, ...) server_log("WARN", __func__, __LINE__, MSG, __VA_ARGS__)
 #define LOG_INFO(   MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__)
 
-enum server_state {
-    SERVER_STATE_LOADING_MODEL,  // Server is starting up, model not fully loaded yet
-    SERVER_STATE_READY,          // Server is ready and model is loaded
-    SERVER_STATE_ERROR           // An error occurred, load_model failed
-};
-
-enum task_type {
-    TASK_TYPE_COMPLETION,
-    TASK_TYPE_CANCEL,
-    TASK_TYPE_NEXT_RESPONSE,
-    TASK_TYPE_METRICS
-};
-
-struct task_server {
-    int id = -1; // to be filled by llama_server_queue
-    int target_id;
-    task_type type;
-    json data;
-    bool infill_mode = false;
-    bool embedding_mode = false;
-    int multitask_id = -1;
-};
-
-struct task_result {
-    int id;
-    int multitask_id = -1;
-    bool stop;
-    bool error;
-    json result_json;
-};
-
-struct task_multi {
-    int id;
-    std::set<int> subtasks_remaining{};
-    std::vector<task_result> results{};
-};
-
-// completion token output with probabilities
-struct completion_token_output {
-    struct token_prob
-    {
-        llama_token tok;
-        float prob;
-    };
-
-    std::vector<token_prob> probs;
-    llama_token tok;
-    std::string text_to_send;
-};
-
-struct token_translator {
-    llama_context * ctx;
-    std::string operator()(llama_token tok)                    const { return llama_token_to_piece(ctx, tok); }
-    std::string operator()(const completion_token_output &cto) const { return (*this)(cto.tok); }
-};
+template <typename T>
+static T json_value(const json &body, const std::string &key, const T &default_value) {
+    // Fallback null to default value
+    return body.contains(key) && !body.at(key).is_null()
+        ? body.value(key, default_value)
+        : default_value;
+}
 
 static inline void server_log(const char *level, const char *function, int line, const char *message, const nlohmann::ordered_json &extra) {
     std::stringstream ss_tid;
     ss_tid << std::this_thread::get_id();
     json log = nlohmann::ordered_json{
-        {"tid", ss_tid.str()},
+        {"tid",       ss_tid.str()},
         {"timestamp", time(nullptr)},
     };
 
     if (server_log_json) {
-        log.merge_patch(
-                {
-                        {"level",     level},
-                        {"function",  function},
-                        {"line",      line},
-                        {"msg",       message},
-                });
+        log.merge_patch( {
+            {"level",    level},
+            {"function", function},
+            {"line",     line},
+            {"msg",      message},
+        });
+
         if (!extra.empty()) {
             log.merge_patch(extra);
         }
 
-        std::cout << log.dump(-1, ' ', false, json::error_handler_t::replace) << "\n" << std::flush;
+        printf("%s\n", log.dump(-1, ' ', false, json::error_handler_t::replace).c_str());
     } else {
         char buf[1024];
         snprintf(buf, 1024, "%4s [%24s] %s", level, function, message);
@@ -136,22 +89,13 @@ static inline void server_log(const char *level, const char *function, int line,
 }
 
 //
-// server utils
+// chat template utils
 //
 
-template <typename T>
-static T json_value(const json &body, const std::string &key, const T &default_value) {
-    // Fallback null to default value
-    return body.contains(key) && !body.at(key).is_null()
-        ? body.value(key, default_value)
-        : default_value;
-}
-
 // Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
 inline bool verify_custom_template(const std::string & tmpl) {
     llama_chat_message chat[] = {{"user", "test"}};
-    std::vector<char> buf(1);
-    int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, buf.data(), buf.size());
+    int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, nullptr, 0);
     return res >= 0;
 }
 
@@ -163,7 +107,7 @@ inline std::string format_chat(const struct llama_model * model, const std::stri
     std::vector<llama_chat_message> chat(messages.size());
 
     for (size_t i = 0; i < messages.size(); ++i) {
-        auto &curr_msg = messages[i];
+        const auto & curr_msg = messages[i];
         str[i*2 + 0]    = json_value(curr_msg, "role",    std::string(""));
         str[i*2 + 1]    = json_value(curr_msg, "content", std::string(""));
         alloc_size     += str[i*2 + 1].length();
@@ -183,261 +127,13 @@ inline std::string format_chat(const struct llama_model * model, const std::stri
         res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size());
     }
 
-    std::string formatted_chat(buf.data(), res);
+    const std::string formatted_chat(buf.data(), res);
+
     LOG_VERBOSE("formatted_chat", {{"text", formatted_chat.c_str()}});
 
     return formatted_chat;
 }
 
-//
-// work queue utils
-//
-
-struct llama_server_queue {
-    int id = 0;
-    std::mutex mutex_tasks;
-    bool running;
-    // queues
-    std::vector<task_server> queue_tasks;
-    std::vector<task_server> queue_tasks_deferred;
-    std::vector<task_multi> queue_multitasks;
-    std::condition_variable condition_tasks;
-    // callback functions
-    std::function<void(task_server&)> callback_new_task;
-    std::function<void(task_multi&)> callback_finish_multitask;
-    std::function<void(void)> callback_run_slots;
-
-    // Add a new task to the end of the queue
-    int post(task_server task) {
-        std::unique_lock<std::mutex> lock(mutex_tasks);
-        if (task.id == -1) {
-            task.id = id++;
-            LOG_VERBOSE("new task id", {{"new_id", task.id}});
-        }
-        queue_tasks.push_back(std::move(task));
-        condition_tasks.notify_one();
-        return task.id;
-    }
-
-    // Add a new task, but defer until one slot is available
-    void defer(task_server task) {
-        std::unique_lock<std::mutex> lock(mutex_tasks);
-        queue_tasks_deferred.push_back(std::move(task));
-    }
-
-    // Get the next id for creating anew task
-    int get_new_id() {
-        std::unique_lock<std::mutex> lock(mutex_tasks);
-        int new_id = id++;
-        LOG_VERBOSE("new task id", {{"new_id", new_id}});
-        return new_id;
-    }
-
-    // Register function to process a new task
-    void on_new_task(std::function<void(task_server&)> callback) {
-        callback_new_task = callback;
-    }
-
-    // Register function to process a multitask when it is finished
-    void on_finish_multitask(std::function<void(task_multi&)> callback) {
-        callback_finish_multitask = callback;
-    }
-
-    // Register the function to be called when all slots data is ready to be processed
-    void on_run_slots(std::function<void(void)> callback) {
-        callback_run_slots = callback;
-    }
-
-    // Call when the state of one slot is changed
-    void notify_slot_changed() {
-        // move deferred tasks back to main loop
-        std::unique_lock<std::mutex> lock(mutex_tasks);
-        for (auto & task : queue_tasks_deferred) {
-            queue_tasks.push_back(std::move(task));
-        }
-        queue_tasks_deferred.clear();
-    }
-
-    // end the start_loop routine
-    void terminate() {
-        {
-            std::unique_lock<std::mutex> lock(mutex_tasks);
-            running = false;
-        }
-        condition_tasks.notify_all();
-    }
-
-    /**
-     * Main loop consists of these steps:
-     * - Wait until a new task arrives
-     * - Process the task (i.e. maybe copy data into slot)
-     * - Check if multitask is finished
-     * - Run all slots
-     */
-    void start_loop() {
-        running = true;
-        while (true) {
-            LOG_VERBOSE("new task may arrive", {});
-            {
-                while (true)
-                {
-                    std::unique_lock<std::mutex> lock(mutex_tasks);
-                    if (queue_tasks.empty()) {
-                        lock.unlock();
-                        break;
-                    }
-                    task_server task = queue_tasks.front();
-                    queue_tasks.erase(queue_tasks.begin());
-                    lock.unlock();
-                    LOG_VERBOSE("callback_new_task", {{"task_id", task.id}});
-                    callback_new_task(task);
-                }
-                LOG_VERBOSE("update_multitasks", {});
-                // check if we have any finished multitasks
-                auto queue_iterator = queue_multitasks.begin();
-                while (queue_iterator != queue_multitasks.end())
-                {
-                    if (queue_iterator->subtasks_remaining.empty())
-                    {
-                        // all subtasks done == multitask is done
-                        task_multi current_multitask = *queue_iterator;
-                        callback_finish_multitask(current_multitask);
-                        // remove this multitask
-                        queue_iterator = queue_multitasks.erase(queue_iterator);
-                    }
-                    else
-                    {
-                        ++queue_iterator;
-                    }
-                }
-                // all tasks in the current loop is processed, slots data is now ready
-                LOG_VERBOSE("callback_run_slots", {});
-                callback_run_slots();
-            }
-            LOG_VERBOSE("wait for new task", {});
-            // wait for new task
-            {
-                std::unique_lock<std::mutex> lock(mutex_tasks);
-                if (queue_tasks.empty()) {
-                    if (!running) {
-                        LOG_VERBOSE("ending start_loop", {});
-                        return;
-                    }
-                    condition_tasks.wait(lock, [&]{
-                        return (!queue_tasks.empty() || !running);
-                    });
-                }
-            }
-        }
-    }
-
-    //
-    // functions to manage multitasks
-    //
-
-    // add a multitask by specifying the id of all subtask (subtask is a task_server)
-    void add_multitask(int multitask_id, std::vector<int>& sub_ids)
-    {
-        std::lock_guard<std::mutex> lock(mutex_tasks);
-        task_multi multi;
-        multi.id = multitask_id;
-        std::copy(sub_ids.begin(), sub_ids.end(), std::inserter(multi.subtasks_remaining, multi.subtasks_remaining.end()));
-        queue_multitasks.push_back(multi);
-    }
-
-    // updatethe remaining subtasks, while appending results to multitask
-    void update_multitask(int multitask_id, int subtask_id, task_result& result)
-    {
-        std::lock_guard<std::mutex> lock(mutex_tasks);
-        for (auto& multitask : queue_multitasks)
-        {
-            if (multitask.id == multitask_id)
-            {
-                multitask.subtasks_remaining.erase(subtask_id);
-                multitask.results.push_back(result);
-            }
-        }
-    }
-};
-
-struct llama_server_response {
-    typedef std::function<void(int, int, task_result&)> callback_multitask_t;
-    callback_multitask_t callback_update_multitask;
-    // for keeping track of all tasks waiting for the result
-    std::set<int> waiting_task_ids;
-    // the main result queue
-    std::vector<task_result> queue_results;
-    std::mutex mutex_results;
-    std::condition_variable condition_results;
-
-    // add the task_id to the list of tasks waiting for response
-    void add_waiting_task_id(int task_id) {
-        LOG_VERBOSE("waiting for task id", {{"task_id", task_id}});
-        std::unique_lock<std::mutex> lock(mutex_results);
-        waiting_task_ids.insert(task_id);
-    }
-
-    // when the request is finished, we can remove task associated with it
-    void remove_waiting_task_id(int task_id) {
-        LOG_VERBOSE("remove waiting for task id", {{"task_id", task_id}});
-        std::unique_lock<std::mutex> lock(mutex_results);
-        waiting_task_ids.erase(task_id);
-    }
-
-    // This function blocks the thread until there is a response for this task_id
-    task_result recv(int task_id) {
-        while (true)
-        {
-            std::unique_lock<std::mutex> lock(mutex_results);
-            condition_results.wait(lock, [&]{
-                return !queue_results.empty();
-            });
-
-            for (int i = 0; i < (int) queue_results.size(); i++)
-            {
-                if (queue_results[i].id == task_id)
-                {
-                    assert(queue_results[i].multitask_id == -1);
-                    task_result res = queue_results[i];
-                    queue_results.erase(queue_results.begin() + i);
-                    return res;
-                }
-            }
-        }
-
-        // should never reach here
-    }
-
-    // Register the function to update multitask
-    void on_multitask_update(callback_multitask_t callback) {
-        callback_update_multitask = callback;
-    }
-
-    // Send a new result to a waiting task_id
-    void send(task_result result) {
-        std::unique_lock<std::mutex> lock(mutex_results);
-        LOG_VERBOSE("send new result", {{"task_id", result.id}});
-        for (auto& task_id : waiting_task_ids) {
-            // LOG_TEE("waiting task id %i \n", task_id);
-            // for now, tasks that have associated parent multitasks just get erased once multitask picks up the result
-            if (result.multitask_id == task_id)
-            {
-                LOG_VERBOSE("callback_update_multitask", {{"task_id", task_id}});
-                callback_update_multitask(task_id, result.id, result);
-                continue;
-            }
-
-            if (result.id == task_id)
-            {
-                LOG_VERBOSE("queue_results.push_back", {{"task_id", task_id}});
-                queue_results.push_back(result);
-                condition_results.notify_all();
-                return;
-            }
-        }
-    }
-};
-
 //
 // base64 utils (TODO: move to common in the future)
 //
@@ -447,13 +143,11 @@ static const std::string base64_chars =
              "abcdefghijklmnopqrstuvwxyz"
              "0123456789+/";
 
-static inline bool is_base64(uint8_t c)
-{
+static inline bool is_base64(uint8_t c) {
     return (isalnum(c) || (c == '+') || (c == '/'));
 }
 
-static inline std::vector<uint8_t> base64_decode(const std::string & encoded_string)
-{
+static inline std::vector<uint8_t> base64_decode(const std::string & encoded_string) {
     int i = 0;
     int j = 0;
     int in_ = 0;
@@ -465,13 +159,10 @@ static inline std::vector<uint8_t> base64_decode(const std::string & encoded_str
 
     std::vector<uint8_t> ret;
 
-    while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_]))
-    {
+    while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_])) {
         char_array_4[i++] = encoded_string[in_]; in_++;
-        if (i == 4)
-        {
-            for (i = 0; i <4; i++)
-            {
+        if (i == 4) {
+            for (i = 0; i < 4; i++) {
                 char_array_4[i] = base64_chars.find(char_array_4[i]);
             }
 
@@ -479,23 +170,20 @@ static inline std::vector<uint8_t> base64_decode(const std::string & encoded_str
             char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
             char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];
 
-            for (i = 0; (i < 3); i++)
-            {
+            for (i = 0; (i < 3); i++) {
                 ret.push_back(char_array_3[i]);
             }
+
             i = 0;
         }
     }
 
-    if (i)
-    {
-        for (j = i; j <4; j++)
-        {
+    if (i) {
+        for (j = i; j < 4; j++) {
             char_array_4[j] = 0;
         }
 
-        for (j = 0; j <4; j++)
-        {
+        for (j = 0; j < 4; j++) {
             char_array_4[j] = base64_chars.find(char_array_4[j]);
         }
 
@@ -503,8 +191,7 @@ static inline std::vector<uint8_t> base64_decode(const std::string & encoded_str
         char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
         char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];
 
-        for (j = 0; (j < i - 1); j++)
-        {
+        for (j = 0; j < i - 1; j++) {
             ret.push_back(char_array_3[j]);
         }
     }
@@ -516,8 +203,7 @@ static inline std::vector<uint8_t> base64_decode(const std::string & encoded_str
 // random string / id
 //
 
-static std::string random_string()
-{
+static std::string random_string() {
     static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz");
 
     std::random_device rd;
@@ -532,10 +218,10 @@ static std::string random_string()
     return result;
 }
 
-static std::string gen_chatcmplid()
-{
+static std::string gen_chatcmplid() {
     std::stringstream chatcmplid;
     chatcmplid << "chatcmpl-" << random_string();
+
     return chatcmplid.str();
 }
 
@@ -543,91 +229,316 @@ static std::string gen_chatcmplid()
 // other common utils
 //
 
-static size_t common_part(const std::vector<llama_token> &a, const std::vector<llama_token> &b)
-{
+static size_t common_part(const std::vector<llama_token> & a, const std::vector<llama_token> & b) {
     size_t i;
-    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++)
-    {
-    }
+    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
+
     return i;
 }
 
-static bool ends_with(const std::string &str, const std::string &suffix)
-{
-    return str.size() >= suffix.size() &&
-           0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix);
+static bool ends_with(const std::string & str, const std::string & suffix) {
+    return str.size() >= suffix.size() && 0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix);
 }
 
-static size_t find_partial_stop_string(const std::string &stop,
-                                       const std::string &text)
-{
-    if (!text.empty() && !stop.empty())
-    {
+static size_t find_partial_stop_string(const std::string &stop, const std::string &text) {
+    if (!text.empty() && !stop.empty()) {
         const char text_last_char = text.back();
-        for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--)
-        {
-            if (stop[char_index] == text_last_char)
-            {
+        for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--) {
+            if (stop[char_index] == text_last_char) {
                 const std::string current_partial = stop.substr(0, char_index + 1);
-                if (ends_with(text, current_partial))
-                {
+                if (ends_with(text, current_partial)) {
                     return text.size() - char_index - 1;
                 }
             }
         }
     }
+
     return std::string::npos;
 }
 
 // TODO: reuse llama_detokenize
 template <class Iter>
-static std::string tokens_to_str(llama_context *ctx, Iter begin, Iter end)
-{
+static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
     std::string ret;
-    for (; begin != end; ++begin)
-    {
+    for (; begin != end; ++begin) {
         ret += llama_token_to_piece(ctx, *begin);
     }
+
     return ret;
 }
 
 // format incomplete utf-8 multibyte character for output
-static std::string tokens_to_output_formatted_string(const llama_context *ctx, const llama_token token)
-{
+static std::string tokens_to_output_formatted_string(const llama_context * ctx, const llama_token token) {
     std::string out = token == -1 ? "" : llama_token_to_piece(ctx, token);
+
     // if the size is 1 and first bit is 1, meaning it's a partial character
     //   (size > 1 meaning it's already a known token)
-    if (out.size() == 1 && (out[0] & 0x80) == 0x80)
-    {
+    if (out.size() == 1 && (out[0] & 0x80) == 0x80) {
         std::stringstream ss;
         ss << std::hex << (out[0] & 0xff);
         std::string res(ss.str());
         out = "byte: \\x" + res;
     }
+
     return out;
 }
 
+struct completion_token_output {
+    llama_token tok;
+    std::string text_to_send;
+
+    struct token_prob {
+        llama_token tok;
+        float prob;
+    };
+
+    std::vector<token_prob> probs;
+};
+
 // convert a vector of completion_token_output to json
-static json probs_vector_to_json(const llama_context *ctx, const std::vector<completion_token_output> &probs)
-{
+static json probs_vector_to_json(const llama_context * ctx, const std::vector<completion_token_output> & probs) {
     json out = json::array();
-    for (const auto &prob : probs)
-    {
+
+    for (const auto & prob : probs) {
         json probs_for_token = json::array();
-        for (const auto &p : prob.probs)
-        {
-            std::string tok_str = tokens_to_output_formatted_string(ctx, p.tok);
-            probs_for_token.push_back(json
-            {
+
+        for (const auto & p : prob.probs) {
+            const std::string tok_str = tokens_to_output_formatted_string(ctx, p.tok);
+            probs_for_token.push_back(json {
                 {"tok_str", tok_str},
                 {"prob",    p.prob},
             });
         }
-        std::string tok_str = tokens_to_output_formatted_string(ctx, prob.tok);
-        out.push_back(json{
+
+        const std::string tok_str = tokens_to_output_formatted_string(ctx, prob.tok);
+        out.push_back(json {
             {"content", tok_str},
             {"probs",   probs_for_token},
         });
     }
+
     return out;
 }
+
+//
+// OAI utils
+//
+
+static json oaicompat_completion_params_parse(
+    const struct llama_model * model,
+    const json & body, /* openai api json semantics */
+    const std::string & chat_template) {
+    json llama_params;
+
+    llama_params["__oaicompat"] = true;
+
+    // Map OpenAI parameters to llama.cpp parameters
+    //
+    // For parameters that are defined by the OpenAI documentation (e.g.
+    // temperature), we explicitly specify OpenAI's intended default; we
+    // need to do that because sometimes OpenAI disagrees with llama.cpp
+    //
+    // https://platform.openai.com/docs/api-reference/chat/create
+    llama_sampling_params default_sparams;
+    llama_params["model"]             = json_value(body,   "model",             std::string("unknown"));
+    llama_params["prompt"]            = format_chat(model, chat_template,       body["messages"]);
+    llama_params["cache_prompt"]      = json_value(body,   "cache_prompt",      false);
+    llama_params["temperature"]       = json_value(body,   "temperature",       0.0);
+    llama_params["top_k"]             = json_value(body,   "top_k",             default_sparams.top_k);
+    llama_params["top_p"]             = json_value(body,   "top_p",             1.0);
+    llama_params["n_predict"]         = json_value(body,   "max_tokens",        -1);
+    llama_params["logit_bias"]        = json_value(body,   "logit_bias",        json::object());
+    llama_params["frequency_penalty"] = json_value(body,   "frequency_penalty", 0.0);
+    llama_params["presence_penalty"]  = json_value(body,   "presence_penalty",  0.0);
+    llama_params["seed"]              = json_value(body,   "seed",              LLAMA_DEFAULT_SEED);
+    llama_params["stream"]            = json_value(body,   "stream",            false);
+    llama_params["mirostat"]          = json_value(body,   "mirostat",          default_sparams.mirostat);
+    llama_params["mirostat_tau"]      = json_value(body,   "mirostat_tau",      default_sparams.mirostat_tau);
+    llama_params["mirostat_eta"]      = json_value(body,   "mirostat_eta",      default_sparams.mirostat_eta);
+    llama_params["penalize_nl"]       = json_value(body,   "penalize_nl",       default_sparams.penalize_nl);
+    llama_params["typical_p"]         = json_value(body,   "typical_p",         default_sparams.typical_p);
+    llama_params["repeat_last_n"]     = json_value(body,   "repeat_last_n",     default_sparams.penalty_last_n);
+    llama_params["ignore_eos"]        = json_value(body,   "ignore_eos",        false);
+    llama_params["tfs_z"]             = json_value(body,   "tfs_z",             default_sparams.tfs_z);
+
+    if (body.count("grammar") != 0) {
+        llama_params["grammar"] = json_value(body, "grammar", json::object());
+    }
+
+    // Handle 'stop' field
+    if (body.contains("stop") && body["stop"].is_string()) {
+        llama_params["stop"] = json::array({body["stop"].get<std::string>()});
+    } else {
+        llama_params["stop"] = json_value(body, "stop", json::array());
+    }
+
+    // Ensure there is ChatML-specific end sequence among stop words
+    llama_params["stop"].push_back("<|im_end|>");
+
+    return llama_params;
+}
+
+static json format_final_response_oaicompat(const json & request, json result, bool streaming = false) {
+    bool stopped_word        = result.count("stopped_word") != 0;
+    bool stopped_eos         = json_value(result, "stopped_eos", false);
+    int num_tokens_predicted = json_value(result, "tokens_predicted", 0);
+    int num_prompt_tokens    = json_value(result, "tokens_evaluated", 0);
+    std::string content      = json_value(result, "content", std::string(""));
+
+    std::string finish_reason = "length";
+    if (stopped_word || stopped_eos) {
+        finish_reason = "stop";
+    }
+
+    json choices =
+        streaming ? json::array({json{{"finish_reason", finish_reason},
+                                        {"index", 0},
+                                        {"delta", json::object()}}})
+                  : json::array({json{{"finish_reason", finish_reason},
+                                        {"index", 0},
+                                        {"message", json{{"content", content},
+                                                         {"role", "assistant"}}}}});
+
+    std::time_t t = std::time(0);
+
+    json res = json {
+        {"choices", choices},
+        {"created", t},
+        {"model",
+            json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
+        {"object", streaming ? "chat.completion.chunk" : "chat.completion"},
+        {"usage", json {
+            {"completion_tokens", num_tokens_predicted},
+            {"prompt_tokens",     num_prompt_tokens},
+            {"total_tokens",      num_tokens_predicted + num_prompt_tokens}
+        }},
+        {"id", gen_chatcmplid()}
+    };
+
+    if (server_verbose) {
+        res["__verbose"] = result;
+    }
+
+    if (result.contains("completion_probabilities")) {
+        res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array());
+    }
+
+    return res;
+}
+
+// return value is vector as there is one case where we might need to generate two responses
+static std::vector<json> format_partial_response_oaicompat(json result) {
+    if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) {
+        return std::vector<json>({result});
+    }
+
+    bool first = json_value(result, "oaicompat_token_ctr", 0) == 0;
+    std::string modelname = json_value(result, "model", std::string(DEFAULT_OAICOMPAT_MODEL));
+
+    bool stopped_word   = json_value(result, "stopped_word",  false);
+    bool stopped_eos    = json_value(result, "stopped_eos",   false);
+    bool stopped_limit  = json_value(result, "stopped_limit", false);
+    std::string content = json_value(result, "content",       std::string(""));
+
+    std::string finish_reason;
+    if (stopped_word || stopped_eos) {
+        finish_reason = "stop";
+    }
+    if (stopped_limit) {
+        finish_reason = "length";
+    }
+
+    std::time_t t = std::time(0);
+
+    json choices;
+
+    if (!finish_reason.empty()) {
+        choices = json::array({json{{"finish_reason", finish_reason},
+                                    {"index", 0},
+                                    {"delta", json::object()}}});
+    } else {
+        if (first) {
+            if (content.empty()) {
+                choices = json::array({json{{"finish_reason", nullptr},
+                                            {"index", 0},
+                                            {"delta", json{{"role", "assistant"}}}}});
+            } else {
+                // We have to send this as two updates to conform to openai behavior
+                json initial_ret = json{{"choices", json::array({json{
+                                        {"finish_reason", nullptr},
+                                        {"index", 0},
+                                        {"delta", json{
+                                            {"role", "assistant"}
+                                        }}}})},
+                            {"created", t},
+                            {"id", gen_chatcmplid()},
+                            {"model", modelname},
+                            {"object", "chat.completion.chunk"}};
+
+                json second_ret = json{
+                            {"choices", json::array({json{{"finish_reason", nullptr},
+                                                            {"index", 0},
+                                                            {"delta", json{
+                                                            {"content", content}}}
+                                                            }})},
+                            {"created", t},
+                            {"id", gen_chatcmplid()},
+                            {"model", modelname},
+                            {"object", "chat.completion.chunk"}};
+
+                return std::vector<json>({initial_ret, second_ret});
+            }
+        } else {
+            // Some idiosyncrasy in task processing logic makes several trailing calls
+            // with empty content, we ignore these at the calee site.
+            if (content.empty()) {
+                return std::vector<json>({json::object()});
+            }
+
+            choices = json::array({json{
+                {"finish_reason", nullptr},
+                {"index", 0},
+                {"delta",
+                json{
+                    {"content", content},
+                }},
+            }});
+        }
+    }
+
+    json ret = json {
+        {"choices", choices},
+        {"created", t},
+        {"id",      gen_chatcmplid()},
+        {"model",   modelname},
+        {"object",  "chat.completion.chunk"}
+    };
+
+    return std::vector<json>({ret});
+}
+
+static json format_embeddings_response_oaicompat(const json & request, const json & embeddings) {
+    json res = json {
+        {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
+        {"object", "list"},
+        {"usage", json {
+            {"prompt_tokens", 0},
+            {"total_tokens", 0}
+        }},
+        {"data", embeddings}
+    };
+
+    return res;
+}
+
+static json format_tokenizer_response(const std::vector<llama_token> & tokens) {
+    return json {
+        {"tokens", tokens}
+    };
+}
+
+static json format_detokenized_response(const std::string & content) {
+    return json {
+        {"content", content}
+    };
+}

examples/speculative/README.md

@@ -6,3 +6,4 @@ More info:
 
 - https://github.com/ggerganov/llama.cpp/pull/2926
 - https://github.com/ggerganov/llama.cpp/pull/3624
+- https://github.com/ggerganov/llama.cpp/pull/5625

examples/speculative/speculative.cpp

@@ -5,6 +5,7 @@
 #include <cstdio>
 #include <string>
 #include <vector>
+#include <set>
 
 #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  100
 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
@@ -18,6 +19,7 @@ struct seq_draft {
     std::vector<int> i_batch_tgt;
 
     std::vector<llama_token> tokens;
+    std::vector<std::vector<llama_token_data>> dists;
 
     struct llama_sampling_context * ctx_sampling;
 };
@@ -37,12 +39,15 @@ int main(int argc, char ** argv) {
     // max number of parallel drafting sequences (i.e. tree branches)
     const int n_seq_dft = params.n_parallel;
 
-    // probability threshold for accepting a token from the draft model
-    const float p_accept = params.p_accept;
-
     // probability threshold for splitting a draft branch (only for n_seq_dft > 1)
     const float p_split  = params.p_split;
 
+    if (params.seed == LLAMA_DEFAULT_SEED) {
+        params.seed = time(NULL);
+    }
+    std::default_random_engine rng(params.seed);
+    std::uniform_real_distribution<> u_dist;
+
 #ifndef LOG_DISABLE_LOGS
     log_set_target(log_filename_generator("speculative", "log"));
     LOG_TEE("Log start\n");
@@ -166,7 +171,9 @@ int main(int argc, char ** argv) {
     std::vector<seq_draft> drafts(n_seq_dft);
 
     params.sparams.grammar.clear(); // the draft samplers will copy the target sampler's grammar
-    params.sparams.temp = -1.0f;    // force greedy sampling with probs for the draft model
+    if (params.sparams.temp == 0) {
+        params.sparams.temp = -1.0f; // force greedy sampling with probs for the draft model
+    }
 
     for (int s = 0; s < n_seq_dft; ++s) {
         drafts[s].ctx_sampling = llama_sampling_init(params.sparams);
@@ -182,12 +189,15 @@ int main(int argc, char ** argv) {
     drafts[0].i_batch_tgt[0] = 0;
 
     while (true) {
+        std::set<int> active_seqs = {};
+
         // print current draft sequences
         for (int s = 0; s < n_seq_dft; ++s) {
             if (!drafts[s].active) {
                 continue;
             }
 
+            active_seqs.insert(s);
             const auto & tokens = drafts[s].tokens;
 
             LOG("draft %d: %s\n", s, LOG_TOKENS_TOSTR_PRETTY(ctx_dft, tokens).c_str());
@@ -196,48 +206,156 @@ int main(int argc, char ** argv) {
         int i_dft  = 0;
         int s_keep = 0;
 
+        llama_token token_id;
+        std::string token_str;
+
+        // loop until we fail to accept a drafted token or we run out of drafted tokens
         while (true) {
-            LOG("sampling target: s_keep = %3d, i_dft = %3d, i_batch_tgt = %3d\n", s_keep, i_dft, drafts[s_keep].i_batch_tgt[i_dft]);
-
-            // sample from the target model
-            llama_token id = llama_sampling_sample(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft]);
-
-            llama_sampling_accept(ctx_sampling, ctx_tgt, id, true);
-
-            //LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, ctx_sampling->prev).c_str());
-
-            const std::string token_str = llama_token_to_piece(ctx_tgt, id);
-
-            if (!params.use_color) {
-                printf("%s", token_str.c_str());
-            }
-
-            if (id == llama_token_eos(model_tgt)) {
-                has_eos = true;
-            }
-
-            ++n_predict;
 
             // check if the target token matches any of the drafts
+            // for stochastic sampling, attempt to match the token with the drafted tokens
             {
-                bool matches = false;
+                bool accept = false;
+                if (params.sparams.temp > 0) {
+                    // stochastic verification
 
-                for (int s = 0; s < n_seq_dft; ++s) {
-                    if (!drafts[s].active) {
-                        continue;
+                    llama_token_data_array dist_tgt = llama_sampling_probability_distribution(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft]);
+                    float p_tgt = 0, p_dft = 0;
+
+                    // GGML_ASSERT(dist_tgt.size() == dist_dft.size());
+
+                    while (active_seqs.size() > 0) {
+                        // randomly select a sequence to verify from active sequences
+                        std::uniform_int_distribution<unsigned int> u_int_dist(0, active_seqs.size() - 1);
+                        int s = *std::next(active_seqs.begin(), u_int_dist(rng));
+                        if (i_dft >= (int) drafts[s].tokens.size()) {
+                            drafts[s].active = false;
+                            active_seqs.erase(s);
+                            continue;
+                        }
+                        if (accept) {
+                            // if we already accepted a token, we can skip the rest
+                            if (drafts[s].tokens[i_dft] != drafts[s_keep].tokens[i_dft]) {
+                                drafts[s].active = false;
+                                active_seqs.erase(s);
+                            }
+                            continue;
+                        }
+                        LOG("verifying sequence #%d at pos #%d from %d active sequence(s)\n", s, i_dft, (int) active_seqs.size());
+                        float r = u_dist(rng);
+                        llama_token_data_array dist_dft = { drafts[s].dists[i_dft].data() , drafts[s].dists[i_dft].size(), true };
+                        // acquire the token probabilities assigned by the draft and target models
+                        for (size_t i = 0; i < dist_tgt.size; i++) {
+                            if (dist_tgt.data[i].id == drafts[s].tokens[i_dft]) {
+                                p_tgt = dist_tgt.data[i].p;
+                            }
+                            if (dist_dft.data[i].id == drafts[s].tokens[i_dft]) {
+                                p_dft = dist_dft.data[i].p;
+                            }
+                            if (p_tgt && p_dft) {
+                                break;
+                            }
+                        }
+                        LOG("r = %f, p_dft = %f, p_tgt = %f\n", r, p_dft, p_tgt);
+                        if (r <= p_tgt / p_dft) {
+                            s_keep = s;
+                            accept = true;
+                            token_id = drafts[s].tokens[i_dft];
+                            token_str = llama_token_to_piece(ctx_tgt, token_id);
+                            llama_sampling_accept(ctx_sampling, ctx_tgt, token_id, true);
+
+                            LOG("draft token %d of sequence %d (%d, '%s') accepted\n", i_dft, s, token_id, token_str.c_str());
+                            break;
+                        } else {
+                            LOG("draft token %d of sequence %d (%d, '%s') rejected\n", i_dft, s, drafts[s].tokens[i_dft], llama_token_to_piece(ctx_tgt, drafts[s].tokens[i_dft]).c_str());
+                            drafts[s].active = false;
+
+                            // calculate residual probability
+                            GGML_ASSERT(dist_tgt.sorted);
+                            GGML_ASSERT(dist_dft.sorted);
+                            float sum_probs = 0.0f;
+
+                            // sort dist by id
+                            std::sort(dist_tgt.data, dist_tgt.data + dist_tgt.size, [](const llama_token_data &a, const llama_token_data &b) {
+                                return a.id < b.id;
+                            });
+                            std::sort(dist_dft.data, dist_dft.data + dist_dft.size, [](const llama_token_data &a, const llama_token_data &b) {
+                                return a.id < b.id;
+                            });
+
+                            for (size_t i = 0; i < dist_tgt.size; i++) {
+                                dist_tgt.data[i].p = std::max(0.0f, dist_tgt.data[i].p - dist_dft.data[i].p);
+                                sum_probs += dist_tgt.data[i].p;
+                            }
+                            for (size_t i = 0; i < dist_tgt.size; i++) {
+                                dist_tgt.data[i].p /= sum_probs;
+                            }
+
+                            // sort dist_tgt by p desc
+                            std::sort(dist_tgt.data, dist_tgt.data + dist_tgt.size, [](const llama_token_data &a, const llama_token_data &b) {
+                                return a.p > b.p;
+                            });
+                        }
+
+                        active_seqs.erase(s);
+                        for(int i = 0; i < n_seq_dft; i++) {
+                            if (i == s) {
+                                continue;
+                            }
+                            if (drafts[i].tokens[i_dft] == drafts[s].tokens[i_dft]) {
+                                // synchronize active status for sequences with the same drafted token
+                                drafts[i].active = drafts[i].active && accept;
+                                if (!drafts[i].active) {
+                                    active_seqs.erase(s);
+                                }
+                            }
+                        }
                     }
 
-                    if (i_dft < (int) drafts[s].tokens.size() && id == drafts[s].tokens[i_dft]) {
-                        LOG("the sampled target token matches the %dth drafted token of sequence %d (%d, '%s') - accepted\n", i_dft, s, id, token_str.c_str());
+                    if (!accept) {
+                        // all drafted tokens were rejected
+                        // sample from the target model
+                        LOG("all drafted tokens were rejected, sampling from residual distribution\n");
+                        token_id = llama_sample_token(ctx_tgt, &dist_tgt);
+                        llama_sampling_accept(ctx_sampling, ctx_tgt, token_id, true);
+                        token_str = llama_token_to_piece(ctx_tgt, token_id);
+                    }
 
-                        s_keep = s;
-                        matches = true;
-                    } else {
-                        drafts[s].active = false;
+                } else {
+                    // greedy verification
+
+                    // sample from the target model
+                    LOG("sampling target: s_keep = %3d, i_dft = %3d, i_batch_tgt = %3d\n", s_keep, i_dft, drafts[s_keep].i_batch_tgt[i_dft]);
+                    token_id = llama_sampling_sample(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft]);
+
+                    llama_sampling_accept(ctx_sampling, ctx_tgt, token_id, true);
+
+                    //LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, ctx_sampling->prev).c_str());
+
+                    token_str = llama_token_to_piece(ctx_tgt, token_id);
+
+                    for (int s = 0; s < n_seq_dft; ++s) {
+                        if (!drafts[s].active) {
+                            continue;
+                        }
+
+                        if (i_dft < (int) drafts[s].tokens.size() && token_id == drafts[s].tokens[i_dft]) {
+                            LOG("the sampled target token matches the %dth drafted token of sequence %d (%d, '%s') - accepted\n", i_dft, s, token_id, token_str.c_str());
+
+                            s_keep = s;
+                            accept = true;
+                        } else {
+                            drafts[s].active = false;
+                        }
                     }
                 }
 
-                if (matches) {
+                if (token_id == llama_token_eos(model_tgt)) {
+                    has_eos = true;
+                }
+                ++n_predict;
+
+                if (accept) {
                     ++n_accept;
                     ++n_past_tgt;
                     ++n_past_dft;
@@ -245,17 +363,21 @@ int main(int argc, char ** argv) {
                     if (params.use_color) {
                         // Color token according to its origin sequence
                         printf("\u001b[%dm%s\u001b[37m", (36 - s_keep % 6), token_str.c_str());
-                        fflush(stdout);
+                    } else {
+                        printf("%s", token_str.c_str());
                     }
+                    fflush(stdout);
                     continue;
+                } else {
+                    printf("%s", token_str.c_str());
+                    fflush(stdout);
+                    break;
                 }
             }
-            if (params.use_color) {
-                printf("%s", token_str.c_str());
-            }
-            fflush(stdout);
+        }
 
-            LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
+        {
+            LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", token_id, token_str.c_str());
 
             // TODO: simplify
             {
@@ -275,21 +397,21 @@ int main(int argc, char ** argv) {
                 drafts[s].active = false;
                 drafts[s].tokens.clear();
                 drafts[s].i_batch_tgt.clear();
+                drafts[s].dists.clear();
             }
             // note: will be erased after the speculation phase
-            drafts[0].tokens.push_back(id);
+            drafts[0].tokens.push_back(token_id);
+            drafts[0].dists.push_back(std::vector<llama_token_data>());
             drafts[0].i_batch_tgt.push_back(0);
 
             llama_batch_clear(batch_dft);
-            llama_batch_add  (batch_dft, id, n_past_dft, { 0 }, true);
+            llama_batch_add  (batch_dft, token_id, n_past_dft, { 0 }, true);
 
             llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
             // LOG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
-            llama_decode         (ctx_dft, batch_dft);
+            llama_decode(ctx_dft, batch_dft);
 
             ++n_past_dft;
-
-            break;
         }
 
         if (n_predict > params.n_predict || has_eos) {
@@ -334,12 +456,6 @@ int main(int argc, char ** argv) {
                             k, s, i, cur_p[k].id, cur_p[k].p, llama_token_to_piece(ctx_dft, cur_p[k].id).c_str());
                 }
 
-                if (cur_p[0].p < p_accept) {
-                    LOG("stopping drafting for seq %3d, probability too low: %.3f < %.3f\n", s, cur_p[0].p, p_accept);
-                    drafts[s].drafting = false;
-                    continue;
-                }
-
                 std::vector<int> sa(1, s);
 
                 // attempt to split the branch if the probability is high enough
@@ -367,6 +483,7 @@ int main(int argc, char ** argv) {
                         drafts[n_seq_cur].skip     = true;
 
                         drafts[n_seq_cur].tokens      = drafts[s].tokens;
+                        drafts[n_seq_cur].dists       = drafts[s].dists;
                         drafts[n_seq_cur].i_batch_dft = drafts[s].i_batch_dft;
                         drafts[n_seq_cur].i_batch_tgt = drafts[s].i_batch_tgt;
 
@@ -389,6 +506,8 @@ int main(int argc, char ** argv) {
                     llama_sampling_accept(drafts[s].ctx_sampling, ctx_dft, id, true);
 
                     drafts[s].tokens.push_back(id);
+                    // save cur_p.data into drafts[s].dists
+                    drafts[s].dists.push_back(cur_p);
 
                     // add unique drafted tokens to the target batch
                     drafts[s].i_batch_tgt.push_back(batch_tgt.n_tokens);
@@ -440,6 +559,7 @@ int main(int argc, char ** argv) {
             }
 
             drafts[s].tokens.erase(drafts[s].tokens.begin());
+            drafts[s].dists.erase(drafts[s].dists.begin());
         }
     }
 

ggml-backend-impl.h

@@ -91,13 +91,14 @@ extern "C" {
         // (optional) complete all pending operations
         void (*GGML_CALL synchronize)(ggml_backend_t backend);
 
-        // compute graph with a plan
+        // create a plan for ggml_cgraph and free it
         ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
         void                      (*GGML_CALL graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-        void                      (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
 
+        // compute graph with a plan
+        enum ggml_status (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
         // compute graph without a plan (async)
-        bool (*GGML_CALL graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        enum ggml_status (*GGML_CALL graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
 
         // check if the backend supports an operation
         bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);

ggml-backend.c

@@ -262,11 +262,11 @@ void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_pla
     backend->iface.graph_plan_free(backend, plan);
 }
 
-void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    backend->iface.graph_plan_compute(backend, plan);
+enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    return backend->iface.graph_plan_compute(backend, plan);
 }
 
-bool ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+enum ggml_status ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     return backend->iface.graph_compute(backend, cgraph);
 }
 
@@ -732,15 +732,15 @@ GGML_CALL static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, g
     GGML_UNUSED(backend);
 }
 
-GGML_CALL static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+GGML_CALL static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
     struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
 
-    ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
+    return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
 
     GGML_UNUSED(backend);
 }
 
-GGML_CALL static bool ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
 
     struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
@@ -755,8 +755,7 @@ GGML_CALL static bool ggml_backend_cpu_graph_compute(ggml_backend_t backend, str
     cplan.abort_callback      = cpu_ctx->abort_callback;
     cplan.abort_callback_data = cpu_ctx->abort_callback_data;
 
-    ggml_graph_compute(cgraph, &cplan);
-    return true;
+    return ggml_graph_compute(cgraph, &cplan);
 }
 
 GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
@@ -1437,7 +1436,7 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
     return true;
 }
 
-static bool ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
+static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
     uint64_t copy_us[GGML_MAX_BACKENDS] = {0};
     uint64_t compute_us[GGML_MAX_BACKENDS] = {0};
 
@@ -1472,8 +1471,9 @@ static bool ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
 
         uint64_t compute_start_us = ggml_time_us();
         if (!sched->callback_eval) {
-            if (!ggml_backend_graph_compute(split_backend, &split->graph)) {
-                return false;
+            enum ggml_status ec = ggml_backend_graph_compute(split_backend, &split->graph);
+            if (ec != GGML_STATUS_SUCCESS) {
+                return ec;
             }
             //ggml_backend_synchronize(split_backend); // necessary to measure compute time
         } else {
@@ -1494,8 +1494,9 @@ static bool ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
 
                 struct ggml_cgraph gv = ggml_graph_view(&split->graph, j0, j1 + 1);
 
-                if (!ggml_backend_graph_compute(split_backend, &gv)) {
-                    return false;
+                enum ggml_status ec = ggml_backend_graph_compute(split_backend, &gv);
+                if (ec != GGML_STATUS_SUCCESS) {
+                    return ec;
                 }
 
                 if (need && !sched->callback_eval(t, false, sched->callback_eval_user_data)) {
@@ -1519,7 +1520,7 @@ static bool ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
     }
 #endif
 
-    return true;
+    return GGML_STATUS_SUCCESS;
 }
 
 ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size) {
@@ -1581,7 +1582,7 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
     return true;
 }
 
-bool ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
     GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS);
 
     if (!sched->is_reset) {
@@ -1590,14 +1591,10 @@ bool ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cg
 
     ggml_backend_sched_split_graph(sched, graph);
     if (!ggml_backend_sched_alloc_splits(sched)) {
-        return false;
+        return GGML_STATUS_ALLOC_FAILED;
     }
 
-    if (!ggml_backend_sched_compute_splits(sched)) {
-        return false;
-    }
-
-    return true;
+    return ggml_backend_sched_compute_splits(sched);
 }
 
 void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data) {

ggml-backend.h

@@ -66,12 +66,13 @@ extern "C" {
 
     GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
 
-    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API void                      ggml_backend_graph_plan_free  (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
 
-    GGML_API void ggml_backend_graph_plan_free   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API bool ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API bool ggml_backend_supports_op       (ggml_backend_t backend, const struct ggml_tensor * op);
+    GGML_API enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API enum ggml_status ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+    GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
 
     // tensor copy between different backends
     GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
@@ -157,26 +158,26 @@ extern "C" {
     typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
 
     // Initialize a backend scheduler
-    GGML_API ggml_backend_sched_t  ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size);
-    GGML_API void                  ggml_backend_sched_free(ggml_backend_sched_t sched);
+    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size);
+    GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
     // Initialize backend buffers from a measure graph
-    GGML_API bool                  ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+    GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
     // Get the number of splits of the last graph
-    GGML_API int                   ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
+    GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
 
-    GGML_API size_t                ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
+    GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
 
-    GGML_API void                  ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
-    GGML_API ggml_backend_t        ggml_backend_sched_get_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
+    GGML_API void                 ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+    GGML_API ggml_backend_t       ggml_backend_sched_get_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
 
     // Allocate and compute graph on the backend scheduler
-    GGML_API bool                  ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
 
     // Reset all assignments and allocators - must be called before changing the node backends
-    GGML_API void                  ggml_backend_sched_reset(ggml_backend_sched_t sched);
+    GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
 
     // Set a callback to be called for each resulting node during graph compute
-    GGML_API void                  ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
+    GGML_API void                 ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
 
     //
     // Utils

ggml-cuda.cu

@@ -12241,7 +12241,7 @@ GGML_CALL static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
     UNUSED(backend);
 }
 
-GGML_CALL static bool ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
 
     ggml_cuda_set_main_device(cuda_ctx->device);
@@ -12277,7 +12277,7 @@ GGML_CALL static bool ggml_backend_cuda_graph_compute(ggml_backend_t backend, gg
         GGML_ASSERT(ok);
     }
 
-    return true;
+    return GGML_STATUS_SUCCESS;
 }
 
 GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) {

ggml-kompute.cpp

@@ -1927,10 +1927,10 @@ static ggml_backend_buffer_type_t ggml_backend_kompute_get_default_buffer_type(g
     return ggml_backend_kompute_buffer_type(ctx->device);
 }
 
-static bool ggml_backend_kompute_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+static ggml_status ggml_backend_kompute_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
     ggml_vk_graph_compute(ctx, cgraph);
-    return true;
+    return GGML_STATUS_SUCCESS;
 }
 
 static bool ggml_backend_kompute_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {

ggml-metal.m

@@ -748,7 +748,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
     }
 }
 
-static bool ggml_metal_graph_compute(
+static enum ggml_status ggml_metal_graph_compute(
         struct ggml_metal_context * ctx,
                struct ggml_cgraph * gf) {
 
@@ -2484,7 +2484,7 @@ static bool ggml_metal_graph_compute(
         MTLCommandBufferStatus status = [command_buffer status];
         if (status != MTLCommandBufferStatusCompleted) {
             GGML_METAL_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status);
-            return false;
+            return GGML_STATUS_FAILED;
         }
     }
 
@@ -2493,7 +2493,7 @@ static bool ggml_metal_graph_compute(
     }
 
     }
-    return true;
+    return GGML_STATUS_SUCCESS;
 }
 
 ////////////////////////////////////////////////////////////////////////////////
@@ -2795,7 +2795,7 @@ GGML_CALL static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffe
     UNUSED(backend);
 }
 
-GGML_CALL static bool ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+GGML_CALL static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
 
     return ggml_metal_graph_compute(metal_ctx, cgraph);

ggml-opencl.cpp

@@ -2231,7 +2231,7 @@ static ggml_backend_buffer_type_t ggml_backend_opencl_get_default_buffer_type(gg
     GGML_UNUSED(backend);
 }
 
-static bool ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * graph) {
+static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * graph) {
     for (int i = 0; i < graph->n_nodes; ++i) {
         ggml_tensor * node = graph->nodes[i];
         switch (node->op) {
@@ -2246,7 +2246,7 @@ static bool ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgrap
         }
     }
 
-    return true;
+    return GGML_STATUS_SUCCESS;
 
     GGML_UNUSED(backend);
 }

ggml-quants.c

@@ -51,6 +51,7 @@
 
 #define UNUSED GGML_UNUSED
 
+// some compilers don't provide _mm256_set_m128i, e.g. gcc 7
 #define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
 
 #if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
@@ -463,8 +464,8 @@ inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
 }
 
 // NOTE: not tested
-inline static int8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
-    int8x16_t res;
+inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
+    uint8x16_t res;
 
     res[ 0] = a[b[ 0]];
     res[ 1] = a[b[ 1]];
@@ -9563,7 +9564,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
 
         const __m128i odd_bits = _mm_shuffle_epi8(bit_helper, partial_sign_bits_for_counting);
         const __m128i full_sign_bits = _mm_or_si128(partial_sign_bits, odd_bits);
-        const __m256i full_signs = _mm256_set_m128i(full_sign_bits, full_sign_bits);
+        const __m256i full_signs = MM256_SET_M128I(full_sign_bits, full_sign_bits);
 
         const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)y[i].qs);
         const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)(y[i].qs+32));
@@ -9585,8 +9586,8 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
         const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
         const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
 
-        const __m256i sc1 = _mm256_set_m128i(_mm_set1_epi16(2*(x[i].scales[0] >> 4)+1), _mm_set1_epi16(2*(x[i].scales[0] & 0xf)+1));
-        const __m256i sc2 = _mm256_set_m128i(_mm_set1_epi16(2*(x[i].scales[1] >> 4)+1), _mm_set1_epi16(2*(x[i].scales[1] & 0xf)+1));
+        const __m256i sc1 = MM256_SET_M128I(_mm_set1_epi16(2*(x[i].scales[0] >> 4)+1), _mm_set1_epi16(2*(x[i].scales[0] & 0xf)+1));
+        const __m256i sc2 = MM256_SET_M128I(_mm_set1_epi16(2*(x[i].scales[1] >> 4)+1), _mm_set1_epi16(2*(x[i].scales[1] & 0xf)+1));
 
         const __m256i sum = _mm256_add_epi32(_mm256_madd_epi16(sc1, dot1), _mm256_madd_epi16(sc2, dot2));
 
@@ -9653,8 +9654,8 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
 
             const __m128i full_signs_l = _mm256_castsi256_si128(full_sign_bits);
             const __m128i full_signs_h = _mm256_extractf128_si256(full_sign_bits, 1);
-            const __m256i full_signs_1 = _mm256_set_m128i(full_signs_l, full_signs_l);
-            const __m256i full_signs_2 = _mm256_set_m128i(full_signs_h, full_signs_h);
+            const __m256i full_signs_1 = MM256_SET_M128I(full_signs_l, full_signs_l);
+            const __m256i full_signs_2 = MM256_SET_M128I(full_signs_h, full_signs_h);
 
             __m256i signs;
             signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_1);
@@ -10551,10 +10552,10 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
         const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[1].qs);
         const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)y[0].qs);
         const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)y[1].qs);
-        const __m256i q4b_1 = _mm256_set_m128i(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
-                                               _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
-        const __m256i q4b_2 = _mm256_set_m128i(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
-                                               _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
+        const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
+                                              _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
+        const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
+                                              _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
         const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
         const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
         const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
@@ -10661,10 +10662,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void *
             const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)qs);  qs += 16;
             const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
             const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q4b_1 = _mm256_set_m128i(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
-                                                   _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
-            const __m256i q4b_2 = _mm256_set_m128i(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
-                                                   _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
+            const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
+                                                  _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
+            const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
+                                                  _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
             const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
             const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
             const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32;

ggml-sycl.h

@@ -28,6 +28,7 @@ GGML_API GGML_CALL int   ggml_backend_sycl_get_device_count();
 GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
 GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
 GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id);
+GGML_API GGML_CALL void ggml_backend_sycl_set_single_device(int main_gpu);
 
 #ifdef  __cplusplus
 }

ggml-vulkan.h

@@ -10,6 +10,7 @@ extern "C" {
 #define GGML_VK_NAME "Vulkan"
 #define GGML_VK_MAX_DEVICES 16
 
+GGML_API void ggml_vk_instance_init(void);
 GGML_API void ggml_vk_init_cpu_assist(void);
 
 GGML_API void ggml_vk_preallocate_buffers_graph_cpu_assist(struct ggml_tensor * node);

ggml.c

@@ -320,6 +320,17 @@ static ggml_fp16_t ggml_table_exp_f16[1 << 16];
 // precomputed f32 table for f16 (256 KB) (ggml-impl.h)
 float ggml_table_f32_f16[1 << 16];
 
+const char * ggml_status_to_string(enum ggml_status status) {
+    switch (status) {
+        case GGML_STATUS_ALLOC_FAILED: return "GGML status: error (failed to allocate memory)";
+        case GGML_STATUS_FAILED:       return "GGML status: error (operation failed)";
+        case GGML_STATUS_SUCCESS:      return "GGML status: success";
+        case GGML_STATUS_ABORTED:      return "GGML status: warning (operation aborted)";
+    }
+
+    return "GGML status: unknown";
+}
+
 // note: do not use these inside ggml.c
 // these are meant to be used via the ggml.h API
 float ggml_fp16_to_fp32(ggml_fp16_t x) {
@@ -2143,7 +2154,10 @@ void ggml_numa_init(enum ggml_numa_strategy numa_flag) {
     getcpu_ret = getcpu(&current_cpu, &g_state.numa.current_node);
 #else
     // old glibc doesn't have a wrapper for this call. Fall back on direct syscall
-    getcpu_ret = syscall(SYS_getcpu,&current_cpu,&g_state.numa.current_node);
+#   if !defined(SYS_getcpu) && defined(SYS_get_cpu)
+#       define SYS_getcpu SYS_get_cpu // some older glibc versions use this name
+#   endif
+    getcpu_ret = syscall(SYS_getcpu, &current_cpu, &g_state.numa.current_node);
 #endif
 
     if (g_state.numa.n_nodes < 1 || g_state.numa.total_cpus < 1 || getcpu_ret != 0) {
@@ -17400,6 +17414,7 @@ struct ggml_compute_state {
     ggml_thread_t thrd;
     int ith;
     struct ggml_compute_state_shared * shared;
+    enum ggml_status ec;
 };
 
 static void ggml_graph_compute_perf_stats_node(struct ggml_tensor * node, const struct ggml_compute_state_shared * st) {
@@ -17693,7 +17708,8 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
     while (true) {
         if (cplan->abort_callback && cplan->abort_callback(cplan->abort_callback_data)) {
             state->shared->node_n += 1;
-            return (thread_ret_t) GGML_EXIT_ABORTED;
+            state->ec = GGML_STATUS_ABORTED;
+            return 0;
         }
 
         if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) {
@@ -17815,7 +17831,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
         }
     }
 
-    return GGML_EXIT_SUCCESS;
+    return 0;
 }
 
 struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threads) {
@@ -18011,7 +18027,7 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa
     return cplan;
 }
 
-int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) {
+enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) {
     {
         GGML_ASSERT(cplan);
         GGML_ASSERT(cplan->n_threads > 0);
@@ -18055,6 +18071,7 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) {
                 .thrd   = 0,
                 .ith = j,
                 .shared = &state_shared,
+                .ec = GGML_STATUS_SUCCESS,
             };
 
             const int rc = ggml_thread_create(&workers[j].thrd, NULL, ggml_graph_compute_thread, &workers[j]);
@@ -18065,12 +18082,14 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) {
 
     workers[0].ith = 0;
     workers[0].shared = &state_shared;
+    workers[0].ec = GGML_STATUS_SUCCESS;
 
     const int64_t perf_start_cycles  = ggml_perf_cycles();
     const int64_t perf_start_time_us = ggml_perf_time_us();
 
     // this is a work thread too
-    int compute_status = (size_t) ggml_graph_compute_thread(&workers[0]);
+    ggml_graph_compute_thread(&workers[0]);
+    enum ggml_status compute_status = workers[0].ec;
 
     // don't leave affinity set on the main thread
     clear_numa_thread_affinity();
@@ -18080,6 +18099,8 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) {
         for (int j = 1; j < n_threads; j++) {
             const int rc = ggml_thread_join(workers[j].thrd, NULL);
             GGML_ASSERT(rc == 0);
+            if (workers[j].ec != GGML_STATUS_SUCCESS)
+                compute_status = workers[j].ec;
         }
     }
 
@@ -18107,14 +18128,14 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) {
     return compute_status;
 }
 
-void ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads) {
+enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads) {
     struct ggml_cplan cplan = ggml_graph_plan(cgraph, n_threads);
 
     struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_TYPE_WORK_BUFFER, cplan.work_size);
 
     cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs;
 
-    ggml_graph_compute(cgraph, &cplan);
+    return ggml_graph_compute(cgraph, &cplan);
 }
 
 struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name) {

ggml.h

@@ -315,6 +315,16 @@
 extern "C" {
 #endif
 
+    enum ggml_status {
+        GGML_STATUS_ALLOC_FAILED = -2,
+        GGML_STATUS_FAILED = -1,
+        GGML_STATUS_SUCCESS = 0,
+        GGML_STATUS_ABORTED = 1,
+    };
+
+    // get ggml_status name string
+    GGML_API GGML_CALL const char * ggml_status_to_string(enum ggml_status status);
+
     typedef uint16_t ggml_fp16_t;
 
     // convert FP16 <-> FP32
@@ -1940,12 +1950,11 @@ extern "C" {
 
     // ggml_graph_plan() has to be called before ggml_graph_compute()
     // when plan.work_size > 0, caller must allocate memory for plan.work_data
-    GGML_API struct ggml_cplan ggml_graph_plan   (const struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
-    GGML_API int               ggml_graph_compute(      struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
-
+    GGML_API struct ggml_cplan ggml_graph_plan            (const struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
+    GGML_API enum ggml_status  ggml_graph_compute         (      struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
     // same as ggml_graph_compute() but the work data is allocated as a part of the context
     // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
-    GGML_API void ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
+    GGML_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
 
     GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);
 

grammars/json.gbnf

@@ -15,7 +15,7 @@ array  ::=
 
 string ::=
   "\"" (
-    [^"\\] |
+    [^"\\\x7F\x00-\x1F] |
     "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) # escapes
   )* "\"" ws
 

grammars/json_arr.gbnf

@@ -24,7 +24,7 @@ array  ::=
 
 string ::=
   "\"" (
-    [^"\\] |
+    [^"\\\x7F\x00-\x1F] |
     "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) # escapes
   )* "\"" ws
 

llama.cpp

@@ -1665,7 +1665,7 @@ struct llama_hparams {
 };
 
 struct llama_cparams {
-    uint32_t n_ctx;       // context size used during inference
+    uint32_t n_ctx;           // context size used during inference
     uint32_t n_batch;
     uint32_t n_threads;       // number of threads to use for generation
     uint32_t n_threads_batch; // number of threads to use for batch processing
@@ -1682,7 +1682,9 @@ struct llama_cparams {
     float yarn_beta_slow;
     float defrag_thold;
 
+    bool embeddings;
     bool offload_kqv;
+
     enum llama_pooling_type pooling_type;
 
     ggml_backend_sched_eval_callback cb_eval;
@@ -1972,7 +1974,7 @@ struct llama_context {
     int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1)
     int32_t n_eval   = 0; // number of eval calls
 
-    // decode output (2-dimensional array: [n_tokens][n_vocab])
+    // logits output (2-dimensional array: [n_tokens][n_vocab])
     std::vector<float> logits;
 #ifndef NDEBUG
     // guard against access to unset logits
@@ -1980,8 +1982,13 @@ struct llama_context {
 #endif
     bool logits_all = false;
 
-    // input embedding (1-dimensional array: [n_embd])
-    std::vector<float> embedding;
+    // embeddings output (2-dimensional array: [n_tokens][n_embd])
+    // populated only when pooling_type == LLAMA_POOLING_TYPE_NONE
+    std::vector<float> embd;
+
+    // sequence embeddings output (map of [n_embd] vectors)
+    // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE
+    std::map<llama_seq_id, std::vector<float>> embd_seq;
 
     // memory buffers used to evaluate the model
     std::vector<uint8_t> buf_compute_meta;
@@ -3743,6 +3750,14 @@ static bool llm_load_tensors(
     model.main_gpu     = main_gpu;
     model.n_gpu_layers = n_gpu_layers;
 
+#ifdef GGML_USE_SYCL
+    if (split_mode == LLAMA_SPLIT_MODE_NONE) {
+        ggml_backend_sycl_set_single_device(main_gpu);
+        //SYCL use device index (0, 1, 2), instead if device id.
+        main_gpu = ggml_backend_sycl_get_device_index(main_gpu);
+    }
+#endif
+
     const int64_t n_layer     = hparams.n_layer;
     const int64_t i_gpu_start = std::max((int64_t) hparams.n_layer - n_gpu_layers, (int64_t) 0);
 
@@ -5007,8 +5022,8 @@ static struct ggml_tensor * llm_build_kqv(
         ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
     }
 
-#if defined(GGML_USE_VULKAN) || defined(GGML_USE_KOMPUTE)
-#pragma message("TODO: ALiBi support in ggml_soft_max_ext is not implemented for Vulkan, and Kompute")
+#if defined(GGML_USE_KOMPUTE)
+#pragma message("TODO: ALiBi support in ggml_soft_max_ext is not implemented for Kompute")
 #pragma message("      Falling back to ggml_alibi(). Will become an error in Mar 2024")
 #pragma message("ref:  https://github.com/ggerganov/llama.cpp/pull/5488")
     if (hparams.f_max_alibi_bias > 0.0f) {
@@ -5092,6 +5107,7 @@ static struct ggml_tensor * llm_build_kv(
     llm_build_kv_store(ctx, hparams, kv, graph, k_cur, v_cur, n_ctx, n_tokens, kv_head, cb, il);
 
     struct ggml_tensor * cur;
+
     cur  = llm_build_kqv(ctx, model, hparams, kv, graph, wo, wo_b,
             q_cur, kq_mask, kq_pos, n_ctx, n_tokens, n_kv, kq_scale, cb, il);
     cb(cur, "kqv_out", il);
@@ -6085,6 +6101,7 @@ struct llm_build_context {
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
 
         struct ggml_tensor * cur;
@@ -6092,9 +6109,10 @@ struct llm_build_context {
 
         // get input vectors with right size
         const size_t stride1 = n_tokens * ggml_type_size(lctx.inp_tokens->type);
-        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
+
+        struct ggml_tensor * inp_pos  = ggml_view_1d(ctx0, lctx.inp_pos,  n_tokens, 0);
         struct ggml_tensor * inp_mean = ggml_view_2d(ctx0, lctx.inp_mean, n_tokens, n_tokens, stride1, 0);
-        struct ggml_tensor * inp_cls = ggml_view_1d(ctx0, lctx.inp_cls, n_tokens, 0);
+        struct ggml_tensor * inp_cls  = ggml_view_1d(ctx0, lctx.inp_cls,  n_tokens, 0);
 
         // construct input embeddings (token, type, position)
         inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
@@ -6112,39 +6130,38 @@ struct llm_build_context {
         cb(inpL, "inp_norm", -1);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
-        cb(KQ_mask, "KQ_mask", -1); // [n_kv, n_tokens]
+        struct ggml_tensor * KQ_mask = ggml_cont(ctx0, ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_tokens, n_tokens, n_tokens*ggml_type_size(lctx.inp_KQ_mask->type), 0));
+        cb(KQ_mask, "KQ_mask", -1); // [n_tokens, n_tokens]
 
         // iterate layers
         for (int il = 0; il < n_layer; ++il) {
             struct ggml_tensor * cur = inpL;
 
+            struct ggml_tensor * Qcur;
+            struct ggml_tensor * Kcur;
+            struct ggml_tensor * Vcur;
+
             // self-attention
             if (model.arch == LLM_ARCH_BERT) {
-                struct ggml_tensor * Qcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, cur), model.layers[il].bq);
+                Qcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, cur), model.layers[il].bq);
                 cb(Qcur, "Qcur", il);
 
-                struct ggml_tensor * Kcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, cur), model.layers[il].bk);
+                Kcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, cur), model.layers[il].bk);
                 cb(Kcur, "Kcur", il);
 
-                struct ggml_tensor * Vcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, cur), model.layers[il].bv);
+                Vcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, cur), model.layers[il].bv);
                 cb(Vcur, "Vcur", il);
 
-                // seems like we just need to do this for Q?
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-
-                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-                cb(cur, "kqv_out", il);
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
             } else {
                 // compute Q and K and RoPE them
                 cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
                 cb(cur, "wqkv", il);
 
-                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
 
                 cb(Qcur, "Qcur", il);
                 cb(Kcur, "Kcur", il);
@@ -6163,13 +6180,41 @@ struct llm_build_context {
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-                cb(cur, "kqv_out", il);
             }
 
+            struct ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+            struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
+
+            struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+            cb(kq, "kq", il);
+
+            kq = ggml_soft_max_ext(ctx0, kq, KQ_mask, nullptr, 1.0f/sqrtf(float(n_embd_head)), hparams.f_max_alibi_bias);
+            cb(kq, "kq_soft_max_ext", il);
+
+            struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens)));
+            cb(v, "v", il);
+
+            struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_tokens, n_embd_head, n_head_kv), kq);
+            cb(kqv, "kqv", il);
+
+            struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+            cb(kqv_merged, "kqv_merged", il);
+
+            cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
+            cb(cur, "kqv_merged_cont", il);
+
+            ggml_build_forward_expand(gf, cur);
+
+            cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur);
+            if (model.layers[il].bo) {
+                cb(cur, "kqv_wo", il);
+            }
+
+            if (model.layers[il].bo) {
+                cur = ggml_add(ctx0, cur, model.layers[il].bo);
+            }
+            cb(cur, "kqv_out", il);
+
             // re-add the layer input
             cur = ggml_add(ctx0, cur, inpL);
 
@@ -6209,16 +6254,29 @@ struct llm_build_context {
 
         // final output
         cur = inpL;
+        cb(cur, "result_embd", -1);
 
         // pooling layer
-        if (pooling_type == LLAMA_POOLING_TYPE_MEAN) {
-            cur = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, cur)), inp_mean);
-        } else if (pooling_type == LLAMA_POOLING_TYPE_CLS) {
-            cur = ggml_get_rows(ctx0, cur, inp_cls);
-        } else {
-            GGML_ASSERT(pooling_type == LLAMA_POOLING_TYPE_NONE && "Invalid pooling type");
+        switch (pooling_type) {
+            case LLAMA_POOLING_TYPE_NONE:
+                {
+                    // nop
+                } break;
+            case LLAMA_POOLING_TYPE_MEAN:
+                {
+                    cur = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, cur)), inp_mean);
+                    cb(cur, "result_embd_pooled", -1);
+                } break;
+            case LLAMA_POOLING_TYPE_CLS:
+                {
+                    cur = ggml_get_rows(ctx0, cur, inp_cls);
+                    cb(cur, "result_embd_pooled", -1);
+                } break;
+            case LLAMA_POOLING_TYPE_UNSPECIFIED:
+                {
+                    GGML_ASSERT(false && "Invalid pooling type");
+                } break;
         }
-        cb(cur, "result_embd", -1);
 
         ggml_build_forward_expand(gf, cur);
 
@@ -7980,7 +8038,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
         ggml_backend_tensor_set(lctx.inp_pos, batch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos));
     }
 
-    {
+    if (hparams.causal_attn) {
         const int64_t n_kv     = kv_self.n;
         const int64_t n_tokens = batch.n_tokens;
 
@@ -7995,16 +8053,40 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
 
                 for (int i = 0; i < n_kv; ++i) {
                     float f;
-                    if (!lctx.kv_self.cells[i].has_seq_id(seq_id) ||
-                        (hparams.causal_attn && lctx.kv_self.cells[i].pos > pos)) {
+                    if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || lctx.kv_self.cells[i].pos > pos) {
                         f = -INFINITY;
                     } else {
-                        f = 0;
+                        f = 0.0f;
                     }
                     data[h*(n_kv*n_tokens) + j*n_kv + i] = f;
                 }
             }
         }
+    } else {
+        // non-causal attention attends only the tokens within the batch (i.e. the KV cache is not used)
+        const int64_t n_tokens = batch.n_tokens;
+
+        assert(ggml_backend_buffer_is_host(lctx.inp_KQ_mask->buffer));
+
+        float * data = (float *) lctx.inp_KQ_mask->data;
+
+        for (int h = 0; h < 1; ++h) {
+            for (int j = 0; j < n_tokens; ++j) {
+                const llama_seq_id seq_id = batch.seq_id[j][0];
+
+                for (int i = 0; i < n_tokens; ++i) {
+                    float f = -INFINITY;
+                    for (int s = 0; s < batch.n_seq_id[i]; ++s) {
+                        if (batch.seq_id[i][s] == seq_id) {
+                            f = 0.0f;
+                            break;
+                        }
+                    }
+
+                    data[h*(n_tokens*n_tokens) + j*n_tokens + i] = f;
+                }
+            }
+        }
     }
 
     if (hparams.need_kq_pos) {
@@ -8023,13 +8105,16 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
         const int64_t n_tokens = batch.n_tokens;
 
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_mean->buffer));
-        float * data = (float *) lctx.inp_mean->data;
 
+        float * data = (float *) lctx.inp_mean->data;
         memset(lctx.inp_mean->data, 0, n_tokens * n_tokens * ggml_element_size(lctx.inp_mean));
 
         std::vector<uint64_t> sum(n_tokens, 0);
         for (int i = 0; i < n_tokens; ++i) {
             const llama_seq_id seq_id = batch.seq_id[i][0];
+
+            GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN");
+
             sum[seq_id] += 1;
         }
 
@@ -8051,11 +8136,16 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
         const int64_t n_tokens = batch.n_tokens;
 
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer));
+
         uint32_t * data = (uint32_t *) lctx.inp_cls->data;
+        memset(lctx.inp_cls->data, 0, n_tokens * ggml_element_size(lctx.inp_cls));
 
         for (int i = 0; i < n_tokens; ++i) {
             const llama_seq_id seq_id = batch.seq_id[i][0];
-            const llama_pos pos = batch.pos[i];
+            const llama_pos    pos    = batch.pos[i];
+
+            GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS");
+
             if (pos == 0) {
                 data[seq_id] = i;
             }
@@ -8169,24 +8259,27 @@ static int llama_decode_internal(
         batch.seq_id = seq_id_arr.data();
     }
 
-    llama_kv_cache_update(&lctx);
+    // non-causal masks do not use the KV cache
+    if (hparams.causal_attn) {
+        llama_kv_cache_update(&lctx);
 
-    // if we have enough unused cells before the current head ->
-    //   better to start searching from the beginning of the cache, hoping to fill it
-    if (kv_self.head > kv_self.used + 2*n_tokens) {
-        kv_self.head = 0;
+        // if we have enough unused cells before the current head ->
+        //   better to start searching from the beginning of the cache, hoping to fill it
+        if (kv_self.head > kv_self.used + 2*n_tokens) {
+            kv_self.head = 0;
+        }
+
+        if (!llama_kv_cache_find_slot(kv_self, batch)) {
+            return 1;
+        }
+
+        // a heuristic, to avoid attending the full cache if it is not yet utilized
+        // after enough generations, the benefit from this heuristic disappears
+        // if we start defragmenting the cache, the benefit from this will be more important
+        kv_self.n = std::min(cparams.n_ctx, std::max(32u, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32)));
+        //kv_self.n = llama_kv_cache_cell_max(kv_self);
     }
 
-    if (!llama_kv_cache_find_slot(kv_self, batch)) {
-        return 1;
-    }
-
-    // a heuristic, to avoid attending the full cache if it is not yet utilized
-    // after enough generations, the benefit from this heuristic disappears
-    // if we start defragmenting the cache, the benefit from this will be more important
-    kv_self.n = std::min(cparams.n_ctx, std::max(32u, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32)));
-    //kv_self.n = llama_kv_cache_cell_max(kv_self);
-
     //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head);
 
     ggml_backend_sched_reset(lctx.sched);
@@ -8195,20 +8288,26 @@ static int llama_decode_internal(
     ggml_cgraph * gf = llama_build_graph(lctx, batch, false);
 
     // the output is always the last tensor in the graph
-    struct ggml_tensor * res        = gf->nodes[gf->n_nodes - 1];
-    struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2];
+    struct ggml_tensor * res  = gf->nodes[gf->n_nodes - 1];
+    struct ggml_tensor * embd = gf->nodes[gf->n_nodes - 2];
 
-    if (strcmp(res->name, "result_output") == 0) {
-        // the embeddings could be the second to last tensor, or the third to last tensor
-        if (strcmp(embeddings->name, "result_norm") != 0) {
-            embeddings = gf->nodes[gf->n_nodes - 3];
-            GGML_ASSERT(strcmp(embeddings->name, "result_norm") == 0);
-        }
-    } else if (strcmp(res->name, "result_embd") == 0) {
-        embeddings = res;
-        res = nullptr;
+    if (!hparams.causal_attn) {
+        res = nullptr; // do not extract logits for embedding models such as BERT
+
+        // token or sequence embeddings
+        embd = gf->nodes[gf->n_nodes - 1];
+
+        GGML_ASSERT(strcmp(embd->name, "result_embd") == 0 || strcmp(embd->name, "result_embd_pooled") == 0);
     } else {
-        GGML_ASSERT(false);
+        if (strcmp(res->name, "result_output") == 0) {
+            // the token embeddings could be the second to last tensor, or the third to last tensor
+            if (strcmp(embd->name, "result_norm") != 0) {
+                embd = gf->nodes[gf->n_nodes - 3];
+                GGML_ASSERT(strcmp(embd->name, "result_norm") == 0);
+            }
+        } else {
+            GGML_ASSERT(false && "missing result_output tensor");
+        }
     }
 
     // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
@@ -8275,46 +8374,82 @@ static int llama_decode_internal(
         logits_out.clear();
 #endif
 
-        ggml_backend_t res_backend = ggml_backend_sched_get_node_backend(lctx.sched, res);
-        GGML_ASSERT(res_backend != nullptr);
+        ggml_backend_t backend_res = ggml_backend_sched_get_node_backend(lctx.sched, res);
+        GGML_ASSERT(backend_res != nullptr);
+
         if (batch.logits) {
             logits_out.resize(n_vocab * n_tokens);
             for (uint32_t i = 0; i < n_tokens; i++) {
                 if (batch.logits[i] == 0) {
                     continue;
                 }
-                ggml_backend_tensor_get_async(res_backend, res, logits_out.data() + (n_vocab*i), (n_vocab*i)*sizeof(float), n_vocab*sizeof(float));
+                ggml_backend_tensor_get_async(backend_res, res, logits_out.data() + (n_vocab*i), (n_vocab*i)*sizeof(float), n_vocab*sizeof(float));
 #ifndef NDEBUG
                 logits_valid[i] = true;
 #endif
             }
         } else if (lctx.logits_all) {
             logits_out.resize(n_vocab * n_tokens);
-            ggml_backend_tensor_get_async(res_backend, res, logits_out.data(), 0, n_vocab*n_tokens*sizeof(float));
+            ggml_backend_tensor_get_async(backend_res, res, logits_out.data(), 0, n_vocab*n_tokens*sizeof(float));
 #ifndef NDEBUG
             std::fill(logits_valid.begin(), logits_valid.end(), true);
 #endif
         } else {
             logits_out.resize(n_vocab);
-            ggml_backend_tensor_get_async(res_backend, res, logits_out.data(), (n_vocab*(n_tokens - 1))*sizeof(float), n_vocab*sizeof(float));
+            ggml_backend_tensor_get_async(backend_res, res, logits_out.data(), (n_vocab*(n_tokens - 1))*sizeof(float), n_vocab*sizeof(float));
 #ifndef NDEBUG
             logits_valid[0] = true;
 #endif
         }
-        ggml_backend_synchronize(res_backend);
+        ggml_backend_synchronize(backend_res);
     }
 
     // extract embeddings
-    if (!lctx.embedding.empty()) {
-        auto & embedding_out = lctx.embedding;
+    if (cparams.embeddings && embd) {
+        ggml_backend_t backend_embd = ggml_backend_sched_get_node_backend(lctx.sched, embd);
+        GGML_ASSERT(backend_embd != nullptr);
 
-        const int64_t embd_pos  = res ? n_embd * (n_tokens-1) : 0;
-        const int64_t embd_size = res ? n_embd : n_embd * n_tokens;
+        switch (cparams.pooling_type) {
+            case LLAMA_POOLING_TYPE_NONE:
+                {
+                    // extract token embeddings
+                    auto & embd_out = lctx.embd;
 
-        embedding_out.resize(embd_size);
-        ggml_backend_t embeddings_backend = ggml_backend_sched_get_node_backend(lctx.sched, embeddings);
-        ggml_backend_tensor_get_async(embeddings_backend, embeddings, embedding_out.data(), embd_pos*sizeof(float), embd_size*sizeof(float));
-        ggml_backend_synchronize(embeddings_backend);
+                    if (batch.logits) {
+                        embd_out.resize(n_embd * n_tokens);
+                        for (uint32_t i = 0; i < n_tokens; i++) {
+                            if (batch.logits[i] == 0) {
+                                continue;
+                            }
+
+                            ggml_backend_tensor_get_async(backend_embd, embd, embd_out.data() + (n_embd*i), (n_embd*i)*sizeof(float), n_embd*sizeof(float));
+                        }
+                    }
+                } break;
+            case LLAMA_POOLING_TYPE_CLS:
+            case LLAMA_POOLING_TYPE_MEAN:
+                {
+                    GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0);
+
+                    // extract sequence embeddings
+                    auto & embd_seq_out = lctx.embd_seq;
+                    embd_seq_out.clear();
+
+                    for (uint32_t i = 0; i < n_tokens; i++) {
+                        const llama_seq_id seq_id = batch.seq_id[i][0];
+                        if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
+                            continue;
+                        }
+                        embd_seq_out[seq_id].resize(n_embd);
+                        ggml_backend_tensor_get_async(backend_embd, embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float));
+                    }
+                } break;
+            case LLAMA_POOLING_TYPE_UNSPECIFIED:
+                {
+                    GGML_ASSERT(false && "unknown pooling type");
+                } break;
+        }
+        ggml_backend_synchronize(backend_embd);
     }
 
     // measure the performance only for the single-token evals
@@ -8608,19 +8743,19 @@ static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) {
     GGML_ASSERT(llama_is_byte_token(vocab, id));
     const auto& token_data = vocab.id_to_token.at(id);
     switch (llama_vocab_get_type(vocab)) {
-    case LLAMA_VOCAB_TYPE_SPM: {
-        auto buf = token_data.text.substr(3, 2);
-        return strtol(buf.c_str(), NULL, 16);
-    }
-    case LLAMA_VOCAB_TYPE_BPE: {
-        GGML_ASSERT(false);
-        return unicode_to_bytes_bpe(token_data.text);
-    }
-    case LLAMA_VOCAB_TYPE_WPM: {
-        GGML_ASSERT(false);
-    }
-    default:
-        GGML_ASSERT(false);
+        case LLAMA_VOCAB_TYPE_SPM: {
+            auto buf = token_data.text.substr(3, 2);
+            return strtol(buf.c_str(), NULL, 16);
+        }
+        case LLAMA_VOCAB_TYPE_BPE: {
+            GGML_ASSERT(false);
+            return unicode_to_bytes_bpe(token_data.text);
+        }
+        case LLAMA_VOCAB_TYPE_WPM: {
+            GGML_ASSERT(false);
+        }
+        default:
+            GGML_ASSERT(false);
     }
 }
 
@@ -11864,7 +11999,7 @@ struct llama_context_params llama_context_default_params() {
         /*.type_k                      =*/ GGML_TYPE_F16,
         /*.type_v                      =*/ GGML_TYPE_F16,
         /*.logits_all                  =*/ false,
-        /*.embedding                   =*/ false,
+        /*.embeddings                  =*/ false,
         /*.offload_kqv                 =*/ true,
         /*.abort_callback              =*/ nullptr,
         /*.abort_callback_data         =*/ nullptr,
@@ -12015,6 +12150,7 @@ struct llama_context * llama_new_context_with_model(
     cparams.yarn_beta_fast   = params.yarn_beta_fast;
     cparams.yarn_beta_slow   = params.yarn_beta_slow;
     cparams.defrag_thold     = params.defrag_thold;
+    cparams.embeddings       = params.embeddings;
     cparams.offload_kqv      = params.offload_kqv;
     cparams.pooling_type     = params.pooling_type;
 
@@ -12132,13 +12268,13 @@ struct llama_context * llama_new_context_with_model(
                 ctx->backends.push_back(backend);
             } else {
                 // LLAMA_SPLIT_LAYER requires a backend for each GPU
-                int id_list[GGML_SYCL_MAX_DEVICES];
-                ggml_sycl_get_gpu_list(id_list, GGML_SYCL_MAX_DEVICES);
+
                 for (int i = 0; i < ggml_backend_sycl_get_device_count(); ++i) {
-                    int device_id = id_list[i];
                     ggml_backend_t backend = ggml_backend_sycl_init(i);
                     if (backend == nullptr) {
-                        LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d)backend\n", __func__, device_id, i);
+                        int id_list[GGML_SYCL_MAX_DEVICES];
+                        ggml_sycl_get_gpu_list(id_list, GGML_SYCL_MAX_DEVICES);
+                        LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d)backend\n", __func__, id_list[i], i);
                         llama_free(ctx);
                         return nullptr;
                     }
@@ -12192,8 +12328,8 @@ struct llama_context * llama_new_context_with_model(
         // resized during inference, reserve maximum
         ctx->logits.reserve(hparams.n_vocab*cparams.n_batch);
 
-        if (params.embedding) {
-            ctx->embedding.resize(hparams.n_embd);
+        if (params.embeddings) {
+            ctx->embd.reserve(hparams.n_embd*cparams.n_batch);
         }
 
         // graph inputs
@@ -12628,7 +12764,7 @@ size_t llama_get_state_size(const struct llama_context * ctx) {
     // assume worst case for logits although only currently set ones are serialized
     const size_t s_logits          = ctx->logits.capacity() * sizeof(float);
     const size_t s_embedding_size  = sizeof(size_t);
-    const size_t s_embedding       = ctx->embedding.size() * sizeof(float);
+    const size_t s_embedding       = ctx->embd.capacity() * sizeof(float);
     const size_t s_kv_buf_size     = sizeof(size_t);
     const size_t s_kv_head         = sizeof(uint32_t);
     const size_t s_kv_size         = sizeof(uint32_t);
@@ -12737,12 +12873,12 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
 
     // copy embeddings
     {
-        const size_t embedding_size = ctx->embedding.size();
+        const size_t embeddings_size = ctx->embd.size();
 
-        data_ctx->write(&embedding_size, sizeof(embedding_size));
+        data_ctx->write(&embeddings_size, sizeof(embeddings_size));
 
-        if (embedding_size) {
-            data_ctx->write(ctx->embedding.data(), embedding_size * sizeof(float));
+        if (embeddings_size) {
+            data_ctx->write(ctx->embd.data(), embeddings_size * sizeof(float));
         }
     }
 
@@ -12846,15 +12982,17 @@ size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) {
 
     // set embeddings
     {
-        size_t embedding_size;
+        size_t embeddings_size;
 
-        memcpy(&embedding_size, inp, sizeof(embedding_size)); inp += sizeof(embedding_size);
+        memcpy(&embeddings_size, inp, sizeof(embeddings_size)); inp += sizeof(embeddings_size);
 
-        GGML_ASSERT(ctx->embedding.capacity() == embedding_size);
+        GGML_ASSERT(ctx->embd.capacity() == embeddings_size);
 
-        if (embedding_size) {
-            memcpy(ctx->embedding.data(), inp, embedding_size * sizeof(float));
-            inp += embedding_size * sizeof(float);
+        if (embeddings_size) {
+            ctx->embd.resize(embeddings_size);
+
+            memcpy(ctx->embd.data(), inp, embeddings_size * sizeof(float));
+            inp += embeddings_size * sizeof(float);
         }
     }
 
@@ -13104,11 +13242,20 @@ float * llama_get_logits_ith(struct llama_context * ctx, int32_t i) {
 }
 
 float * llama_get_embeddings(struct llama_context * ctx) {
-    return ctx->embedding.data();
+    return ctx->embd.data();
 }
 
 float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i) {
-    return ctx->embedding.data() + i*ctx->model.hparams.n_embd;
+    return ctx->embd.data() + i*ctx->model.hparams.n_embd;
+}
+
+float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id) {
+    auto it = ctx->embd_seq.find(seq_id);
+    if (it == ctx->embd_seq.end()) {
+        return nullptr;
+    }
+
+    return it->second.data();
 }
 
 const char * llama_token_get_text(const struct llama_model * model, llama_token token) {
@@ -13394,18 +13541,22 @@ LLAMA_API int32_t llama_chat_apply_template(
             curr_tmpl = std::string(model_template.data(), model_template.size());
         }
     }
+
     // format the chat to string
     std::vector<const llama_chat_message *> chat_vec;
     chat_vec.resize(n_msg);
     for (size_t i = 0; i < n_msg; i++) {
         chat_vec[i] = &chat[i];
     }
+
     std::string formatted_chat;
     int32_t res = llama_chat_apply_template_internal(curr_tmpl, chat_vec, formatted_chat, add_ass);
     if (res < 0) {
         return res;
     }
-    strncpy(buf, formatted_chat.c_str(), length);
+    if (buf && length > 0) {
+        strncpy(buf, formatted_chat.c_str(), length);
+    }
     return res;
 }
 

llama.h

@@ -163,7 +163,7 @@ extern "C" {
     // - embd   : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
     // - pos    : the positions of the respective token in the sequence
     // - seq_id : the sequence to which the respective token belongs
-    // - logits : if zero, the logits for the respective token will not be output
+    // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output
     //
     typedef struct llama_batch {
         int32_t n_tokens;
@@ -173,7 +173,7 @@ extern "C" {
         llama_pos    *  pos;
         int32_t      *  n_seq_id;
         llama_seq_id ** seq_id;
-        int8_t       *  logits;
+        int8_t       *  logits; // TODO: rename this to "output"
 
         // NOTE: helpers for smooth API transition - can be deprecated in the future
         //       for future-proof code, use the above fields instead and ignore everything below
@@ -260,7 +260,7 @@ extern "C" {
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
         bool logits_all;  // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
-        bool embedding;   // embedding mode only
+        bool embeddings;  // if true, extract embeddings (together with logits)
         bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
 
         // Abort callback
@@ -655,14 +655,20 @@ extern "C" {
     // llama_get_logits(ctx) + i*n_vocab
     LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i);
 
-    // Get the embeddings for the input
-    // shape: [n_embd] (1-dimensional)
+    // Get all output token embeddings
+    // shape: [n_tokens*n_embd] (1-dimensional)
     LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
 
-    // Get the embeddings for the ith sequence
+    // Get the embeddings for the ith token
     // llama_get_embeddings(ctx) + i*n_embd
+    // shape: [n_embd] (1-dimensional)
     LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i);
 
+    // Get the embeddings for a sequence id
+    // Returns NULL if pooling_type is LLAMA_POOLING_TYPE_NONE
+    // shape: [n_embd] (1-dimensional)
+    LLAMA_API float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id);
+
     //
     // Vocab
     //

scripts/compare-llama-bench.py

@@ -18,7 +18,7 @@ except ImportError as e:
 KEY_PROPERTIES = [
     "cpu_info", "gpu_info", "n_gpu_layers", "main_gpu", "cuda", "opencl", "metal", "gpu_blas",
     "blas", "model_filename", "model_type", "model_size", "model_n_params", "n_batch", "n_threads",
-    "type_k", "type_v", "no_kv_offload", "mul_mat_q", "tensor_split", "n_prompt", "n_gen"
+    "type_k", "type_v", "no_kv_offload", "tensor_split", "n_prompt", "n_gen"
 ]
 
 # Properties that are boolean and are converted to Yes/No for the table:

scripts/sync-ggml.last

@@ -1 +1 @@
-274680868e12427373bab4bec87554431b954704
+8695910a39102609073d0e099aa7c97d6bcb3bf9