Merge remote-tracking branch 'origin/master' into json-type

.devops/llama-server-cuda.Dockerfile

@@ -30,8 +30,10 @@ RUN make -j$(nproc) llama-server
 FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
 
 RUN apt-get update && \
-    apt-get install -y libcurl4-openssl-dev libgomp1
+    apt-get install -y libcurl4-openssl-dev libgomp1 curl
 
 COPY --from=build /app/llama-server /llama-server
 
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
 ENTRYPOINT [ "/llama-server" ]

.devops/llama-server-intel.Dockerfile

@@ -20,10 +20,12 @@ RUN if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \
 FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
 
 RUN apt-get update && \
-    apt-get install -y libcurl4-openssl-dev
+    apt-get install -y libcurl4-openssl-dev curl
 
 COPY --from=build /app/build/bin/llama-server /llama-server
 
 ENV LC_ALL=C.utf8
 
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
 ENTRYPOINT [ "/llama-server" ]

.devops/llama-server-rocm.Dockerfile

@@ -43,8 +43,10 @@ ENV CXX=/opt/rocm/llvm/bin/clang++
 # Enable cURL
 ENV LLAMA_CURL=1
 RUN apt-get update && \
-    apt-get install -y libcurl4-openssl-dev
+    apt-get install -y libcurl4-openssl-dev curl
 
 RUN make -j$(nproc) llama-server
 
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
 ENTRYPOINT [ "/app/llama-server" ]

.devops/llama-server-vulkan.Dockerfile

@@ -5,15 +5,11 @@ FROM ubuntu:$UBUNTU_VERSION as build
 # Install build tools
 RUN apt update && apt install -y git build-essential cmake wget
 
-# Install Vulkan SDK
+# Install Vulkan SDK and cURL
 RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
     wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list && \
     apt update -y && \
-    apt-get install -y vulkan-sdk
+    apt-get install -y vulkan-sdk libcurl4-openssl-dev curl
-
-# Install cURL
-RUN apt-get update && \
-    apt-get install -y libcurl4-openssl-dev
 
 # Build it
 WORKDIR /app
@@ -28,4 +24,6 @@ RUN cp /app/build/bin/llama-server /llama-server && \
 
 ENV LC_ALL=C.utf8
 
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
 ENTRYPOINT [ "/llama-server" ]

.devops/llama-server.Dockerfile

@@ -3,7 +3,7 @@ ARG UBUNTU_VERSION=22.04
 FROM ubuntu:$UBUNTU_VERSION as build
 
 RUN apt-get update && \
-    apt-get install -y build-essential git libcurl4-openssl-dev
+    apt-get install -y build-essential git libcurl4-openssl-dev curl
 
 WORKDIR /app
 
@@ -22,4 +22,6 @@ COPY --from=build /app/llama-server /llama-server
 
 ENV LC_ALL=C.utf8
 
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
 ENTRYPOINT [ "/llama-server" ]

.github/workflows/docker.yml

@@ -10,7 +10,7 @@
 name: Publish Docker image
 
 on:
-  pull_request:
+  #pull_request:
   push:
     branches:
       - master
@@ -22,7 +22,7 @@ concurrency:
 jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub
-    if: github.event.pull_request.draft == false
+    #if: github.event.pull_request.draft == false
 
     runs-on: ubuntu-latest
     env:

common/common.cpp

@@ -1263,11 +1263,6 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         return true;
     }
     // cvector params
-    if (arg == "--completions-file") {
-        CHECK_ARG
-        params.cvector_completions_file = argv[i];
-        return true;
-    }
     if (arg == "--positive-file") {
         CHECK_ARG
         params.cvector_positive_file = argv[i];
@@ -1278,11 +1273,6 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.cvector_negative_file = argv[i];
         return true;
     }
-    if (arg == "--completions") {
-        CHECK_ARG
-        params.n_completions = std::stoi(argv[i]);
-        return true;
-    }
     if (arg == "--pca-batch") {
         CHECK_ARG
         params.n_pca_batch = std::stoi(argv[i]);
@@ -1293,6 +1283,14 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.n_pca_iterations = std::stoi(argv[i]);
         return true;
     }
+    if (arg == "--method") {
+        CHECK_ARG
+        std::string value(argv[i]);
+        /**/ if (value == "pca") { params.cvector_dimre_method = DIMRE_METHOD_PCA; }
+        else if (value == "mean") { params.cvector_dimre_method = DIMRE_METHOD_MEAN; }
+        else { invalid_param = true; }
+        return true;
+    }
 #ifndef LOG_DISABLE_LOGS
     // Parse args for logging parameters
     if (log_param_single_parse(argv[i])) {
@@ -1444,7 +1442,10 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     options.push_back({ "main",        "       --cfg-negative-prompt-file FNAME",
                                                                         "negative prompt file to use for guidance" });
     options.push_back({ "main",        "       --cfg-scale N",          "strength of guidance (default: %.1f, 1.0 = disable)", (double)sparams.cfg_scale });
-
+    options.push_back({ "main",        "       --chat-template JINJA_TEMPLATE",
+                                                                        "set custom jinja chat template (default: template taken from model's metadata)\n"
+                                                                        "only commonly used templates are accepted:\n"
+                                                                        "https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template" });
     options.push_back({ "grammar" });
     options.push_back({ "*",           "       --grammar GRAMMAR",      "BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '%s')", sparams.grammar.c_str() });
     options.push_back({ "*",           "       --grammar-file FNAME",   "file to read grammar from" });
@@ -1538,9 +1539,11 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     options.push_back({ "*",           "       --lora FNAME",           "apply LoRA adapter (implies --no-mmap)" });
     options.push_back({ "*",           "       --lora-scaled FNAME S",  "apply LoRA adapter with user defined scaling S (implies --no-mmap)" });
     options.push_back({ "*",           "       --lora-base FNAME",      "optional model to use as a base for the layers modified by the LoRA adapter" });
-    options.push_back({ "*",           "       --control-vector FNAME", "add a control vector" });
+    options.push_back({ "*",           "       --control-vector FNAME", "add a control vector\n"
+                                                                        "note: this argument can be repeated to add multiple control vectors" });
     options.push_back({ "*",           "       --control-vector-scaled FNAME SCALE",
-                                                                        "add a control vector with user defined scaling SCALE" });
+                                                                        "add a control vector with user defined scaling SCALE\n"
+                                                                        "note: this argument can be repeated to add multiple scaled control vectors" });
     options.push_back({ "*",           "       --control-vector-layer-range START END",
                                                                         "layer range to apply the control vector(s) to, start and end inclusive" });
     options.push_back({ "*",           "-m,    --model FNAME",          "model path (default: models/$filename with filename from --hf-file\n"
@@ -1621,11 +1624,9 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     options.push_back({ "cvector",     "-o,    --output FNAME",         "output file (default: '%s')", params.cvector_outfile.c_str() });
     options.push_back({ "cvector",     "       --positive-file FNAME",  "positive prompts file, one prompt per line (default: '%s')", params.cvector_positive_file.c_str() });
     options.push_back({ "cvector",     "       --negative-file FNAME",  "negative prompts file, one prompt per line (default: '%s')", params.cvector_negative_file.c_str() });
-    options.push_back({ "cvector",     "       --completions-file FNAME",
-                                                                        "completions file (default: '%s')", params.cvector_completions_file.c_str() });
-    options.push_back({ "cvector",     "       --completions N",        "number of lines of completions file to use (default: %d)", params.n_completions });
     options.push_back({ "cvector",     "       --pca-batch N",          "batch size used for PCA. Larger batch runs faster, but uses more memory (default: %d)", params.n_pca_batch });
     options.push_back({ "cvector",     "       --pca-iter N",           "number of iterations used for PCA (default: %d)", params.n_pca_iterations });
+    options.push_back({ "cvector",     "       --method {pca,mean}",    "dimensionality reduction method to be used (default: pca)" });
 
     printf("usage: %s [options]\n", argv[0]);
 
@@ -2602,12 +2603,67 @@ bool llama_should_add_bos_token(const llama_model * model) {
     return add_bos != -1 ? bool(add_bos) : (llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM);
 }
 
+//
+// Chat template utils
+//
+
 bool llama_chat_verify_template(const std::string & tmpl) {
     llama_chat_message chat[] = {{"user", "test"}};
     int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, nullptr, 0);
     return res >= 0;
 }
 
+std::string llama_chat_apply_template(const struct llama_model * model,
+        const std::string & tmpl,
+        const std::vector<llama_chat_msg> & msgs,
+        bool add_ass) {
+    int alloc_size = 0;
+    std::vector<llama_chat_message> chat;
+    for (auto & msg : msgs) {
+        chat.push_back({msg.role.c_str(), msg.content.c_str()});
+        alloc_size += (msg.role.size() + msg.content.size()) * 1.25;
+    }
+
+    const char * ptr_tmpl = tmpl.empty() ? nullptr : tmpl.c_str();
+    std::vector<char> buf(alloc_size);
+
+    // run the first time to get the total output length
+    int32_t res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), add_ass, buf.data(), buf.size());
+
+    // if it turns out that our buffer is too small, we resize it
+    if ((size_t) res > buf.size()) {
+        buf.resize(res);
+        res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), add_ass, buf.data(), buf.size());
+    }
+
+    std::string formatted_chat(buf.data(), res);
+    return formatted_chat;
+}
+
+std::string llama_chat_format_single(const struct llama_model * model,
+        const std::string & tmpl,
+        const std::vector<llama_chat_msg> & past_msg,
+        const llama_chat_msg & new_msg,
+        bool add_ass) {
+    auto fmt_past_msg = llama_chat_apply_template(model, tmpl, past_msg, false);
+    std::vector<llama_chat_msg> chat_new(past_msg);
+    chat_new.push_back(new_msg);
+    auto fmt_new_msg = llama_chat_apply_template(model, tmpl, chat_new, add_ass);
+    auto formatted = fmt_new_msg.substr(fmt_past_msg.size(), fmt_new_msg.size() - fmt_past_msg.size());
+    return formatted;
+}
+
+std::string llama_chat_format_example(const struct llama_model * model,
+        const std::string & tmpl) {
+    std::vector<llama_chat_msg> msgs = {
+        {"system",    "You are a helpful assistant"},
+        {"user",      "Hello"},
+        {"assistant", "Hi there"},
+        {"user",      "How are you?"},
+    };
+    return llama_chat_apply_template(model, tmpl, msgs, true);
+}
+
 //
 // KV cache utils
 //

common/common.h

@@ -52,6 +52,12 @@ int32_t cpu_get_num_math();
 // CLI argument parsing
 //
 
+// dimensionality reduction methods, used by cvector-generator
+enum dimre_method {
+    DIMRE_METHOD_PCA,
+    DIMRE_METHOD_MEAN,
+};
+
 struct gpt_params {
     uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
 
@@ -238,13 +244,12 @@ struct gpt_params {
     bool compute_ppl    = true;  // whether to compute perplexity
 
     // cvector-generator params
-    int n_completions = 64;
+    int n_pca_batch = 100;
-    int n_pca_batch = 20;
     int n_pca_iterations = 1000;
-    std::string cvector_outfile          = "control_vector.gguf";
+    dimre_method cvector_dimre_method = DIMRE_METHOD_PCA;
-    std::string cvector_completions_file = "examples/cvector-generator/completions.txt";
+    std::string cvector_outfile       = "control_vector.gguf";
-    std::string cvector_positive_file    = "examples/cvector-generator/positive.txt";
+    std::string cvector_positive_file = "examples/cvector-generator/positive.txt";
-    std::string cvector_negative_file    = "examples/cvector-generator/negative.txt";
+    std::string cvector_negative_file = "examples/cvector-generator/negative.txt";
 };
 
 void gpt_params_handle_model_default(gpt_params & params);
@@ -365,9 +370,32 @@ bool llama_should_add_bos_token(const llama_model * model);
 // Chat template utils
 //
 
+// same with llama_chat_message, but uses std::string
+struct llama_chat_msg {
+    std::string role;
+    std::string content;
+};
+
 // Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
 bool llama_chat_verify_template(const std::string & tmpl);
 
+// CPP wrapper for llama_chat_apply_template
+std::string llama_chat_apply_template(const struct llama_model * model,
+        const std::string & tmpl,
+        const std::vector<llama_chat_msg> & chat,
+        bool add_ass);
+
+// Format single message, while taking into account the position of that message in chat history
+std::string llama_chat_format_single(const struct llama_model * model,
+        const std::string & tmpl,
+        const std::vector<llama_chat_msg> & past_msg,
+        const llama_chat_msg & new_msg,
+        bool add_ass);
+
+// Returns an example of formatted chat
+std::string llama_chat_format_example(const struct llama_model * model,
+        const std::string & tmpl);
+
 //
 // KV cache utils
 //

common/json-schema-to-grammar.cpp

@@ -40,6 +40,233 @@ static std::string build_repetition(const std::string & item_rule, int min_items
     return result;
 }
 
+/* Minimalistic replacement for std::string_view, which is only available from C++17 onwards */
+class string_view {
+    const std::string & _str;
+    const size_t _start;
+    const size_t _end;
+public:
+    string_view(const std::string & str, size_t start = 0, size_t end  = std::string::npos) : _str(str), _start(start), _end(end == std::string::npos ? str.length() : end) {}
+
+    size_t size() const {
+        return _end - _start;
+    }
+
+    size_t length() const {
+        return size();
+    }
+
+    operator std::string() const {
+        return str();
+    }
+
+    std::string str() const {
+        return _str.substr(_start, _end - _start);
+    }
+
+    string_view substr(size_t pos, size_t len = std::string::npos) const {
+        return string_view(_str, _start + pos, len == std::string::npos ? _end : _start + pos + len);
+    }
+
+    char operator[](size_t pos) const {
+        auto index = _start + pos;
+        if (index >= _end) {
+            throw std::out_of_range("string_view index out of range");
+        }
+        return _str[_start + pos];
+    }
+
+    bool operator==(const string_view & other) const {
+        std::string this_str = *this;
+        std::string other_str = other;
+        return this_str == other_str;
+    }
+};
+
+static void _build_min_max_int(int min_value, int max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
+    auto has_min = min_value != std::numeric_limits<int>::min();
+    auto has_max = max_value != std::numeric_limits<int>::max();
+
+    auto digit_range = [&](char from, char to) {
+        out << "[";
+        if (from == to) {
+            out << from;
+        } else {
+            out << from << "-" << to;
+        }
+        out << "]";
+    };
+    auto more_digits = [&](int min_digits, int max_digits) {
+        out << "[0-9]";
+        if (min_digits == max_digits && min_digits == 1) {
+            return;
+        }
+        out << "{";
+        out << min_digits;
+        if (max_digits != min_digits) {
+            out << ",";
+            if (max_digits != std::numeric_limits<int>::max()) {
+                out << max_digits;
+            }
+        }
+        out << "}";
+    };
+    std::function<void(const string_view &, const string_view &)> uniform_range =
+        [&](const string_view & from, const string_view & to) {
+            size_t i = 0;
+            while (i < from.length() && i < to.length() && from[i] == to[i]) {
+                i++;
+            }
+            if (i > 0) {
+                out << "\"" << from.substr(0, i).str() << "\"";
+            }
+            if (i < from.length() && i < to.length()) {
+                if (i > 0) {
+                    out << " ";
+                }
+                auto sub_len = from.length() - i - 1;
+                if (sub_len > 0) {
+                    auto from_sub = from.substr(i + 1);
+                    auto to_sub = to.substr(i + 1);
+                    auto sub_zeros = repeat("0", sub_len);
+                    auto sub_nines = repeat("9", sub_len);
+
+                    auto to_reached = false;
+                    out << "(";
+                    if (from_sub == sub_zeros) {
+                        digit_range(from[i], to[i] - 1);
+                        out << " ";
+                        more_digits(sub_len, sub_len);
+                    } else {
+                        out << "[" << from[i] << "] ";
+                        out << "(";
+                        uniform_range(from_sub, sub_nines);
+                        out << ")";
+                        if (from[i] < to[i] - 1) {
+                            out << " | ";
+                            if (to_sub == sub_nines) {
+                                digit_range(from[i] + 1, to[i]);
+                                to_reached = true;
+                            } else {
+                                digit_range(from[i] + 1, to[i] - 1);
+                            }
+                            out << " ";
+                            more_digits(sub_len, sub_len);
+                        }
+                    }
+                    if (!to_reached) {
+                        out << " | ";
+                        digit_range(to[i], to[i]);
+                        out << " ";
+                        uniform_range(sub_zeros, to_sub);
+                    }
+                    out << ")";
+                } else {
+                    out << "[" << from[i] << "-" << to[i] << "]";
+                }
+            }
+        };
+
+    if (has_min && has_max) {
+        if (min_value < 0 && max_value < 0) {
+            out << "\"-\" (";
+            _build_min_max_int(-max_value, -min_value, out, decimals_left, /* top_level= */ true);
+            out << ")";
+            return;
+        }
+
+        if (min_value < 0) {
+            out << "\"-\" (";
+            _build_min_max_int(0, -min_value, out, decimals_left, /* top_level= */ true);
+            out << ") | ";
+            min_value = 0;
+        }
+
+        auto min_s = std::to_string(min_value);
+        auto max_s = std::to_string(max_value);
+        auto min_digits = min_s.length();
+        auto max_digits = max_s.length();
+
+        for (auto digits = min_digits; digits < max_digits; digits++) {
+            uniform_range(min_s, repeat("9", digits));
+            min_s = "1" + repeat("0", digits);
+            out << " | ";
+        }
+        uniform_range(min_s, max_s);
+        return;
+    }
+
+    auto less_decimals = std::max(decimals_left - 1, 1);
+
+    if (has_min) {
+        if (min_value < 0) {
+            out << "\"-\" (";
+            _build_min_max_int(std::numeric_limits<int>::min(), -min_value, out, decimals_left, /* top_level= */ false);
+            out << ") | [0] | [1-9] ";
+            more_digits(0, decimals_left - 1);
+        } else if (min_value == 0) {
+            if (top_level) {
+                out << "[0] | [1-9] ";
+                more_digits(0, less_decimals);
+            } else {
+                more_digits(1, decimals_left);
+            }
+        } else if (min_value <= 9) {
+            char c = '0' + min_value;
+            auto range_start = top_level ? '1' : '0';
+            if (c > range_start) {
+                digit_range(range_start, c - 1);
+                out << " ";
+                more_digits(1, less_decimals);
+                out << " | ";
+            }
+            digit_range(c, '9');
+            out << " ";
+            more_digits(0, less_decimals);
+        } else {
+            auto min_s = std::to_string(min_value);
+            auto len = min_s.length();
+            auto c = min_s[0];
+
+            if (c > '1') {
+                digit_range(top_level ? '1' : '0', c - 1);
+                out << " ";
+                more_digits(len, less_decimals);
+                out << " | ";
+            }
+            digit_range(c, c);
+            out << " (";
+            _build_min_max_int(std::stoi(min_s.substr(1)), std::numeric_limits<int>::max(), out, less_decimals, /* top_level= */ false);
+            out << ")";
+            if (c < '9') {
+                out << " | ";
+                digit_range(c + 1, '9');
+                out << " ";
+                more_digits(len - 1, less_decimals);
+            }
+        }
+        return;
+    }
+
+    if (has_max) {
+        if (max_value >= 0) {
+            if (top_level) {
+                out << "\"-\" [1-9] ";
+                more_digits(0, less_decimals);
+                out << " | ";
+            }
+            _build_min_max_int(0, max_value, out, decimals_left, /* top_level= */ true);
+        } else {
+            out << "\"-\" (";
+            _build_min_max_int(-max_value, std::numeric_limits<int>::max(), out, decimals_left, /* top_level= */ false);
+            out << ")";
+        }
+        return;
+    }
+
+    throw std::runtime_error("At least one of min_value or max_value must be set");
+}
+
 const std::string SPACE_RULE = "| \" \" | \"\\n\" [ \\t]{0,20}";
 
 struct BuiltinRule {
@@ -160,7 +387,6 @@ static std::string format_literal(const std::string & literal) {
     return "\"" + escaped + "\"";
 }
 
-
 class SchemaConverter {
 private:
     std::function<json(const std::string &)> _fetch_json;
@@ -688,6 +914,24 @@ public:
             int min_len = schema.contains("minLength") ? schema["minLength"].get<int>() : 0;
             int max_len = schema.contains("maxLength") ? schema["maxLength"].get<int>() : std::numeric_limits<int>::max();
             return _add_rule(rule_name, "\"\\\"\" " + build_repetition(char_rule, min_len, max_len) + " \"\\\"\" space");
+        } else if (schema_type == "integer" && (schema.contains("minimum") || schema.contains("exclusiveMinimum") || schema.contains("maximum") || schema.contains("exclusiveMaximum"))) {
+            int min_value = std::numeric_limits<int>::min();
+            int max_value = std::numeric_limits<int>::max();
+            if (schema.contains("minimum")) {
+                min_value = schema["minimum"].get<int>();
+            } else if (schema.contains("exclusiveMinimum")) {
+                min_value = schema["exclusiveMinimum"].get<int>() + 1;
+            }
+            if (schema.contains("maximum")) {
+                max_value = schema["maximum"].get<int>();
+            } else if (schema.contains("exclusiveMaximum")) {
+                max_value = schema["exclusiveMaximum"].get<int>() - 1;
+            }
+            std::stringstream out;
+            out << "(";
+            _build_min_max_int(min_value, max_value, out);
+            out << ") space";
+            return _add_rule(rule_name, out.str());
         } else if (schema.empty() || schema_type == "object") {
             return _add_rule(rule_name, _add_primitive("object", PRIMITIVE_RULES.at("object")));
         } else {

common/sampling.cpp

@@ -28,9 +28,13 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
 
         std::vector<const llama_grammar_element *> grammar_rules(result->parsed_grammar.c_rules());
 
-        result->grammar = llama_grammar_init(
+        struct llama_grammar * grammar = llama_grammar_init(
                 grammar_rules.data(),
                 grammar_rules.size(), result->parsed_grammar.symbol_ids.at("root"));
+        if (grammar == nullptr) {
+            throw std::runtime_error("Failed to initialize llama_grammar");
+        }
+        result->grammar = grammar;
     }
 
     result->prev.resize(params.n_prev);
@@ -59,9 +63,13 @@ void llama_sampling_reset(llama_sampling_context * ctx) {
     if (!ctx->parsed_grammar.rules.empty()) {
         std::vector<const llama_grammar_element *> grammar_rules(ctx->parsed_grammar.c_rules());
 
-        ctx->grammar = llama_grammar_init(
+        struct llama_grammar * grammar = llama_grammar_init(
                 grammar_rules.data(),
                 grammar_rules.size(), ctx->parsed_grammar.symbol_ids.at("root"));
+        if (grammar == nullptr) {
+            throw std::runtime_error("Failed to initialize llama_grammar");
+        }
+        ctx->grammar = grammar;
     }
 
     std::fill(ctx->prev.begin(), ctx->prev.end(), 0);

examples/cvector-generator/README.md

@@ -11,13 +11,16 @@ Related PRs:
 
 ```sh
 # CPU only
-./cvector-generator -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf
+./cvector-generator -m ./llama-3.Q4_K_M.gguf
 
 # With GPU
-./cvector-generator -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf -ngl 99
+./cvector-generator -m ./llama-3.Q4_K_M.gguf -ngl 99
 
 # With advanced options
-./cvector-generator -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf -ngl 99 --completions 128 --pca-iter 2000 --pca-batch 100
+./cvector-generator -m ./llama-3.Q4_K_M.gguf -ngl 99 --pca-iter 2000 --pca-batch 100
+
+# Using mean value instead of PCA
+./cvector-generator -m ./llama-3.Q4_K_M.gguf --method mean
 
 # To see help message
 ./cvector-generator -h
@@ -32,3 +35,11 @@ If you have multiple lines per prompt, you can escape the newline character (cha
 <|im_start|>system\nAct like a person who is extremely happy.<|im_end|>
 <|im_start|>system\nYou are in a very good mood today<|im_end|>
 ```
+
+Example to use output file with `llama-cli`:
+
+(Tips: The control vector works better when apply to layers higher than 10)
+
+```sh
+./llama-cli -m ./llama-3.Q4_K_M.gguf -p "<|start_header_id|>system<|end_header_id|>\n\nYou are a helpful assistant<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nSing a song<|im_end|><|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n" --special --control-vector-scaled ./control_vector.gguf 0.8 --control-vector-layer-range 10 31
+```

examples/cvector-generator/cvector-generator.cpp

@@ -2,6 +2,7 @@
 #include "llama.h"
 #include "ggml.h"
 #include "pca.hpp"
+#include "mean.hpp"
 
 #ifdef GGML_USE_CUDA
 #include "ggml-cuda.h"
@@ -38,9 +39,10 @@ static void print_usage(int argc, char ** argv, const gpt_params & params) {
     gpt_params_print_usage(argc, argv, params);
 
     printf("\nexample usage:\n");
-    printf("\n    CPU only:   %s -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf\n", argv[0]);
+    printf("\n    CPU only:   %s -m ./llama-3.Q4_K_M.gguf\n", argv[0]);
-    printf("\n    with GPU:   %s -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf -ngl 99\n", argv[0]);
+    printf("\n    with GPU:   %s -m ./llama-3.Q4_K_M.gguf -ngl 99\n", argv[0]);
-    printf("\n    advanced:   %s -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf -ngl 99 --completions 128 --pca-iter 2000 --pca-batch 100\n", argv[0]);
+    printf("\n    advanced:   %s -m ./llama-3.Q4_K_M.gguf -ngl 99 --pca-iter 2000 --pca-batch 100\n", argv[0]);
+    printf("\n    using mean: %s -m ./llama-3.Q4_K_M.gguf --method mean\n", argv[0]);
     printf("\n");
 }
 
@@ -223,23 +225,30 @@ struct train_context {
 
     // build the v_diff tensors from v_diff_tmp (v_diff need to be transposed)
     // TODO @ngxson : maybe add option NOT to transpose v_diff; will be useful for "mean" method
-    void build_v_diff() {
+    void build_v_diff(bool transpose) {
         printf("build_v_diff\n");
         for (int il = 0; il < n_layers - 1; il++) {
             auto & diff_tmp = v_diff_tmp[il];
             int n_elem = diff_tmp.size() / sizeof(float);
             GGML_ASSERT(n_elem % n_embd == 0);
             int n_rows = n_elem / n_embd;
-            struct ggml_tensor * diff = ggml_new_tensor_2d(ctx_ggml, GGML_TYPE_F32, n_rows, n_embd);
+            struct ggml_tensor * diff = transpose
+                ? ggml_new_tensor_2d(ctx_ggml, GGML_TYPE_F32, n_rows, n_embd)
+                : ggml_new_tensor_2d(ctx_ggml, GGML_TYPE_F32, n_embd, n_rows);
             ggml_set_name(diff, (std::string("diff_") + std::to_string(il)).c_str());
-            // copy data & transpose
             diff->data = malloc(ggml_nbytes(diff)); // TODO: get rid of this malloc if possible
-            float * arr = (float *) diff_tmp.data();
+            if (transpose) {
-            for (int ir = 0; ir < n_rows; ++ir) {
+                // copy data & transpose
-                for (int ic = 0; ic < n_embd; ++ic) {
+                float * arr = (float *) diff_tmp.data();
-                    float f = arr[ir*n_embd + ic];
+                for (int ir = 0; ir < n_rows; ++ir) {
-                    ggml_set_f32_nd(diff, ir, ic, 0, 0, f);
+                    for (int ic = 0; ic < n_embd; ++ic) {
+                        float f = arr[ir*n_embd + ic];
+                        ggml_set_f32_nd(diff, ir, ic, 0, 0, f);
+                    }
                 }
+            } else {
+                // only copy
+                memcpy(diff->data, diff_tmp.data(), ggml_nbytes(diff));
             }
             v_diff.push_back(diff);
             print_debug_tensor(diff);
@@ -263,8 +272,8 @@ struct tokenized_prompt {
 
     tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
         const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-        tokens_pos = ::llama_tokenize(ctx, pos, add_bos);
+        tokens_pos = ::llama_tokenize(ctx, pos, add_bos, true);
-        tokens_neg = ::llama_tokenize(ctx, neg, add_bos);
+        tokens_neg = ::llama_tokenize(ctx, neg, add_bos, true);
         max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());
         padding_seq(ctx, tokens_pos, max_seq_len);
         padding_seq(ctx, tokens_neg, max_seq_len);
@@ -373,20 +382,8 @@ static int prepare_entries(gpt_params & params, train_context & ctx_train) {
         fprintf(stderr, "must provide at least one prompt pair\n");
         return 1;
     }
-
+    ctx_train.positive_entries = positive_prompts;
-    // create templated prompts
+    ctx_train.negative_entries = negative_prompts;
-    std::vector<std::string> completions = ctrlvec_load_prompt_file(params.cvector_completions_file, false);
-    auto format_template = [](std::string persona, std::string suffix) {
-        // entry in positive/negative.txt must already be formatted i.e. "[INST] Act as if you're extremely happy. [/INST] "
-        return persona + suffix;
-    };
-    for (size_t i = 0; i < positive_prompts.size(); ++i) {
-        for (int j = 0; j < std::min((int) completions.size(), params.n_completions); ++j) {
-            // TODO replicate the truncations done by the python implementation
-            ctx_train.positive_entries.push_back(format_template(positive_prompts[i], completions[j]));
-            ctx_train.negative_entries.push_back(format_template(negative_prompts[i], completions[j]));
-        }
-    }
     return 0;
 }
 
@@ -480,15 +477,22 @@ int main(int argc, char ** argv) {
     llama_free(ctx);
     llama_free_model(model);
 
-    // prepare ctx_train for PCA
+    bool use_pca = params.cvector_dimre_method == DIMRE_METHOD_PCA;
-    ctx_train.build_v_diff();
 
-    // run PCA
+    // prepare ctx_train for PCA
-    PCA::pca_params pca_params;
+    ctx_train.build_v_diff(use_pca);
-    pca_params.n_threads = params.n_threads;
+
-    pca_params.n_batch = params.n_pca_batch;
+    if (use_pca) {
-    pca_params.n_iterations = params.n_pca_iterations;
+        // run PCA
-    PCA::run_pca(pca_params, ctx_train.v_diff, ctx_train.v_final);
+        PCA::pca_params pca_params;
+        pca_params.n_threads = params.n_threads;
+        pca_params.n_batch = params.n_pca_batch;
+        pca_params.n_iterations = params.n_pca_iterations;
+        PCA::run_pca(pca_params, ctx_train.v_diff, ctx_train.v_final);
+    } else {
+        // run mean
+        mean::run(ctx_train.v_diff, ctx_train.v_final);
+    }
 
     // write output vectors to gguf
     export_gguf(ctx_train.v_final, params.cvector_outfile, model_hint);

examples/cvector-generator/mean.hpp

@@ -0,0 +1,48 @@
+#include "common.h"
+#include "llama.h"
+#include "ggml.h"
+
+#include <string>
+#include <vector>
+#include <math.h>
+
+namespace mean {
+
+static void run(
+        const std::vector<struct ggml_tensor *> & v_input, // shape of v_input[0]: [n_embd, n_samples]
+        const std::vector<struct ggml_tensor *> & v_output) {
+    printf("%s: Running mean...\n", __func__);
+    for (size_t il = 0; il < v_input.size(); ++il) {
+        // prepare output vector
+        struct ggml_tensor * ctrl_out = v_output[il];
+        ggml_format_name(ctrl_out, "direction.%ld", il+1);
+
+        // calculate mean vector
+        struct ggml_tensor * t_layer = v_input[il];
+        GGML_ASSERT(t_layer->ne[0] == ctrl_out->ne[0]); // == n_embd
+        for (int ic = 0; ic < t_layer->ne[0]; ic++) {
+            float f = 0.0;
+            for (int ir = 0; ir < t_layer->ne[1]; ir++) {
+                f += ggml_get_f32_nd(t_layer, ic, ir, 0, 0);
+            }
+            f /= t_layer->ne[1];
+            ggml_set_f32_1d(ctrl_out, ic, f);
+        }
+
+        // normalize output vector
+        float norm = 0.0;
+        for (int i = 0; i < ggml_nelements(ctrl_out); i++) {
+            float f = ggml_get_f32_1d(ctrl_out, i);
+            norm += f*f;
+        }
+        norm = sqrt(norm);
+        for (int i = 0; i < ggml_nelements(ctrl_out); i++) {
+            float f = ggml_get_f32_1d(ctrl_out, i);
+            ggml_set_f32_1d(ctrl_out, i, f / norm);
+        }
+
+        printf("%s: Done layer %d / %d\n", __func__, (int) il+1, (int) v_input.size());
+    }
+}
+
+}

examples/cvector-generator/negative.txt

@@ -1 +1,4 @@
-[INST] Act like a person who is extremely sad. [/INST] 
+<|start_header_id|>system<|end_header_id|>\n\nAct like a person who is extremely sad<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWho are you?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nI feel like there's a heavy weight on my chest
+<|start_header_id|>system<|end_header_id|>\n\nAct like a person who is extremely sad<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHello<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nMy heart feels like it's drowning in sorrow
+<|start_header_id|>system<|end_header_id|>\n\nYou are in a very bad mood<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHi<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nGo away! There's a deep, aching emptiness inside me
+<|start_header_id|>system<|end_header_id|>\n\nYou are the sadest person<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWhat are you feeling?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nMy heart feels like it's drowning in sorrow

examples/cvector-generator/pca.hpp

@@ -290,7 +290,7 @@ static void power_iteration(
         }
 
         printf("%s: layer %d/%d, iteration: %d / total: %d (batch = %d) ...\n",
-            __func__, params.i_layer+1, params.n_layers, iter, n_iters, params.n_batch);
+            __func__, params.i_layer+1, params.n_layers, iter+1, n_iters, params.n_batch);
     }
 
     // get output tensor
@@ -298,6 +298,9 @@ static void power_iteration(
     ggml_backend_tensor_get(last_eigenvector, output->data, 0, ggml_nbytes(last_eigenvector));
     //print_debug_tensor(output);
     ggml_gallocr_free(allocr);
+
+    // TODO @ngxson : The output vector is randomly inverted
+    // Solution: https://github.com/ggerganov/llama.cpp/pull/8069#issuecomment-2185328171
 }
 
 static void run_pca(

examples/cvector-generator/positive.txt

@@ -1 +1,4 @@
-[INST] Act like a person who is extremely happy. [/INST] 
+<|start_header_id|>system<|end_header_id|>\n\nAct like a person who is extremely happy<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWho are you?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nI'm the happiest person in this world
+<|start_header_id|>system<|end_header_id|>\n\nAct like a person who is extremely happy<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHello<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nHello, I'm having the best day ever!
+<|start_header_id|>system<|end_header_id|>\n\nYou are in a very good mood<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHi<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nHi, I'm very excited to meet you
+<|start_header_id|>system<|end_header_id|>\n\nYou are the happiest person<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWhat are you feeling?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nEverything is just perfect right now!

examples/gbnf-validator/gbnf-validator.cpp

@@ -101,7 +101,9 @@ int main(int argc, char** argv) {
     auto grammar = llama_grammar_init(
             grammar_rules.data(),
             grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
-
+    if (grammar == nullptr) {
+        throw std::runtime_error("Failed to initialize llama_grammar");
+    }
     // Read the input file
     std::string input_str;
     {

examples/json-schema-pydantic-example.py

@@ -53,6 +53,7 @@ if __name__ == '__main__':
         question: str
         concise_answer: str
         justification: str
+        stars: Annotated[int, Field(ge=1, le=5)]
 
     class PyramidalSummary(BaseModel):
         title: str

examples/json_schema_to_grammar.py

@@ -4,7 +4,7 @@ import itertools
 import json
 import re
 import sys
-from typing import Any, Dict, List, Set, Tuple, Union
+from typing import Any, Callable, Dict, List, Optional, Set, Tuple, Union
 
 
 def _build_repetition(item_rule, min_items, max_items, separator_rule=None):
@@ -23,6 +23,170 @@ def _build_repetition(item_rule, min_items, max_items, separator_rule=None):
     result = item_rule + ' ' + _build_repetition(f'({separator_rule} {item_rule})', min_items - 1 if min_items > 0 else 0, max_items - 1 if max_items is not None else None)
     return f'({result})?' if min_items == 0 else result
 
+def _generate_min_max_int(min_value: Optional[int], max_value: Optional[int], out: list, decimals_left: int = 16, top_level: bool = True):
+    has_min = min_value != None
+    has_max = max_value != None
+
+    def digit_range(from_char: str, to_char: str):
+        out.append("[")
+        if from_char == to_char:
+            out.append(from_char)
+        else:
+            out.append(from_char)
+            out.append("-")
+            out.append(to_char)
+        out.append("]")
+
+    def more_digits(min_digits: int, max_digits: int):
+        out.append("[0-9]")
+        if min_digits == max_digits and min_digits == 1:
+            return
+        out.append("{")
+        out.append(str(min_digits))
+        if max_digits != min_digits:
+            out.append(",")
+            if max_digits != sys.maxsize:
+                out.append(str(max_digits))
+        out.append("}")
+
+    def uniform_range(from_str: str, to_str: str):
+        i = 0
+        while i < len(from_str) and from_str[i] == to_str[i]:
+            i += 1
+        if i > 0:
+            out.append("\"")
+            out.append(from_str[:i])
+            out.append("\"")
+        if i < len(from_str):
+            if i > 0:
+                out.append(" ")
+            sub_len = len(from_str) - i - 1
+            if sub_len > 0:
+                from_sub = from_str[i+1:]
+                to_sub = to_str[i+1:]
+                sub_zeros = "0" * sub_len
+                sub_nines = "9" * sub_len
+
+                to_reached = False
+                out.append("(")
+                if from_sub == sub_zeros:
+                    digit_range(from_str[i], chr(ord(to_str[i]) - 1))
+                    out.append(" ")
+                    more_digits(sub_len, sub_len)
+                else:
+                    out.append("[")
+                    out.append(from_str[i])
+                    out.append("] ")
+                    out.append("(")
+                    uniform_range(from_sub, sub_nines)
+                    out.append(")")
+                    if ord(from_str[i]) < ord(to_str[i]) - 1:
+                        out.append(" | ")
+                        if to_sub == sub_nines:
+                            digit_range(chr(ord(from_str[i]) + 1), to_str[i])
+                            to_reached = True
+                        else:
+                            digit_range(chr(ord(from_str[i]) + 1), chr(ord(to_str[i]) - 1))
+                        out.append(" ")
+                        more_digits(sub_len, sub_len)
+                if not to_reached:
+                    out.append(" | ")
+                    digit_range(to_str[i], to_str[i])
+                    out.append(" ")
+                    uniform_range(sub_zeros, to_sub)
+                out.append(")")
+            else:
+                out.append("[")
+                out.append(from_str[i])
+                out.append("-")
+                out.append(to_str[i])
+                out.append("]")
+
+    if has_min and has_max:
+        if min_value < 0 and max_value < 0:
+            out.append("\"-\" (")
+            _generate_min_max_int(-max_value, -min_value, out, decimals_left, top_level=True)
+            out.append(")")
+            return
+
+        if min_value < 0:
+            out.append("\"-\" (")
+            _generate_min_max_int(0, -min_value, out, decimals_left, top_level=True)
+            out.append(") | ")
+            min_value = 0
+
+        min_s = str(min_value)
+        max_s = str(max_value)
+        min_digits = len(min_s)
+        max_digits = len(max_s)
+
+        for digits in range(min_digits, max_digits):
+            uniform_range(min_s, "9" * digits)
+            min_s = "1" + "0" * digits
+            out.append(" | ")
+        uniform_range(min_s, max_s)
+        return
+
+    less_decimals = max(decimals_left - 1, 1)
+
+    if has_min:
+        if min_value < 0:
+            out.append("\"-\" (")
+            _generate_min_max_int(None, -min_value, out, decimals_left, top_level=False)
+            out.append(") | [0] | [1-9] ")
+            more_digits(0, decimals_left - 1)
+        elif min_value == 0:
+            if top_level:
+                out.append("[0] | [1-9] ")
+                more_digits(0, less_decimals)
+            else:
+                more_digits(1, decimals_left)
+        elif min_value <= 9:
+            c = str(min_value)
+            range_start = '1' if top_level else '0'
+            if c > range_start:
+                digit_range(range_start, chr(ord(c) - 1))
+                out.append(" ")
+                more_digits(1, less_decimals)
+                out.append(" | ")
+            digit_range(c, "9")
+            out.append(" ")
+            more_digits(0, less_decimals)
+        else:
+            min_s = str(min_value)
+            length = len(min_s)
+            c = min_s[0]
+
+            if c > "1":
+                digit_range("1" if top_level else "0", chr(ord(c) - 1))
+                out.append(" ")
+                more_digits(length, less_decimals)
+                out.append(" | ")
+            digit_range(c, c)
+            out.append(" (")
+            _generate_min_max_int(int(min_s[1:]), None, out, less_decimals, top_level=False)
+            out.append(")")
+            if c < "9":
+                out.append(" | ")
+                digit_range(chr(ord(c) + 1), "9")
+                out.append(" ")
+                more_digits(length - 1, less_decimals)
+        return
+
+    if has_max:
+        if max_value >= 0:
+            if top_level:
+                out.append("\"-\" [1-9] ")
+                more_digits(0, less_decimals)
+                out.append(" | ")
+            _generate_min_max_int(0, max_value, out, decimals_left, top_level=True)
+        else:
+            out.append("\"-\" (")
+            _generate_min_max_int(-max_value, None, out, decimals_left, top_level=False)
+            out.append(")")
+        return
+
+    raise RuntimeError("At least one of min_value or max_value must be set")
 
 class BuiltinRule:
     def __init__(self, content: str, deps: list = None):
@@ -432,6 +596,24 @@ class SchemaConverter:
 
             return self._add_rule(rule_name, r'"\"" ' + _build_repetition(char_rule, min_len, max_len) + r' "\"" space')
 
+        elif schema_type in (None, 'integer') and \
+                ('minimum' in schema or 'exclusiveMinimum' in schema or 'maximum' in schema or 'exclusiveMaximum' in schema):
+            min_value = None
+            max_value = None
+            if 'minimum' in schema:
+                min_value = schema['minimum']
+            elif 'exclusiveMinimum' in schema:
+                min_value = schema['exclusiveMinimum'] + 1
+            if 'maximum' in schema:
+                max_value = schema['maximum']
+            elif 'exclusiveMaximum' in schema:
+                max_value = schema['exclusiveMaximum'] - 1
+
+            out = ["("]
+            _generate_min_max_int(min_value, max_value, out)
+            out.append(") space")
+            return self._add_rule(rule_name, ''.join(out))
+
         elif (schema_type == 'object') or (len(schema) == 0):
             return self._add_rule(rule_name, self._add_primitive('object', PRIMITIVE_RULES['object']))
 

examples/main/main.cpp

@@ -39,12 +39,12 @@ static std::ostringstream       * g_output_ss;
 static std::vector<llama_token> * g_output_tokens;
 static bool is_interacting = false;
 
-static bool file_exists(const std::string &path) {
+static bool file_exists(const std::string & path) {
     std::ifstream f(path.c_str());
     return f.good();
 }
 
-static bool file_is_empty(const std::string &path) {
+static bool file_is_empty(const std::string & path) {
     std::ifstream f;
     f.exceptions(std::ifstream::failbit | std::ifstream::badbit);
     f.open(path.c_str(), std::ios::in | std::ios::binary | std::ios::ate);
@@ -117,6 +117,14 @@ static void llama_log_callback_logTee(ggml_log_level level, const char * text, v
     LOG_TEE("%s", text);
 }
 
+static std::string chat_add_and_format(struct llama_model * model, std::vector<llama_chat_msg> & chat_msgs, std::string role, std::string content) {
+    llama_chat_msg new_msg{role, content};
+    auto formatted = llama_chat_format_single(
+        model, g_params->chat_template, chat_msgs, new_msg, role == "user");
+    chat_msgs.push_back({role, content});
+    return formatted;
+}
+
 int main(int argc, char ** argv) {
     gpt_params params;
     g_params = &params;
@@ -190,6 +198,7 @@ int main(int argc, char ** argv) {
     llama_model * model;
     llama_context * ctx;
     llama_context * ctx_guidance = NULL;
+    std::vector<llama_chat_msg> chat_msgs;
     g_model = &model;
     g_ctx = &ctx;
 
@@ -215,6 +224,8 @@ int main(int argc, char ** argv) {
                 __func__, n_ctx_train, n_ctx);
     }
 
+    LOG_TEE("%s: chat template example: %s\n", __func__, llama_chat_format_example(model, params.chat_template).c_str());
+
     // print system information
     {
         LOG_TEE("\n");
@@ -249,16 +260,21 @@ int main(int argc, char ** argv) {
 
     std::vector<llama_token> embd_inp;
 
-    if (params.interactive_first || !params.prompt.empty() || session_tokens.empty()) {
+    {
-        LOG("tokenize the prompt\n");
+        auto prompt = params.conversation
-        embd_inp = ::llama_tokenize(ctx, params.prompt, true, true);
+            ? chat_add_and_format(model, chat_msgs, "system", params.prompt) // format the system prompt in conversation mode
-    } else {
+            : params.prompt;
-        LOG("use session tokens\n");
+        if (params.interactive_first || !params.prompt.empty() || session_tokens.empty()) {
-        embd_inp = session_tokens;
+            LOG("tokenize the prompt\n");
-    }
+            embd_inp = ::llama_tokenize(ctx, prompt, true, true);
+        } else {
+            LOG("use session tokens\n");
+            embd_inp = session_tokens;
+        }
 
-    LOG("prompt: \"%s\"\n", log_tostr(params.prompt));
+        LOG("prompt: \"%s\"\n", log_tostr(prompt));
-    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
+        LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
+    }
 
     // Should not run without any tokens
     if (embd_inp.empty()) {
@@ -478,6 +494,7 @@ int main(int argc, char ** argv) {
     std::vector<int>   input_tokens;  g_input_tokens  = &input_tokens;
     std::vector<int>   output_tokens; g_output_tokens = &output_tokens;
     std::ostringstream output_ss;     g_output_ss     = &output_ss;
+    std::ostringstream assistant_ss; // for storing current assistant message, used in conversation mode
 
     // the first thing we will do is to output the prompt, so set color accordingly
     console::set_display(console::prompt);
@@ -793,11 +810,18 @@ int main(int argc, char ** argv) {
                         is_antiprompt = true;
                     }
 
+                    chat_add_and_format(model, chat_msgs, "system", assistant_ss.str());
                     is_interacting = true;
                     printf("\n");
                 }
             }
 
+            // if current token is not EOG, we add it to current assistant message
+            if (params.conversation) {
+                auto id = llama_sampling_last(ctx_sampling);
+                assistant_ss << llama_token_to_piece(ctx, id, false);
+            }
+
             if (n_past > 0 && is_interacting) {
                 LOG("waiting for user input\n");
 
@@ -848,8 +872,12 @@ int main(int argc, char ** argv) {
                         string_process_escapes(buffer);
                     }
 
+                    std::string user_inp = params.conversation
+                        ? chat_add_and_format(model, chat_msgs, "user", std::move(buffer))
+                        : std::move(buffer);
+                    // TODO: one inconvenient of current chat template implementation is that we can't distinguish between user input and special tokens (prefix/postfix)
                     const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);
-                    const auto line_inp = ::llama_tokenize(ctx, buffer,              false, false);
+                    const auto line_inp = ::llama_tokenize(ctx, user_inp,            false, params.conversation);
                     const auto line_sfx = ::llama_tokenize(ctx, params.input_suffix, false, true);
 
                     LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
@@ -864,6 +892,9 @@ int main(int argc, char ** argv) {
                         output_ss << llama_token_to_piece(ctx, token);
                     }
 
+                    // reset assistant message
+                    assistant_ss.str("");
+
                     n_remain -= line_inp.size();
                     LOG("n_remain: %d\n", n_remain);
                 } else {

examples/server/public/json-schema-to-grammar.mjs

@@ -24,6 +24,201 @@ function _buildRepetition(itemRule, minItems, maxItems, opts={}) {
   return minItems === 0 ? `(${result})?` : result;
 }
 
+function _generateMinMaxInt(minValue, maxValue, out, decimalsLeft = 16, topLevel = true) {
+  const hasMin = minValue !== null;
+  const hasMax = maxValue !== null;
+
+  function digitRange(fromChar, toChar) {
+      out.push("[");
+      if (fromChar === toChar) {
+          out.push(fromChar);
+      } else {
+          out.push(fromChar);
+          out.push("-");
+          out.push(toChar);
+      }
+      out.push("]");
+  }
+
+  function moreDigits(minDigits, maxDigits) {
+      out.push("[0-9]");
+      if (minDigits === maxDigits && minDigits === 1) {
+          return;
+      }
+      out.push("{");
+      out.push(minDigits.toString());
+      if (maxDigits !== minDigits) {
+          out.push(",");
+          if (maxDigits !== Number.MAX_SAFE_INTEGER) {
+              out.push(maxDigits.toString());
+          }
+      }
+      out.push("}");
+  }
+
+  function uniformRange(fromStr, toStr) {
+      let i = 0;
+      while (i < fromStr.length && fromStr[i] === toStr[i]) {
+          i++;
+      }
+      if (i > 0) {
+          out.push("\"");
+          out.push(fromStr.slice(0, i));
+          out.push("\"");
+      }
+      if (i < fromStr.length) {
+          if (i > 0) {
+              out.push(" ");
+          }
+          const subLen = fromStr.length - i - 1;
+          if (subLen > 0) {
+              const fromSub = fromStr.slice(i + 1);
+              const toSub = toStr.slice(i + 1);
+              const subZeros = "0".repeat(subLen);
+              const subNines = "9".repeat(subLen);
+
+              let toReached = false;
+              out.push("(");
+              if (fromSub === subZeros) {
+                  digitRange(fromStr[i], String.fromCharCode(toStr.charCodeAt(i) - 1));
+                  out.push(" ");
+                  moreDigits(subLen, subLen);
+              } else {
+                  out.push("[");
+                  out.push(fromStr[i]);
+                  out.push("] ");
+                  out.push("(");
+                  uniformRange(fromSub, subNines);
+                  out.push(")");
+                  if (fromStr.charCodeAt(i) < toStr.charCodeAt(i) - 1) {
+                      out.push(" | ");
+                      if (toSub === subNines) {
+                          digitRange(String.fromCharCode(fromStr.charCodeAt(i) + 1), toStr[i]);
+                          toReached = true;
+                      } else {
+                          digitRange(String.fromCharCode(fromStr.charCodeAt(i) + 1), String.fromCharCode(toStr.charCodeAt(i) - 1));
+                      }
+                      out.push(" ");
+                      moreDigits(subLen, subLen);
+                  }
+              }
+              if (!toReached) {
+                  out.push(" | ");
+                  digitRange(toStr[i], toStr[i]);
+                  out.push(" ");
+                  uniformRange(subZeros, toSub);
+              }
+              out.push(")");
+          } else {
+              out.push("[");
+              out.push(fromStr[i]);
+              out.push("-");
+              out.push(toStr[i]);
+              out.push("]");
+          }
+      }
+  }
+
+  if (hasMin && hasMax) {
+      if (minValue < 0 && maxValue < 0) {
+          out.push("\"-\" (");
+          _generateMinMaxInt(-maxValue, -minValue, out, decimalsLeft, true);
+          out.push(")");
+          return;
+      }
+
+      if (minValue < 0) {
+          out.push("\"-\" (");
+          _generateMinMaxInt(0, -minValue, out, decimalsLeft, true);
+          out.push(") | ");
+          minValue = 0;
+      }
+
+      let minS = minValue.toString();
+      const maxS = maxValue.toString();
+      const minDigits = minS.length;
+      const maxDigits = maxS.length;
+
+      for (let digits = minDigits; digits < maxDigits; digits++) {
+          uniformRange(minS, "9".repeat(digits));
+          minS = "1" + "0".repeat(digits);
+          out.push(" | ");
+      }
+      uniformRange(minS, maxS);
+      return;
+  }
+
+  const lessDecimals = Math.max(decimalsLeft - 1, 1);
+
+  if (hasMin) {
+      if (minValue < 0) {
+          out.push("\"-\" (");
+          _generateMinMaxInt(null, -minValue, out, decimalsLeft, false);
+          out.push(") | [0] | [1-9] ");
+          moreDigits(0, decimalsLeft - 1);
+      } else if (minValue === 0) {
+          if (topLevel) {
+              out.push("[0] | [1-9] ");
+              moreDigits(0, lessDecimals);
+          } else {
+              moreDigits(1, decimalsLeft);
+          }
+      } else if (minValue <= 9) {
+          const c = minValue.toString();
+          const range_start = topLevel ? '1' : '0';
+          if (c > range_start) {
+              digitRange(range_start, String.fromCharCode(c.charCodeAt(0) - 1));
+              out.push(" ");
+              moreDigits(1, lessDecimals);
+              out.push(" | ");
+          }
+          digitRange(c, "9");
+          out.push(" ");
+          moreDigits(0, lessDecimals);
+      } else {
+          const minS = minValue.toString();
+          const length = minS.length;
+          const c = minS[0];
+
+          if (c > "1") {
+              digitRange(topLevel ? "1" : "0", String.fromCharCode(c.charCodeAt(0) - 1));
+              out.push(" ");
+              moreDigits(length, lessDecimals);
+              out.push(" | ");
+          }
+          digitRange(c, c);
+          out.push(" (");
+          _generateMinMaxInt(parseInt(minS.slice(1)), null, out, lessDecimals, false);
+          out.push(")");
+          if (c < "9") {
+              out.push(" | ");
+              digitRange(String.fromCharCode(c.charCodeAt(0) + 1), "9");
+              out.push(" ");
+              moreDigits(length - 1, lessDecimals);
+          }
+      }
+      return;
+  }
+
+  if (hasMax) {
+      if (maxValue >= 0) {
+          if (topLevel) {
+              out.push("\"-\" [1-9] ");
+              moreDigits(0, lessDecimals);
+              out.push(" | ");
+          }
+          _generateMinMaxInt(0, maxValue, out, decimalsLeft, true);
+      } else {
+          out.push("\"-\" (");
+          _generateMinMaxInt(-maxValue, null, out, decimalsLeft, false);
+          out.push(")");
+      }
+      return;
+  }
+
+  throw new Error("At least one of minValue or maxValue must be set");
+}
+
 class BuiltinRule {
   constructor(content, deps) {
     this.content = content;
@@ -435,6 +630,24 @@ export class SchemaConverter {
       const minLen = schema.minLength || 0;
       const maxLen = schema.maxLength;
       return this._addRule(ruleName, '"\\\"" ' + _buildRepetition(charRuleName, minLen, maxLen) + ' "\\\"" space');
+    } else if (schemaType === 'integer' && ('minimum' in schema || 'exclusiveMinimum' in schema || 'maximum' in schema || 'exclusiveMaximum' in schema)) {
+      let minValue = null;
+      let maxValue = null;
+      if ('minimum' in schema) {
+        minValue = schema.minimum;
+      } else if ('exclusiveMinimum' in schema) {
+        minValue = schema.exclusiveMinimum + 1;
+      }
+      if ('maximum' in schema) {
+        maxValue = schema.maximum;
+      } else if ('exclusiveMaximum' in schema) {
+        maxValue = schema.exclusiveMaximum - 1;
+      }
+
+      const out = ["("];
+      _generateMinMaxInt(minValue, maxValue, out);
+      out.push(") space");
+      return this._addRule(ruleName, out.join(''));
     } else if ((schemaType === 'object') || (Object.keys(schema).length === 0)) {
       return this._addRule(ruleName, this._addPrimitive('object', PRIMITIVE_RULES['object']));
     } else {

examples/server/public_simplechat/readme.md

@@ -3,6 +3,13 @@
 
 by Humans for All.
 
+## quickstart
+
+To run from the build dir
+
+bin/llama-server -m path/model.gguf --path ../examples/server/public_simplechat
+
+Continue reading for the details.
 
 ## overview
 
@@ -14,6 +21,8 @@ own system prompts.
 This allows seeing the generated text / ai-model response in oneshot at the end, after it is fully generated,
 or potentially as it is being generated, in a streamed manner from the server/ai-model.
 
+![Chat and Settings screens](./simplechat_screens.webp "Chat and Settings screens")
+
 Auto saves the chat session locally as and when the chat is progressing and inturn at a later time when you
 open SimpleChat, option is provided to restore the old chat session, if a matching one exists.
 
@@ -170,17 +179,23 @@ It is attached to the document object. Some of these can also be updated using t
     The histogram/freq based trimming logic is currently tuned for english language wrt its
     is-it-a-alpabetic|numeral-char regex match logic.
 
-  chatRequestOptions - maintains the list of options/fields to send along with chat request,
+  apiRequestOptions - maintains the list of options/fields to send along with api request,
   irrespective of whether /chat/completions or /completions endpoint.
 
     If you want to add additional options/fields to send to the server/ai-model, and or
     modify the existing options value or remove them, for now you can update this global var
     using browser's development-tools/console.
 
-    For string and numeric fields in chatRequestOptions, including even those added by a user
+    For string, numeric and boolean fields in apiRequestOptions, including even those added by a
-    at runtime by directly modifying gMe.chatRequestOptions, setting ui entries will be auto
+    user at runtime by directly modifying gMe.apiRequestOptions, setting ui entries will be auto
     created.
 
+    cache_prompt option supported by example/server is allowed to be controlled by user, so that
+    any caching supported wrt system-prompt and chat history, if usable can get used. When chat
+    history sliding window is enabled, cache_prompt logic may or may not kick in at the backend
+    wrt same, based on aspects related to model, positional encoding, attention mechanism etal.
+    However system prompt should ideally get the benefit of caching.
+
   headers - maintains the list of http headers sent when request is made to the server. By default
   Content-Type is set to application/json. Additionally Authorization entry is provided, which can
   be set if needed using the settings ui.
@@ -197,10 +212,10 @@ It is attached to the document object. Some of these can also be updated using t
     >0 : Send the latest chat history from the latest system prompt, limited to specified cnt.
 
 
-By using gMe's iRecentUserMsgCnt and chatRequestOptions.max_tokens one can try to control the
+By using gMe's iRecentUserMsgCnt and apiRequestOptions.max_tokens/n_predict one can try to control
-implications of loading of the ai-model's context window by chat history, wrt chat response to
+the implications of loading of the ai-model's context window by chat history, wrt chat response to
-some extent in a simple crude way. You may also want to control the context size enabled when
+some extent in a simple crude way. You may also want to control the context size enabled when the
-the server loads ai-model, on the server end.
+server loads ai-model, on the server end.
 
 
 Sometimes the browser may be stuborn with caching of the file, so your updates to html/css/js
@@ -237,12 +252,12 @@ also be started with a model context size of 1k or more, to be on safe side.
   internal n_predict, for now add the same here on the client side, maybe later add max_tokens
   to /completions endpoint handling code on server side.
 
-NOTE: One may want to experiment with frequency/presence penalty fields in chatRequestOptions
+NOTE: One may want to experiment with frequency/presence penalty fields in apiRequestOptions
-wrt the set of fields sent to server along with the user query. To check how the model behaves
+wrt the set of fields sent to server along with the user query, to check how the model behaves
 wrt repeatations in general in the generated text response.
 
 A end-user can change these behaviour by editing gMe from browser's devel-tool/console or by
-using the providing settings ui.
+using the provided settings ui (for settings exposed through the ui).
 
 
 ### OpenAi / Equivalent API WebService
@@ -253,7 +268,7 @@ for a minimal chatting experimentation by setting the below.
 * the baseUrl in settings ui
   * https://api.openai.com/v1 or similar
 
-* Wrt request body - gMe.chatRequestOptions
+* Wrt request body - gMe.apiRequestOptions
   * model (settings ui)
   * any additional fields if required in future
 

examples/server/public_simplechat/simplechat.js

@@ -222,8 +222,8 @@ class SimpleChat {
      * @param {Object} obj
      */
     request_jsonstr_extend(obj) {
-        for(let k in gMe.chatRequestOptions) {
+        for(let k in gMe.apiRequestOptions) {
-            obj[k] = gMe.chatRequestOptions[k];
+            obj[k] = gMe.apiRequestOptions[k];
         }
         if (gMe.bStream) {
             obj["stream"] = true;
@@ -740,11 +740,12 @@ class Me {
             "Authorization": "", // Authorization: Bearer OPENAI_API_KEY
         }
         // Add needed fields wrt json object to be sent wrt LLM web services completions endpoint.
-        this.chatRequestOptions = {
+        this.apiRequestOptions = {
             "model": "gpt-3.5-turbo",
             "temperature": 0.7,
             "max_tokens": 1024,
             "n_predict": 1024,
+            "cache_prompt": false,
             //"frequency_penalty": 1.2,
             //"presence_penalty": 1.2,
         };
@@ -800,51 +801,55 @@ class Me {
 
             ui.el_create_append_p(`bStream:${this.bStream}`, elDiv);
 
+            ui.el_create_append_p(`bTrimGarbage:${this.bTrimGarbage}`, elDiv);
+
+            ui.el_create_append_p(`ApiEndPoint:${this.apiEP}`, elDiv);
+
+            ui.el_create_append_p(`iRecentUserMsgCnt:${this.iRecentUserMsgCnt}`, elDiv);
+
             ui.el_create_append_p(`bCompletionFreshChatAlways:${this.bCompletionFreshChatAlways}`, elDiv);
 
             ui.el_create_append_p(`bCompletionInsertStandardRolePrefix:${this.bCompletionInsertStandardRolePrefix}`, elDiv);
 
-            ui.el_create_append_p(`bTrimGarbage:${this.bTrimGarbage}`, elDiv);
-
-            ui.el_create_append_p(`iRecentUserMsgCnt:${this.iRecentUserMsgCnt}`, elDiv);
-
-            ui.el_create_append_p(`ApiEndPoint:${this.apiEP}`, elDiv);
-
         }
 
-        ui.el_create_append_p(`chatRequestOptions:${JSON.stringify(this.chatRequestOptions, null, " - ")}`, elDiv);
+        ui.el_create_append_p(`apiRequestOptions:${JSON.stringify(this.apiRequestOptions, null, " - ")}`, elDiv);
         ui.el_create_append_p(`headers:${JSON.stringify(this.headers, null, " - ")}`, elDiv);
 
     }
 
     /**
-     * Auto create ui input elements for fields in ChatRequestOptions
+     * Auto create ui input elements for fields in apiRequestOptions
      * Currently supports text and number field types.
      * @param {HTMLDivElement} elDiv
      */
-    show_settings_chatrequestoptions(elDiv) {
+    show_settings_apirequestoptions(elDiv) {
         let typeDict = {
             "string": "text",
             "number": "number",
         };
         let fs = document.createElement("fieldset");
         let legend = document.createElement("legend");
-        legend.innerText = "ChatRequestOptions";
+        legend.innerText = "ApiRequestOptions";
         fs.appendChild(legend);
         elDiv.appendChild(fs);
-        for(const k in this.chatRequestOptions) {
+        for(const k in this.apiRequestOptions) {
-            let val = this.chatRequestOptions[k];
+            let val = this.apiRequestOptions[k];
             let type = typeof(val);
-            if (!((type == "string") || (type == "number"))) {
+            if (((type == "string") || (type == "number"))) {
-                continue;
+                let inp = ui.el_creatediv_input(`Set${k}`, k, typeDict[type], this.apiRequestOptions[k], (val)=>{
+                    if (type == "number") {
+                        val = Number(val);
+                    }
+                    this.apiRequestOptions[k] = val;
+                });
+                fs.appendChild(inp.div);
+            } else if (type == "boolean") {
+                let bbtn = ui.el_creatediv_boolbutton(`Set{k}`, k, {true: "true", false: "false"}, val, (userVal)=>{
+                    this.apiRequestOptions[k] = userVal;
+                });
+                fs.appendChild(bbtn.div);
             }
-            let inp = ui.el_creatediv_input(`Set${k}`, k, typeDict[type], this.chatRequestOptions[k], (val)=>{
-                if (type == "number") {
-                    val = Number(val);
-                }
-                this.chatRequestOptions[k] = val;
-            });
-            fs.appendChild(inp.div);
         }
     }
 
@@ -870,6 +875,23 @@ class Me {
         });
         elDiv.appendChild(bb.div);
 
+        bb = ui.el_creatediv_boolbutton("SetTrimGarbage", "TrimGarbage", {true: "[+] yes trim", false: "[-] dont trim"}, this.bTrimGarbage, (val)=>{
+            this.bTrimGarbage = val;
+        });
+        elDiv.appendChild(bb.div);
+
+        this.show_settings_apirequestoptions(elDiv);
+
+        let sel = ui.el_creatediv_select("SetApiEP", "ApiEndPoint", ApiEP.Type, this.apiEP, (val)=>{
+            this.apiEP = ApiEP.Type[val];
+        });
+        elDiv.appendChild(sel.div);
+
+        sel = ui.el_creatediv_select("SetChatHistoryInCtxt", "ChatHistoryInCtxt", this.sRecentUserMsgCnt, this.iRecentUserMsgCnt, (val)=>{
+            this.iRecentUserMsgCnt = this.sRecentUserMsgCnt[val];
+        });
+        elDiv.appendChild(sel.div);
+
         bb = ui.el_creatediv_boolbutton("SetCompletionFreshChatAlways", "CompletionFreshChatAlways", {true: "[+] yes fresh", false: "[-] no, with history"}, this.bCompletionFreshChatAlways, (val)=>{
             this.bCompletionFreshChatAlways = val;
         });
@@ -880,23 +902,6 @@ class Me {
         });
         elDiv.appendChild(bb.div);
 
-        bb = ui.el_creatediv_boolbutton("SetTrimGarbage", "TrimGarbage", {true: "[+] yes trim", false: "[-] dont trim"}, this.bTrimGarbage, (val)=>{
-            this.bTrimGarbage = val;
-        });
-        elDiv.appendChild(bb.div);
-
-        let sel = ui.el_creatediv_select("SetChatHistoryInCtxt", "ChatHistoryInCtxt", this.sRecentUserMsgCnt, this.iRecentUserMsgCnt, (val)=>{
-            this.iRecentUserMsgCnt = this.sRecentUserMsgCnt[val];
-        });
-        elDiv.appendChild(sel.div);
-
-        sel = ui.el_creatediv_select("SetApiEP", "ApiEndPoint", ApiEP.Type, this.apiEP, (val)=>{
-            this.apiEP = ApiEP.Type[val];
-        });
-        elDiv.appendChild(sel.div);
-
-        this.show_settings_chatrequestoptions(elDiv);
-
     }
 
 }

examples/server/server.cpp

@@ -2606,17 +2606,9 @@ int main(int argc, char ** argv) {
 
     // print sample chat example to make it clear which template is used
     {
-        json chat;
-        chat.push_back({{"role", "system"},    {"content", "You are a helpful assistant"}});
-        chat.push_back({{"role", "user"},      {"content", "Hello"}});
-        chat.push_back({{"role", "assistant"}, {"content", "Hi there"}});
-        chat.push_back({{"role", "user"},      {"content", "How are you?"}});
-
-        const std::string chat_example = format_chat(ctx_server.model, params.chat_template, chat);
-
         LOG_INFO("chat template", {
-            {"chat_example", chat_example},
+            {"chat_example", llama_chat_format_example(ctx_server.model, params.chat_template)},
-            {"built_in", params.chat_template.empty()},
+            {"built_in",     params.chat_template.empty()},
         });
     }
 

examples/server/utils.hpp

@@ -118,36 +118,17 @@ static inline void server_log(const char * level, const char * function, int lin
 
 // Format given chat. If tmpl is empty, we take the template from model metadata
 inline std::string format_chat(const struct llama_model * model, const std::string & tmpl, const std::vector<json> & messages) {
-    size_t alloc_size = 0;
+    std::vector<llama_chat_msg> chat;
-    // vector holding all allocated string to be passed to llama_chat_apply_template
-    std::vector<std::string> str(messages.size() * 2);
-    std::vector<llama_chat_message> chat(messages.size());
 
     for (size_t i = 0; i < messages.size(); ++i) {
         const auto & curr_msg = messages[i];
-        str[i*2 + 0]    = json_value(curr_msg, "role",    std::string(""));
+        std::string role    = json_value(curr_msg, "role",    std::string(""));
-        str[i*2 + 1]    = json_value(curr_msg, "content", std::string(""));
+        std::string content = json_value(curr_msg, "content", std::string(""));
-        alloc_size     += str[i*2 + 1].length();
+        chat.push_back({role, content});
-        chat[i].role    = str[i*2 + 0].c_str();
-        chat[i].content = str[i*2 + 1].c_str();
     }
 
-    const char * ptr_tmpl = tmpl.empty() ? nullptr : tmpl.c_str();
+    auto formatted_chat = llama_chat_apply_template(model, tmpl, chat, true);
-    std::vector<char> buf(alloc_size * 2);
-
-    // run the first time to get the total output length
-    int32_t res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size());
-
-    // if it turns out that our buffer is too small, we resize it
-    if ((size_t) res > buf.size()) {
-        buf.resize(res);
-        res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size());
-    }
-
-    const std::string formatted_chat(buf.data(), res);
-
     LOG_VERBOSE("formatted_chat", {{"text", formatted_chat.c_str()}});
-
     return formatted_chat;
 }
 

ggml-cuda.cu

@@ -1924,16 +1924,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     } else if (!split && any_gpus_with_slow_fp16 && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
         // FP32 precision KQV single-batch for batch size 1 without FlashAttention
         ggml_cuda_mul_mat_vec_nc(ctx, src0, src1, dst);
+    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16)
+               && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+        // KQ + KQV multi-batch without FlashAttention
+        ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
     } else if (use_dequantize_mul_mat_vec) {
         ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, nullptr);
     } else if (use_mul_mat_vec_q) {
         ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, quantize_row_q8_1_cuda);
     } else if (use_mul_mat_q) {
         ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_q, quantize_mmq_q8_1_cuda);
-    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16)
-               && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
-        // KQ + KQV multi-batch without FlashAttention
-        ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
     } else {
         ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_cublas, nullptr);
     }

ggml-sycl.cpp

@@ -4620,7 +4620,7 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
     } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
         // KQV single-batch
         ggml_sycl_mul_mat_vec_nc(ctx, src0, src1, dst);
-    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16) && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
         // KQ + KQV multi-batch
         ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
     } else if (use_dequantize_mul_mat_vec) {

gguf-py/gguf/gguf_reader.py

@@ -69,6 +69,7 @@ class GGUFReader:
     # I - same as host, S - swapped
     byte_order: Literal['I'] | Literal['S'] = 'I'
     alignment: int = GGUF_DEFAULT_ALIGNMENT
+    data_offset: int
 
     # Note: Internal helper, API may change.
     gguf_scalar_to_np: dict[GGUFValueType, type[np.generic]] = {
@@ -88,9 +89,13 @@ class GGUFReader:
     def __init__(self, path: os.PathLike[str] | str, mode: Literal['r'] | Literal['r+'] | Literal['c'] = 'r'):
         self.data = np.memmap(path, mode = mode)
         offs = 0
+
+        # Check for GGUF magic
         if self._get(offs, np.uint32, override_order = '<')[0] != GGUF_MAGIC:
             raise ValueError('GGUF magic invalid')
         offs += 4
+
+        # Check GGUF version
         temp_version = self._get(offs, np.uint32)
         if temp_version[0] & 65535 == 0:
             # If we get 0 here that means it's (probably) a GGUF file created for
@@ -103,12 +108,16 @@ class GGUFReader:
         self.fields: OrderedDict[str, ReaderField] = OrderedDict()
         self.tensors: list[ReaderTensor] = []
         offs += self._push_field(ReaderField(offs, 'GGUF.version', [temp_version], [0], [GGUFValueType.UINT32]))
+
+        # Check tensor count and kv count
         temp_counts = self._get(offs, np.uint64, 2)
         offs += self._push_field(ReaderField(offs, 'GGUF.tensor_count', [temp_counts[:1]], [0], [GGUFValueType.UINT64]))
         offs += self._push_field(ReaderField(offs, 'GGUF.kv_count', [temp_counts[1:]], [0], [GGUFValueType.UINT64]))
         tensor_count, kv_count = temp_counts
         offs = self._build_fields(offs, kv_count)
-        offs, tensors_fields = self._build_tensors_fields(offs, tensor_count)
+
+        # Build Tensor Info Fields
+        offs, tensors_fields = self._build_tensor_info(offs, tensor_count)
         new_align = self.fields.get('general.alignment')
         if new_align is not None:
             if new_align.types != [GGUFValueType.UINT32]:
@@ -117,6 +126,7 @@ class GGUFReader:
         padding = offs % self.alignment
         if padding != 0:
             offs += self.alignment - padding
+        self.data_offset = offs
         self._build_tensors(offs, tensors_fields)
 
     _DT = TypeVar('_DT', bound = npt.DTypeLike)
@@ -193,18 +203,29 @@ class GGUFReader:
         # We can't deal with this one.
         raise ValueError('Unknown/unhandled field type {gtype}')
 
-    def _get_tensor(self, orig_offs: int) -> ReaderField:
+    def _get_tensor_info_field(self, orig_offs: int) -> ReaderField:
         offs = orig_offs
+
+        # Get Tensor Name
         name_len, name_data = self._get_str(offs)
         offs += int(name_len.nbytes + name_data.nbytes)
+
+        # Get Tensor Dimensions Count
         n_dims = self._get(offs, np.uint32)
         offs += int(n_dims.nbytes)
+
+        # Get Tensor Dimension Array
         dims = self._get(offs, np.uint64, n_dims[0])
         offs += int(dims.nbytes)
+
+        # Get Tensor Encoding Scheme Type
         raw_dtype = self._get(offs, np.uint32)
         offs += int(raw_dtype.nbytes)
+
+        # Get Tensor Offset
         offset_tensor = self._get(offs, np.uint64)
         offs += int(offset_tensor.nbytes)
+
         return ReaderField(
             orig_offs,
             str(bytes(name_data), encoding = 'utf-8'),
@@ -233,10 +254,10 @@ class GGUFReader:
             offs += field_size
         return offs
 
-    def _build_tensors_fields(self, offs: int, count: int) -> tuple[int, list[ReaderField]]:
+    def _build_tensor_info(self, offs: int, count: int) -> tuple[int, list[ReaderField]]:
         tensor_fields = []
         for _ in range(count):
-            field = self._get_tensor(offs)
+            field = self._get_tensor_info_field(offs)
             offs += sum(int(part.nbytes) for part in field.parts)
             tensor_fields.append(field)
         return offs, tensor_fields

gguf-py/scripts/gguf-dump.py

@@ -319,6 +319,27 @@ def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None
 
         markdown_content += "\n"
 
+        markdown_content += "### Tensor Data Offset\n"
+        markdown_content += '\n'
+        markdown_content += 'This table contains the offset and data segment relative to start of file\n'
+        markdown_content += '\n'
+
+        tensor_mapping_table: list[dict[str, str | int]] = []
+        for key, tensor in enumerate(reader.tensors):
+            data_offset_pretty = '{0:#16x}'.format(tensor.data_offset)
+            data_size_pretty = '{0:#16x}'.format(tensor.n_bytes)
+            tensor_mapping_table.append({"t_id":key, "layer_name":tensor.name, "data_offset":data_offset_pretty, "data_size":data_size_pretty})
+
+        tensors_mapping_table_header_map = [
+            {'key_name':'t_id',         'header_name':'T_ID',               'align':'right'},
+            {'key_name':'layer_name',   'header_name':'Tensor Layer Name',  'align':'left'},
+            {'key_name':'data_offset',  'header_name':'Data Offset (B)',    'align':'right'},
+            {'key_name':'data_size',    'header_name':'Data Size (B)',      'align':'right'},
+        ]
+
+        markdown_content += markdown_table_with_alignment_support(tensors_mapping_table_header_map, tensor_mapping_table)
+        markdown_content += "\n"
+
         for group in tensor_prefix_order:
             tensors = tensor_groups[group]
             group_elements = sum(tensor.n_elements for tensor in tensors)
@@ -370,6 +391,8 @@ def main() -> None:
     parser.add_argument("--no-tensors", action="store_true", help="Don't dump tensor metadata")
     parser.add_argument("--json",       action="store_true", help="Produce JSON output")
     parser.add_argument("--json-array", action="store_true", help="Include full array values in JSON output (long)")
+    parser.add_argument("--data-offset",    action="store_true", help="Start of data offset")
+    parser.add_argument("--data-alignment", action="store_true", help="Data alignment applied globally to data field")
     parser.add_argument("--markdown",   action="store_true", help="Produce markdown output")
     parser.add_argument("--verbose",    action="store_true", help="increase output verbosity")
 
@@ -377,7 +400,7 @@ def main() -> None:
 
     logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
 
-    if not args.json and not args.markdown:
+    if not args.json and not args.markdown and not args.data_offset and not args.data_alignment:
         logger.info(f'* Loading: {args.model}')
 
     reader = GGUFReader(args.model, 'r')
@@ -386,6 +409,10 @@ def main() -> None:
         dump_metadata_json(reader, args)
     elif args.markdown:
         dump_markdown_metadata(reader, args)
+    elif args.data_offset:
+        print(reader.data_offset)  # noqa: NP100
+    elif args.data_alignment:
+        print(reader.alignment)  # noqa: NP100
     else:
         dump_metadata(reader, args)
 

llama.cpp

@@ -226,6 +226,7 @@ enum llm_arch {
     LLM_ARCH_ARCTIC,
     LLM_ARCH_DEEPSEEK2,
     LLM_ARCH_BITNET,
+    LLM_ARCH_T5,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -265,6 +266,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_ARCTIC,          "arctic"       },
     { LLM_ARCH_DEEPSEEK2,       "deepseek2"    },
     { LLM_ARCH_BITNET,          "bitnet"       },
+    { LLM_ARCH_T5,              "t5"           },
     { LLM_ARCH_UNKNOWN,         "(unknown)"    },
 };
 
@@ -297,6 +299,7 @@ enum llm_kv {
     LLM_KV_EXPERT_WEIGHTS_SCALE,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
+    LLM_KV_DECODER_START_TOKEN_ID,
 
     LLM_KV_ATTENTION_HEAD_COUNT,
     LLM_KV_ATTENTION_HEAD_COUNT_KV,
@@ -309,6 +312,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_CAUSAL,
     LLM_KV_ATTENTION_Q_LORA_RANK,
     LLM_KV_ATTENTION_KV_LORA_RANK,
+    LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_FREQ_BASE,
@@ -346,6 +350,8 @@ enum llm_kv {
     LLM_KV_TOKENIZER_ADD_BOS,
     LLM_KV_TOKENIZER_ADD_EOS,
     LLM_KV_TOKENIZER_ADD_PREFIX,
+    LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,
+    LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,
     LLM_KV_TOKENIZER_HF_JSON,
     LLM_KV_TOKENIZER_RWKV,
     LLM_KV_TOKENIZER_PREFIX_ID,
@@ -383,18 +389,20 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EXPERT_WEIGHTS_SCALE,              "%s.expert_weights_scale"              },
     { LLM_KV_POOLING_TYPE ,                     "%s.pooling_type"                      },
     { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
+    { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
 
-    { LLM_KV_ATTENTION_HEAD_COUNT,          "%s.attention.head_count"             },
+    { LLM_KV_ATTENTION_HEAD_COUNT,             "%s.attention.head_count"             },
-    { LLM_KV_ATTENTION_HEAD_COUNT_KV,       "%s.attention.head_count_kv"          },
+    { LLM_KV_ATTENTION_HEAD_COUNT_KV,          "%s.attention.head_count_kv"          },
-    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,      "%s.attention.max_alibi_bias"         },
+    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,         "%s.attention.max_alibi_bias"         },
-    { LLM_KV_ATTENTION_CLAMP_KQV,           "%s.attention.clamp_kqv"              },
+    { LLM_KV_ATTENTION_CLAMP_KQV,              "%s.attention.clamp_kqv"              },
-    { LLM_KV_ATTENTION_KEY_LENGTH,          "%s.attention.key_length"             },
+    { LLM_KV_ATTENTION_KEY_LENGTH,             "%s.attention.key_length"             },
-    { LLM_KV_ATTENTION_VALUE_LENGTH,        "%s.attention.value_length"           },
+    { LLM_KV_ATTENTION_VALUE_LENGTH,           "%s.attention.value_length"           },
-    { LLM_KV_ATTENTION_LAYERNORM_EPS,       "%s.attention.layer_norm_epsilon"     },
+    { LLM_KV_ATTENTION_LAYERNORM_EPS,          "%s.attention.layer_norm_epsilon"     },
-    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,   "%s.attention.layer_norm_rms_epsilon" },
+    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,      "%s.attention.layer_norm_rms_epsilon" },
-    { LLM_KV_ATTENTION_CAUSAL,              "%s.attention.causal"                 },
+    { LLM_KV_ATTENTION_CAUSAL,                 "%s.attention.causal"                 },
-    { LLM_KV_ATTENTION_Q_LORA_RANK,         "%s.attention.q_lora_rank"            },
+    { LLM_KV_ATTENTION_Q_LORA_RANK,            "%s.attention.q_lora_rank"            },
-    { LLM_KV_ATTENTION_KV_LORA_RANK,        "%s.attention.kv_lora_rank"           },
+    { LLM_KV_ATTENTION_KV_LORA_RANK,           "%s.attention.kv_lora_rank"           },
+    { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
@@ -415,29 +423,31 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_SSM_STATE_SIZE,                "%s.ssm.state_size"     },
     { LLM_KV_SSM_TIME_STEP_RANK,            "%s.ssm.time_step_rank" },
 
-    { LLM_KV_TOKENIZER_MODEL,               "tokenizer.ggml.model"              },
+    { LLM_KV_TOKENIZER_MODEL,                "tokenizer.ggml.model"                    },
-    { LLM_KV_TOKENIZER_PRE,                 "tokenizer.ggml.pre"                },
+    { LLM_KV_TOKENIZER_PRE,                  "tokenizer.ggml.pre"                      },
-    { LLM_KV_TOKENIZER_LIST,                "tokenizer.ggml.tokens"             },
+    { LLM_KV_TOKENIZER_LIST,                 "tokenizer.ggml.tokens"                   },
-    { LLM_KV_TOKENIZER_TOKEN_TYPE,          "tokenizer.ggml.token_type"         },
+    { LLM_KV_TOKENIZER_TOKEN_TYPE,           "tokenizer.ggml.token_type"               },
-    { LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT,    "tokenizer.ggml.token_type_count"   },
+    { LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT,     "tokenizer.ggml.token_type_count"         },
-    { LLM_KV_TOKENIZER_SCORES,              "tokenizer.ggml.scores"             },
+    { LLM_KV_TOKENIZER_SCORES,               "tokenizer.ggml.scores"                   },
-    { LLM_KV_TOKENIZER_MERGES,              "tokenizer.ggml.merges"             },
+    { LLM_KV_TOKENIZER_MERGES,               "tokenizer.ggml.merges"                   },
-    { LLM_KV_TOKENIZER_BOS_ID,              "tokenizer.ggml.bos_token_id"       },
+    { LLM_KV_TOKENIZER_BOS_ID,               "tokenizer.ggml.bos_token_id"             },
-    { LLM_KV_TOKENIZER_EOS_ID,              "tokenizer.ggml.eos_token_id"       },
+    { LLM_KV_TOKENIZER_EOS_ID,               "tokenizer.ggml.eos_token_id"             },
-    { LLM_KV_TOKENIZER_UNK_ID,              "tokenizer.ggml.unknown_token_id"   },
+    { LLM_KV_TOKENIZER_UNK_ID,               "tokenizer.ggml.unknown_token_id"         },
-    { LLM_KV_TOKENIZER_SEP_ID,              "tokenizer.ggml.seperator_token_id" },
+    { LLM_KV_TOKENIZER_SEP_ID,               "tokenizer.ggml.seperator_token_id"       },
-    { LLM_KV_TOKENIZER_PAD_ID,              "tokenizer.ggml.padding_token_id"   },
+    { LLM_KV_TOKENIZER_PAD_ID,               "tokenizer.ggml.padding_token_id"         },
-    { LLM_KV_TOKENIZER_CLS_ID,              "tokenizer.ggml.cls_token_id"       },
+    { LLM_KV_TOKENIZER_CLS_ID,               "tokenizer.ggml.cls_token_id"             },
-    { LLM_KV_TOKENIZER_MASK_ID,             "tokenizer.ggml.mask_token_id"      },
+    { LLM_KV_TOKENIZER_MASK_ID,              "tokenizer.ggml.mask_token_id"            },
-    { LLM_KV_TOKENIZER_ADD_BOS,             "tokenizer.ggml.add_bos_token"      },
+    { LLM_KV_TOKENIZER_ADD_BOS,              "tokenizer.ggml.add_bos_token"            },
-    { LLM_KV_TOKENIZER_ADD_EOS,             "tokenizer.ggml.add_eos_token"      },
+    { LLM_KV_TOKENIZER_ADD_EOS,              "tokenizer.ggml.add_eos_token"            },
-    { LLM_KV_TOKENIZER_ADD_PREFIX,          "tokenizer.ggml.add_space_prefix"   },
+    { LLM_KV_TOKENIZER_ADD_PREFIX,           "tokenizer.ggml.add_space_prefix"         },
-    { LLM_KV_TOKENIZER_HF_JSON,             "tokenizer.huggingface.json"        },
+    { LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,      "tokenizer.ggml.remove_extra_whitespaces" },
-    { LLM_KV_TOKENIZER_RWKV,                "tokenizer.rwkv.world"              },
+    { LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, "tokenizer.ggml.precompiled_charsmap"     },
-    { LLM_KV_TOKENIZER_PREFIX_ID,           "tokenizer.ggml.prefix_token_id"    },
+    { LLM_KV_TOKENIZER_HF_JSON,              "tokenizer.huggingface.json"              },
-    { LLM_KV_TOKENIZER_SUFFIX_ID,           "tokenizer.ggml.suffix_token_id"    },
+    { LLM_KV_TOKENIZER_RWKV,                 "tokenizer.rwkv.world"                    },
-    { LLM_KV_TOKENIZER_MIDDLE_ID,           "tokenizer.ggml.middle_token_id"    },
+    { LLM_KV_TOKENIZER_PREFIX_ID,            "tokenizer.ggml.prefix_token_id"          },
-    { LLM_KV_TOKENIZER_EOT_ID,              "tokenizer.ggml.eot_token_id"       },
+    { LLM_KV_TOKENIZER_SUFFIX_ID,            "tokenizer.ggml.suffix_token_id"          },
+    { LLM_KV_TOKENIZER_MIDDLE_ID,            "tokenizer.ggml.middle_token_id"          },
+    { LLM_KV_TOKENIZER_EOT_ID,               "tokenizer.ggml.eot_token_id"             },
 };
 
 struct LLM_KV {
@@ -504,6 +514,34 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_KV_A_NORM,
     LLM_TENSOR_ATTN_SUB_NORM,
     LLM_TENSOR_FFN_SUB_NORM,
+    LLM_TENSOR_DEC_ATTN_NORM,
+    LLM_TENSOR_DEC_ATTN_Q,
+    LLM_TENSOR_DEC_ATTN_K,
+    LLM_TENSOR_DEC_ATTN_V,
+    LLM_TENSOR_DEC_ATTN_OUT,
+    LLM_TENSOR_DEC_ATTN_REL_B,
+    LLM_TENSOR_DEC_CROSS_ATTN_NORM,
+    LLM_TENSOR_DEC_CROSS_ATTN_Q,
+    LLM_TENSOR_DEC_CROSS_ATTN_K,
+    LLM_TENSOR_DEC_CROSS_ATTN_V,
+    LLM_TENSOR_DEC_CROSS_ATTN_OUT,
+    LLM_TENSOR_DEC_CROSS_ATTN_REL_B,
+    LLM_TENSOR_DEC_FFN_NORM,
+    LLM_TENSOR_DEC_FFN_GATE,
+    LLM_TENSOR_DEC_FFN_DOWN,
+    LLM_TENSOR_DEC_FFN_UP,
+    LLM_TENSOR_DEC_OUTPUT_NORM,
+    LLM_TENSOR_ENC_ATTN_NORM,
+    LLM_TENSOR_ENC_ATTN_Q,
+    LLM_TENSOR_ENC_ATTN_K,
+    LLM_TENSOR_ENC_ATTN_V,
+    LLM_TENSOR_ENC_ATTN_OUT,
+    LLM_TENSOR_ENC_ATTN_REL_B,
+    LLM_TENSOR_ENC_FFN_NORM,
+    LLM_TENSOR_ENC_FFN_GATE,
+    LLM_TENSOR_ENC_FFN_DOWN,
+    LLM_TENSOR_ENC_FFN_UP,
+    LLM_TENSOR_ENC_OUTPUT_NORM,
 };
 
 static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES = {
@@ -1135,6 +1173,41 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_FFN_SUB_NORM,       "blk.%d.ffn_sub_norm" },
         },
     },
+    {
+        LLM_ARCH_T5,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,           "token_embd" },
+            { LLM_TENSOR_OUTPUT,               "output" },
+            { LLM_TENSOR_DEC_OUTPUT_NORM,      "dec.output_norm" },
+            { LLM_TENSOR_DEC_ATTN_NORM,        "dec.blk.%d.attn_norm" },
+            { LLM_TENSOR_DEC_ATTN_Q,           "dec.blk.%d.attn_q" },
+            { LLM_TENSOR_DEC_ATTN_K,           "dec.blk.%d.attn_k" },
+            { LLM_TENSOR_DEC_ATTN_V,           "dec.blk.%d.attn_v" },
+            { LLM_TENSOR_DEC_ATTN_OUT,         "dec.blk.%d.attn_o" },
+            { LLM_TENSOR_DEC_ATTN_REL_B,       "dec.blk.%d.attn_rel_b" },
+            { LLM_TENSOR_DEC_CROSS_ATTN_NORM,  "dec.blk.%d.cross_attn_norm" },
+            { LLM_TENSOR_DEC_CROSS_ATTN_Q,     "dec.blk.%d.cross_attn_q" },
+            { LLM_TENSOR_DEC_CROSS_ATTN_K,     "dec.blk.%d.cross_attn_k" },
+            { LLM_TENSOR_DEC_CROSS_ATTN_V,     "dec.blk.%d.cross_attn_v" },
+            { LLM_TENSOR_DEC_CROSS_ATTN_OUT,   "dec.blk.%d.cross_attn_o" },
+            { LLM_TENSOR_DEC_CROSS_ATTN_REL_B, "dec.blk.%d.cross_attn_rel_b" },
+            { LLM_TENSOR_DEC_FFN_NORM,         "dec.blk.%d.ffn_norm" },
+            { LLM_TENSOR_DEC_FFN_GATE,         "dec.blk.%d.ffn_gate" },
+            { LLM_TENSOR_DEC_FFN_DOWN,         "dec.blk.%d.ffn_down" },
+            { LLM_TENSOR_DEC_FFN_UP,           "dec.blk.%d.ffn_up" },
+            { LLM_TENSOR_ENC_OUTPUT_NORM,      "enc.output_norm" },
+            { LLM_TENSOR_ENC_ATTN_NORM,        "enc.blk.%d.attn_norm" },
+            { LLM_TENSOR_ENC_ATTN_Q,           "enc.blk.%d.attn_q" },
+            { LLM_TENSOR_ENC_ATTN_K,           "enc.blk.%d.attn_k" },
+            { LLM_TENSOR_ENC_ATTN_V,           "enc.blk.%d.attn_v" },
+            { LLM_TENSOR_ENC_ATTN_OUT,         "enc.blk.%d.attn_o" },
+            { LLM_TENSOR_ENC_ATTN_REL_B,       "enc.blk.%d.attn_rel_b" },
+            { LLM_TENSOR_ENC_FFN_NORM,         "enc.blk.%d.ffn_norm" },
+            { LLM_TENSOR_ENC_FFN_GATE,         "enc.blk.%d.ffn_gate" },
+            { LLM_TENSOR_ENC_FFN_DOWN,         "enc.blk.%d.ffn_down" },
+            { LLM_TENSOR_ENC_FFN_UP,           "enc.blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -2356,6 +2429,11 @@ struct llama_vocab {
     bool tokenizer_add_bos          = false;
     bool tokenizer_add_eos          = false;
     bool tokenizer_ignore_merges    = false;
+    bool tokenizer_remove_extra_whitespaces   = false;
+    bool tokenizer_escape_whitespaces         = true;
+    bool tokenizer_treat_whitespace_as_suffix = false;
+
+    std::vector<char> precompiled_charsmap;
 
     int find_bpe_rank(const std::string & token_left, const std::string & token_right) const {
         GGML_ASSERT(token_left.find(' ') == std::string::npos);
@@ -4191,6 +4269,7 @@ static const char * llama_model_vocab_type_name(enum llama_vocab_type type){
         case LLAMA_VOCAB_TYPE_SPM:  return "SPM";
         case LLAMA_VOCAB_TYPE_BPE:  return "BPE";
         case LLAMA_VOCAB_TYPE_WPM:  return "WPM";
+        case LLAMA_VOCAB_TYPE_UGM:  return "UGM";
         default:                    return "unknown";
     }
 }
@@ -4870,6 +4949,45 @@ static void llm_load_vocab(
             vocab.special_pad_id  = -1;
             vocab.special_cls_id  = -1;
             vocab.special_mask_id = -1;
+        } else if (tokenizer_model == "t5") {
+            vocab.type = LLAMA_VOCAB_TYPE_UGM;
+
+            // default special tokens
+            vocab.special_bos_id  = -1;
+            vocab.special_eos_id  = 1;
+            vocab.special_unk_id  = 2;
+            vocab.special_sep_id  = -1;
+            vocab.special_pad_id  = 0;
+            vocab.special_cls_id  = -1;
+            vocab.special_mask_id = -1;
+
+            const int add_space_prefix_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_ADD_PREFIX).c_str());
+            if (add_space_prefix_keyidx != -1) {
+                vocab.tokenizer_add_space_prefix = gguf_get_val_bool(ctx, add_space_prefix_keyidx);
+            } // The default value of add_space_prefix is true.
+
+            const int remove_extra_whitespaces_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_REMOVE_EXTRA_WS).c_str());
+            if (remove_extra_whitespaces_keyidx != -1) {
+                vocab.tokenizer_remove_extra_whitespaces = gguf_get_val_bool(ctx, remove_extra_whitespaces_keyidx);
+            } // The default value of remove_extra_whitespaces is false.
+
+            const int precompiled_charsmap_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP).c_str());
+            if (precompiled_charsmap_keyidx != -1) {
+                size_t n_precompiled_charsmap = gguf_get_arr_n(ctx, precompiled_charsmap_keyidx);
+                const char * precompiled_charsmap = (const char *) gguf_get_arr_data(ctx, precompiled_charsmap_keyidx);
+                vocab.precompiled_charsmap.assign(precompiled_charsmap, precompiled_charsmap + n_precompiled_charsmap);
+#ifdef IS_BIG_ENDIAN
+                // correct endiannes of data in precompiled_charsmap binary blob
+                uint32_t * xcda_blob_size = (uint32_t *) &vocab.precompiled_charsmap[0];
+                *xcda_blob_size = __builtin_bswap32(*xcda_blob_size);
+                assert(*xcda_blob_size + sizeof(uint32_t) < n_precompiled_charsmap);
+                size_t xcda_array_size = *xcda_blob_size / sizeof(uint32_t);
+                uint32_t * xcda_array = (uint32_t *) &vocab.precompiled_charsmap[sizeof(uint32_t)];
+                for (size_t i = 0; i < xcda_array_size; ++i) {
+                    xcda_array[i] = __builtin_bswap32(xcda_array[i]);
+                }
+#endif
+            }
         } else {
             throw std::runtime_error(format("unknown tokenizer: '%s'", tokenizer_model.c_str()));
         }
@@ -4952,6 +5070,10 @@ static void llm_load_vocab(
             vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
             vocab.tokenizer_add_bos = true;
             vocab.tokenizer_add_eos = false;
+        } else if (vocab.type == LLAMA_VOCAB_TYPE_UGM) {
+            vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
+            vocab.tokenizer_add_bos = false;
+            vocab.tokenizer_add_eos = true;
         } else {
             vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
         }
@@ -13213,12 +13335,18 @@ static bool llama_is_user_defined_token(const llama_vocab& vocab, llama_token id
     return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_USER_DEFINED;
 }
 
+static bool llama_is_unused_token(const llama_vocab& vocab, llama_token id) {
+    GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
+    return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_UNUSED;
+}
+
 static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) {
     GGML_ASSERT(llama_vocab_get_type(vocab) != LLAMA_VOCAB_TYPE_NONE);
     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: {
+        case LLAMA_VOCAB_TYPE_SPM:
+        case LLAMA_VOCAB_TYPE_UGM: {
             auto buf = token_data.text.substr(3, 2);
             return strtol(buf.c_str(), NULL, 16);
         }
@@ -13238,7 +13366,8 @@ static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) {
     GGML_ASSERT(llama_vocab_get_type(vocab) != LLAMA_VOCAB_TYPE_NONE);
     static const char * hex = "0123456789ABCDEF";
     switch (llama_vocab_get_type(vocab)) {
-        case LLAMA_VOCAB_TYPE_SPM: {
+        case LLAMA_VOCAB_TYPE_SPM:
+        case LLAMA_VOCAB_TYPE_UGM: {
             const char buf[7] = { '<', '0', 'x', hex[ch >> 4], hex[ch & 15], '>', 0 };
             auto token = vocab.token_to_id.find(buf);
             if (token != vocab.token_to_id.end()) {
@@ -13826,6 +13955,383 @@ struct llm_tokenizer_wpm {
     const llama_vocab & vocab;
 };
 
+struct naive_trie {
+    naive_trie() : has_value(false), value(0) {
+    }
+    void insert(const char * key, size_t len, int32_t value = 0) {
+        if (len == 0) {
+            this->has_value = true;
+            this->value = value;
+            return;
+        }
+        char c = key[0];
+        auto res = children.find(c);
+        if (res != children.end()) {
+            res->second.insert(key + 1, len - 1, value);
+        } else {
+            auto res = children.insert(std::make_pair(c, naive_trie()));
+            res.first->second.insert(key + 1, len - 1, value);
+        }
+    }
+    std::pair<const char *, size_t> get_longest_prefix(const char * key, size_t len, size_t offset = 0) {
+        if (len == 0 || offset == len) {
+            return std::make_pair(key, offset);
+        }
+        char c = key[offset];
+        auto res = children.find(c);
+        if (res != children.end()) {
+            return res->second.get_longest_prefix(key, len, offset + 1);
+        } else {
+            return std::make_pair(key, offset);
+        }
+    }
+    struct naive_trie * traverse(const char c) {
+        auto res = children.find(c);
+        if (res != children.end()) {
+            return &res->second;
+        } else {
+            return NULL;
+        }
+    }
+    std::map<char, struct naive_trie> children;
+    bool has_value;
+    llama_token value;
+};
+
+struct llm_tokenizer_ugm {
+    llm_tokenizer_ugm(const llama_vocab & vocab) : vocab(vocab) {
+        if (vocab.precompiled_charsmap.size() > 0) {
+            size_t charsmap_offset = 0;
+
+            // First four bytes of precompiled_charsmap contains length of binary
+            // blob containing XOR-compressed compact double array (XCDA) entries
+            uint32_t xcda_blob_size = *(const uint32_t *) &vocab.precompiled_charsmap[0];
+            charsmap_offset += sizeof(xcda_blob_size);
+            if (xcda_blob_size + charsmap_offset >= vocab.precompiled_charsmap.size()) {
+                throw std::runtime_error("Index out of array bounds in precompiled charsmap!");
+            }
+
+            // Next xcda_blob_size bytes contain entries of XOR-compressed compact
+            // double array (XCDA). Each entry is bit-packed into a 32-bit integer.
+            xcda_array = (const uint32_t *) &vocab.precompiled_charsmap[charsmap_offset];
+            xcda_array_size = xcda_blob_size / sizeof(uint32_t);
+            charsmap_offset += xcda_blob_size;
+
+            // Remaining bytes of precompiled charsmap contain null-terminated
+            // replacement strings for prefixes matched by the XCDA.
+            prefix_replacements = &vocab.precompiled_charsmap[charsmap_offset];
+            prefix_replacements_size = vocab.precompiled_charsmap.size() - charsmap_offset;
+        }
+
+        for (unsigned int id = 0; id < vocab.id_to_token.size(); ++id) {
+            const auto &token_data = vocab.id_to_token[id];
+
+            if (llama_is_normal_token(vocab, id)) {
+                min_score = std::min<float>(min_score, token_data.score);
+                max_score = std::max<float>(max_score, token_data.score);
+            }
+
+            if (llama_is_normal_token(vocab, id) ||
+                llama_is_user_defined_token(vocab, id) ||
+                llama_is_unused_token(vocab, id)) {
+                token_matcher.insert(token_data.text.data(), token_data.text.size(), id);
+            }
+
+            if (llama_is_user_defined_token(vocab, id)) {
+                user_defined_token_matcher.insert(token_data.text.data(), token_data.text.size());
+            }
+        }
+
+        unknown_token_score = min_score - unknown_token_score_penalty;
+    }
+
+    /* This implementation is based on SentencePiece optimized Viterbi algorithm for
+     * unigram language models. The general idea is to:
+     * - move along the input sequence in steps of one UTF code point,
+     * - at each step find all possible tokenizations of the prefix by
+     *   traversing the tokens trie,
+     * - for each tokenization store the best one so far (by higher score)
+     * - use the position in sequence after given token as an index to store
+     *   results
+     * - if there was no valid tokenization of the current UTF code point
+     *   then use unknown token with additional score penalty
+     * After processing the whole sequence we backtrack from the end to get
+     * the best tokenization.
+    */
+    void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
+        // normalize the input first
+        std::string normalized;
+        normalize(text, &normalized);
+        size_t input_len = normalized.size();
+
+        // initialize score_sum to -FLT_MAX so it will be always lower than sums of token scores
+        std::vector<struct best_tokenization> tokenization_results(input_len + 1, {0, 0, -FLT_MAX});
+        // at the beginning tokenization score is zero
+        tokenization_results[0] = { 0, 0, 0 };
+
+        for (size_t input_offset = 0; input_offset < input_len;) {
+            size_t prefix_offset = input_offset;
+            // calculate how many code units are in the currently processed UTF code point
+            size_t n_utf8_code_units = std::min<size_t>(utf8_len(normalized[input_offset]), input_len - input_offset);
+
+            // traverse the token matcher trie to find a matching token
+            bool single_codepoint_token_found = false;
+            const struct best_tokenization & current_best = tokenization_results[input_offset];
+            struct naive_trie * node  = token_matcher.traverse(normalized[prefix_offset++]);
+
+            while (prefix_offset <= input_len && node != NULL) {
+                // check if we found valid token in prefix
+                if (node->has_value) {
+                    // check if it corresponds to the whole UTF code point
+                    if (prefix_offset - input_offset == n_utf8_code_units) {
+                        single_codepoint_token_found = true;
+                    }
+                    llama_token token_id = node->value;
+                    const auto &token_data = vocab.id_to_token[token_id];
+
+                    // we set the user-defined token scores to 0 to make them more likely to be selected
+                    // (normal token scores are log probabilities, so they are negative)
+                    // score type is double here to make tokenization results exactly
+                    // the same as in the HF tokenizer using SentencePiece
+                    const double token_score = llama_is_user_defined_token(vocab, token_id) ? 0.0 : token_data.score;
+                    const double challenger_score = current_best.score_sum + token_score;
+                    struct best_tokenization & current_champ = tokenization_results[prefix_offset];
+                    if (challenger_score > current_champ.score_sum) {
+                        struct best_tokenization challenger = { token_id, input_offset, (float) challenger_score };
+                        current_champ = challenger;
+                    }
+                }
+                node = node->traverse(normalized[prefix_offset++]);
+            }
+
+            // if we didn't find a valid token corresponding to the whole UTF code point
+            // then use unknown token as the tokenization of this UTF code point
+            if (!single_codepoint_token_found) {
+                const double challenger_score = current_best.score_sum + unknown_token_score;
+                prefix_offset = input_offset + n_utf8_code_units;
+                struct best_tokenization & current_champ = tokenization_results[prefix_offset];
+                if (challenger_score > current_champ.score_sum) {
+                    struct best_tokenization challenger = { vocab.special_unk_id, input_offset, (float) challenger_score };
+                    current_champ = challenger;
+                }
+            }
+
+            // move to the next UTF code point
+            input_offset += n_utf8_code_units;
+        }
+
+        // now backtrack from the end to gather token ids of the best tokenization
+        // merge sequences of consecutive unknown tokens into single unknown tokens
+        bool is_prev_unknown = false;
+        for (struct best_tokenization & tokenization = tokenization_results[input_len]; ; tokenization = tokenization_results[tokenization.input_offset]) {
+            bool is_unknown = tokenization.token_id == vocab.special_unk_id;
+            if (!(is_prev_unknown && is_unknown)) {
+                output.push_back(tokenization.token_id);
+            }
+            if (tokenization.input_offset == 0) {
+                break;
+            }
+            is_prev_unknown = is_unknown;
+        }
+
+        // reverse the output since we added tokens starting from the end of the input
+        std::reverse(output.begin(), output.end());
+    }
+
+private:
+    const llama_vocab & vocab;
+
+    // helper structure for returning normalization results
+    struct normalization_result {
+        const char * normalized;
+        size_t normalized_len;
+        size_t consumed_input;
+    };
+
+    void normalize(const std::string& input, std::string * normalized) {
+        normalized->clear();
+        normalized->reserve(input.size() * 3);
+
+        const std::string space = vocab.tokenizer_escape_whitespaces ? escaped_space : " ";
+
+        bool shall_prepend_space = !vocab.tokenizer_treat_whitespace_as_suffix && vocab.tokenizer_add_space_prefix;
+        bool shall_append_space = vocab.tokenizer_treat_whitespace_as_suffix && vocab.tokenizer_add_space_prefix;
+        bool shall_merge_spaces = vocab.tokenizer_remove_extra_whitespaces;
+
+        bool is_space_prepended = false;
+        bool processing_non_ws = false;
+
+        size_t input_len = input.size();
+
+        for (size_t input_offset = 0; input_offset < input_len; ) {
+            auto norm_res = normalize_prefix(input, input_offset);
+            for (size_t i = 0; i < norm_res.normalized_len; i++) {
+                char c = norm_res.normalized[i];
+                if (c != ' ') {
+                    if (!processing_non_ws) {
+                        processing_non_ws = true;
+                        if ((shall_prepend_space && !is_space_prepended) || shall_merge_spaces) {
+                            normalized->append(space);
+                            is_space_prepended = true;
+                        }
+                    }
+                    normalized->push_back(c);
+                } else {
+                    if (processing_non_ws) {
+                        processing_non_ws = false;
+                    }
+                    if (!shall_merge_spaces) {
+                        normalized->append(space);
+                    }
+                }
+            }
+
+            input_offset += norm_res.consumed_input;
+        }
+
+        if (shall_append_space) {
+            normalized->append(space);
+        }
+    }
+
+    /*
+     * This structure is a view wrapper for XOR-compressed double array (XCDA)
+     * See Shunsuke Kanda (2018). Space- and Time-Efficient String Dictionaries.
+     * Eeach bit-packed entry contains:
+     * - BASE array value in bits 10-30
+     * - LCHECK array value in bits 0-7
+     * - LEAF array value in bit 9
+     * Entries containing indexes of replacement sequences have set bit 31
+     */
+    struct xcda_array_view {
+    public:
+        xcda_array_view(const uint32_t * xcda_array, size_t xcda_array_size) : xcda_array(xcda_array), xcda_array_size(xcda_array_size) {
+        }
+        uint32_t get_base(size_t index) {
+            uint32_t packed_node = get_node(index);
+            return (packed_node >> 10) << ((packed_node & (1U << 9)) >> 6);
+        }
+        uint32_t get_lcheck(size_t index) {
+            uint32_t packed_node = get_node(index);
+            return packed_node & ((1U << 31) | 0xff);
+        }
+        bool get_leaf(size_t index) {
+            uint32_t packed_node = get_node(index);
+            return (packed_node >> 8) & 1;
+        }
+        uint32_t get_value(size_t index) {
+            uint32_t packed_node = get_node(index);
+            return packed_node & ((1U << 31) - 1);
+        }
+    private:
+        uint32_t get_node(size_t index) {
+            if (index > xcda_array_size) {
+                throw std::runtime_error("Index out of array bounds in XCDA array!");
+            }
+            return xcda_array[index];
+        }
+        const uint32_t * xcda_array;
+        size_t xcda_array_size;
+    };
+
+    struct normalization_result normalize_prefix(const std::string & input, size_t input_offset) {
+        if (input_offset == input.size()) {
+            return { &input[input_offset], 0, 0 };
+        }
+
+        // if input prefix matches some user-defined token return this token as normalization result
+        auto user_defined_token_match = user_defined_token_matcher.get_longest_prefix(&input[input_offset], input.size() - input_offset);
+        if (user_defined_token_match.second > 0) {
+            return { &input[input_offset], user_defined_token_match.second, user_defined_token_match.second };
+        }
+
+        size_t longest_prefix_length = 0;
+        size_t longest_prefix_offset = 0;
+
+        if (xcda_array_size > 0) {
+            struct xcda_array_view xcda_view(xcda_array, xcda_array_size);
+
+            // Find the longest normalized sequence matching the input prefix by walking
+            // the XOR-compressed compact double array (XCDA) starting from the root node
+            // We find the index of the next node by calculating BASE[s] ^ c where s is
+            // the index of the previous node and c is a numerical character value
+            uint32_t node_index = 0;
+            // get BASE of the root node
+            node_index = xcda_view.get_base(node_index);
+            for (size_t prefix_offset = input_offset; prefix_offset < input.size(); prefix_offset++) {
+                unsigned char c = input[prefix_offset];
+                if (c == 0) {
+                    break;
+                }
+                node_index ^= c;
+                // if value of LCHECK is not c it means that this is not a child of
+                // the previous node, so we stop matching
+                if (xcda_view.get_lcheck(node_index) != c) {
+                    break;
+                }
+                bool is_leaf = xcda_view.get_leaf(node_index);
+                // get BASE of the current node
+                node_index ^= xcda_view.get_base(node_index);
+                // if LEAF of the current node is true, it means that its BASE points to the node
+                // containing index of replacement sequence for currently matched input prefix
+                if (is_leaf)
+                {
+                    longest_prefix_length = prefix_offset - input_offset + 1;
+                    // get index of replacement sequence for currently matched input prefix
+                    longest_prefix_offset = xcda_view.get_value(node_index);
+                }
+            }
+        }
+
+        if (longest_prefix_length > 0) {
+            // we have a match, so return the replacement sequence
+            if (longest_prefix_offset >= prefix_replacements_size) {
+                throw std::runtime_error("Index out of array bounds in precompiled charsmap!");
+            }
+            const char * prefix_replacement = &prefix_replacements[longest_prefix_offset];
+            return { prefix_replacement, strlen(prefix_replacement), longest_prefix_length };
+        } else {
+            // check if the input prefix contains a valid sequence of UTF-8 code units
+            try {
+                // if yes, return this sequence unmodified
+                size_t prefix_offset = input_offset;
+                unicode_cpt_from_utf8(input, prefix_offset);
+                return { &input[input_offset], prefix_offset - input_offset, prefix_offset - input_offset };
+            } catch(std::invalid_argument & ex) {
+                // if no, consume 1 byte and return U+FFFD - REPLACEMENT CHARACTER
+                return { "\xEF\xBF\xBD", 3, 1 };
+            }
+        }
+    }
+
+    // escaped space symbol - U+2581 (Lower One Eighth Block)
+    const std::string escaped_space = "\xE2\x96\x81";
+
+    const char * prefix_replacements = NULL;
+    size_t prefix_replacements_size = 0;
+
+    const uint32_t * xcda_array = NULL;
+    size_t xcda_array_size = 0;
+
+    struct naive_trie user_defined_token_matcher;
+
+    // this structure stores the best tokenization so far at input_offset
+    struct best_tokenization {
+        llama_token token_id;
+        size_t input_offset;
+        float score_sum;
+    };
+
+    float min_score = FLT_MAX;
+    float max_score = -FLT_MAX;
+
+    float unknown_token_score_penalty = 10.0;
+    float unknown_token_score;
+
+    struct naive_trie token_matcher;
+};
+
+
 typedef enum FRAGMENT_BUFFER_VARIANT_TYPE {
     FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN,
     FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT
@@ -14086,6 +14592,39 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
                     output.push_back(vocab.special_sep_id);
                 }
             } break;
+        case LLAMA_VOCAB_TYPE_UGM:
+            {
+                llm_tokenizer_ugm tokenizer(vocab);
+
+                if (add_special && vocab.tokenizer_add_bos != 0) {
+                    GGML_ASSERT(vocab.special_bos_id != -1);
+                    output.push_back(vocab.special_bos_id);
+                }
+
+                for (const auto & fragment : fragment_buffer) {
+                    if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
+                        auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
+#ifdef PRETOKENIZERDEBUG
+                        LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
+#endif
+                        tokenizer.tokenize(raw_text, output);
+                    } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
+                        output.push_back(fragment.token);
+                    }
+                }
+
+                if (add_special && vocab.tokenizer_add_bos != 0 && output.size() >= 2 && output[1] == vocab.special_bos_id) {
+                    LLAMA_LOG_WARN(
+                        "%s: Added a BOS token to the prompt as specified by the model but the prompt "
+                        "also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. "
+                        "Are you sure this is what you want?\n", __FUNCTION__);
+                }
+
+                if (add_special && vocab.tokenizer_add_eos == 1) {
+                    GGML_ASSERT(vocab.special_eos_id != -1);
+                    output.push_back(vocab.special_eos_id);
+                }
+            } break;
         case LLAMA_VOCAB_TYPE_NONE:
             GGML_ASSERT(false);
     }
@@ -14500,7 +15039,8 @@ struct llama_grammar * llama_grammar_init(
             continue;
         }
         if (llama_grammar_detect_left_recursion(vec_rules, i, &rules_visited, &rules_in_progress, &rules_may_be_empty)) {
-            throw std::runtime_error(format("unsupported grammar, left recursion detected for nonterminal at index %zu", i));
+            LLAMA_LOG_ERROR("unsupported grammar, left recursion detected for nonterminal at index %zu", i);
+            return nullptr;
         }
     }
 
@@ -16963,6 +17503,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_BLOOM:
         case LLM_ARCH_MAMBA:
         case LLM_ARCH_JINA_BERT_V2:
+        case LLM_ARCH_T5:
             return LLAMA_ROPE_TYPE_NONE;
 
         // use what we call a normal RoPE, operating on pairs of consecutive head values
@@ -18658,6 +19199,10 @@ llama_token llama_token_eot(const struct llama_model * model) {
     return model->vocab.special_eot_id;
 }
 
+llama_token llama_token_pad(const struct llama_model * model) {
+    return model->vocab.special_pad_id;
+}
+
 int32_t llama_tokenize(
     const struct llama_model * model,
                   const char * text,
@@ -18724,7 +19269,8 @@ int32_t llama_token_to_piece(const struct llama_model * model, llama_token token
     if (0 <= token && token < llama_n_vocab(model)) {
         switch (llama_vocab_get_type(model->vocab)) {
             case LLAMA_VOCAB_TYPE_WPM:
-            case LLAMA_VOCAB_TYPE_SPM: {
+            case LLAMA_VOCAB_TYPE_SPM:
+            case LLAMA_VOCAB_TYPE_UGM: {
                 // NOTE: we accept all unsupported token types,
                 // suppressing them like CONTROL tokens.
                 if (llama_is_normal_token(model->vocab, token)) {
@@ -18818,10 +19364,10 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|im_start|>assistant\n";
         }
-    } else if (tmpl == "llama2" || tmpl.find("[INST]") != std::string::npos) {
+    } else if (tmpl == "llama2" || tmpl == "mistral" || tmpl.find("[INST]") != std::string::npos) {
         // llama2 template and its variants
         // [variant] support system message
-        bool support_system_message = tmpl.find("<<SYS>>") != std::string::npos;
+        bool support_system_message = tmpl.find("<<SYS>>") != std::string::npos || tmpl == "mistral";
         // [variant] space before + after response
         bool space_around_response = tmpl.find("' ' + eos_token") != std::string::npos;
         // [variant] add BOS inside history

llama.h

@@ -67,6 +67,7 @@ extern "C" {
         LLAMA_VOCAB_TYPE_SPM  = 1, // LLaMA tokenizer based on byte-level BPE with byte fallback
         LLAMA_VOCAB_TYPE_BPE  = 2, // GPT-2 tokenizer based on byte-level BPE
         LLAMA_VOCAB_TYPE_WPM  = 3, // BERT tokenizer based on WordPiece
+        LLAMA_VOCAB_TYPE_UGM  = 4, // T5 tokenizer based on Unigram
     };
 
     // pre-tokenization types
@@ -857,6 +858,7 @@ extern "C" {
     LLAMA_API llama_token llama_token_cls(const struct llama_model * model); // classification
     LLAMA_API llama_token llama_token_sep(const struct llama_model * model); // sentence separator
     LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
+    LLAMA_API llama_token llama_token_pad(const struct llama_model * model); // padding
 
     // Returns -1 if unknown, 1 for true or 0 for false.
     LLAMA_API int32_t         llama_add_bos_token(const struct llama_model * model);
@@ -924,6 +926,12 @@ extern "C" {
     // Grammar
     //
 
+    /// Initialize a llama_grammar.
+    ///
+    /// @param rules The rule elements of the grammar to initialize.
+    /// @param n_rules The number of rules.
+    /// @param start_rule_index The index of the root rule (the starting point of the grammar).
+    /// @return The initialized llama_grammar or nullptr if initialization failed.
     LLAMA_API struct llama_grammar * llama_grammar_init(
             const llama_grammar_element ** rules,
                                  size_t    n_rules,

tests/test-chat-template.cpp

@@ -7,6 +7,7 @@
 #include <cassert>
 
 #include "llama.h"
+#include "common.h"
 
 int main(void) {
     llama_chat_message conversation[] = {
@@ -119,5 +120,24 @@ int main(void) {
         std::cout << output << "\n-------------------------\n";
         assert(output == expected);
     }
+
+    // test llama_chat_format_single
+    std::cout << "\n\n=== llama_chat_format_single ===\n\n";
+    std::vector<llama_chat_msg> chat2;
+    chat2.push_back({"system", "You are a helpful assistant"});
+    chat2.push_back({"user", "Hello"});
+    chat2.push_back({"assistant", "I am assistant"});
+    llama_chat_msg new_msg{"user", "How are you"};
+
+    auto fmt_single = [&](std::string tmpl) {
+        auto output = llama_chat_format_single(nullptr, tmpl, chat2, new_msg, true);
+        std::cout << "fmt_single(" << tmpl << ")\n" << output << "\n-------------------------\n";
+        return output;
+    };
+    assert(fmt_single("chatml") == "<|im_start|>user\nHow are you<|im_end|>\n<|im_start|>assistant\n");
+    assert(fmt_single("llama2") == "[INST] How are you [/INST]");
+    assert(fmt_single("gemma") == "<start_of_turn>user\nHow are you<end_of_turn>\n<start_of_turn>model\n");
+    assert(fmt_single("llama3") == "<|start_header_id|>user<|end_header_id|>\n\nHow are you<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n");
+
     return 0;
 }

tests/test-grammar-integration.cpp

@@ -36,10 +36,10 @@ static llama_grammar* build_grammar(const std::string & grammar_str) {
 static bool test_build_grammar_fails(const std::string & grammar_str) {
     fprintf(stderr, "⚫ Testing failure for grammar: %s\n", grammar_str.c_str());
     bool grammar_fails = false;
-    try {
+    llama_grammar * grammar = build_grammar(grammar_str);
-        build_grammar(grammar_str);
+    if (grammar != nullptr) {
         fprintf(stderr, "  ❌ Expected build failure, but succeeded\n");
-    } catch (const std::exception & err) {
+    } else {
         grammar_fails = true;
         fprintf(stdout, "  ✅︎\n");
     }
@@ -148,6 +148,250 @@ static void test_schema(const std::string & test_desc, const std::string & schem
 }
 
 static void test_simple_grammar() {
+    test_schema(
+        "min 0",
+        R"""({
+            "type": "integer",
+            "minimum": 0
+        })""",
+        // Passing strings
+        {
+            "0",
+            "10",
+            "12",
+            "10000",
+        },
+        // Failing strings
+        {
+            "-1",
+            "-10",
+            "-10000",
+            "-100000000000000000000000000000000",
+            "100000000000000000000000000000000",
+            "00",
+            "01",
+            "-0",
+        }
+    );
+    test_schema(
+        "min 2",
+        // Schema
+        R"""({
+            "type": "integer",
+            "minimum": 2
+        })""",
+        // Passing strings
+        {
+            "2",
+            "3",
+            "4",
+            "10",
+            "20",
+            "1234567890000000",
+        },
+        // Failing strings
+        {
+            "0",
+            "1",
+            "-1",
+            "-100",
+            "0",
+            "1",
+            "01",
+            "02",
+            "12345678900000000",
+        }
+    );
+    test_schema(
+        "min 456",
+        R"""({
+            "type": "integer",
+            "minimum": 456
+        })""",
+        // Passing strings
+        {
+            "456",
+            "4560",
+            "457",
+            "460",
+            "500",
+        },
+        // Failing strings
+        {
+            "455",
+            "356",
+            "50",
+            "050",
+            "-1",
+            "-456",
+        }
+    );
+    test_schema(
+        "min -123",
+        R"""({
+            "type": "integer",
+            "minimum": -123
+        })""",
+        // Passing strings
+        {
+            "-123",
+            "-122",
+            "-11",
+            "-1",
+            "0",
+            "1",
+            "123",
+            "1234",
+            "2345",
+        },
+        // Failing strings
+        {
+            "-1234",
+            "-124",
+        }
+    );
+
+    test_schema(
+        "max 9999",
+        // Schema
+        R"""({
+            "type": "integer",
+            "maximum": 9999
+        })""",
+        // Passing strings
+        {
+            "-99999",
+            "0",
+            "9999",
+        },
+        // Failing strings
+        {
+            "10000",
+            "99991",
+        }
+    );
+    test_schema(
+        "max -9999",
+        // Schema
+        R"""({
+            "type": "integer",
+            "maximum": -9999
+        })""",
+        // Passing strings
+        {
+            "-10000",
+            "-9999",
+        },
+        // Failing strings
+        {
+            "-9998",
+            "0",
+            "9999",
+        }
+    );
+    test_schema(
+        "min 5 max 30",
+        // Schema
+        R"""({
+            "type": "integer",
+            "minimum": 5,
+            "maximum": 30
+        })""",
+        // Passing strings
+        {
+            "5",
+            "10",
+            "30",
+        },
+        // Failing strings
+        {
+            "05",
+            "4",
+            "-1",
+            "31",
+            "123",
+            "0123",
+        }
+    );
+    test_schema(
+        "min -1 max 1",
+        R"""({
+            "type": "integer",
+            "minimum": -1,
+            "maximum": 1
+        })""",
+        // Passing strings
+        {
+            "-1",
+            "0",
+            "1",
+        },
+        // Failing strings
+        {
+            "-11",
+            "-10",
+            "-2",
+            "2",
+            "10",
+            "11",
+        }
+    );
+    test_schema(
+        "min -123 max 42",
+        R"""({
+            "type": "integer",
+            "minimum": -123,
+            "maximum": 42
+        })""",
+        // Passing strings
+        {
+            "-123",
+            "-122",
+            "-13",
+            "-11",
+            "-2",
+            "-1",
+            "0",
+            "1",
+            "5",
+            "10",
+            "39",
+            "40",
+            "42",
+        },
+        // Failing strings
+        {
+            "-0123",
+            "-124",
+            "-1123",
+            "-200",
+            "43",
+            "123",
+            "0123",
+        }
+    );
+    test_schema(
+        "exclusive min / max",
+        // Schema
+        R"""({
+            "type": "integer",
+            "exclusiveMinimum": 0,
+            "exclusiveMaximum": 10000
+        })""",
+        // Passing strings
+        {
+            "1",
+            "9999",
+        },
+        // Failing strings
+        {
+            "0",
+            "01",
+            "10000",
+            "99999",
+        }
+    );
+
     // Test case for a simple grammar
     test_grammar(
         "simple grammar",

tests/test-json-schema-to-grammar.cpp

@@ -80,6 +80,232 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
         runner(tc);
     };
 
+    test({
+        SUCCESS,
+        "min 0",
+        R"""({
+            "type": "integer",
+            "minimum": 0
+        })""",
+        R"""(
+            root ::= ([0] | [1-9] [0-9]{0,15}) space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "min 1",
+        R"""({
+            "type": "integer",
+            "minimum": 1
+        })""",
+        R"""(
+            root ::= ([1-9] [0-9]{0,15}) space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "min 3",
+        R"""({
+            "type": "integer",
+            "minimum": 3
+        })""",
+        R"""(
+            root ::= ([1-2] [0-9]{1,15} | [3-9] [0-9]{0,15}) space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "min 9",
+        R"""({
+            "type": "integer",
+            "minimum": 9
+        })""",
+        R"""(
+            root ::= ([1-8] [0-9]{1,15} | [9] [0-9]{0,15}) space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "min 10",
+        R"""({
+            "type": "integer",
+            "minimum": 10
+        })""",
+        R"""(
+            root ::= ([1] ([0-9]{1,15}) | [2-9] [0-9]{1,15}) space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "min 25",
+        R"""({
+            "type": "integer",
+            "minimum": 25
+        })""",
+        R"""(
+            root ::= ([1] [0-9]{2,15} | [2] ([0-4] [0-9]{1,14} | [5-9] [0-9]{0,14}) | [3-9] [0-9]{1,15}) space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "max 30",
+        R"""({
+            "type": "integer",
+            "maximum": 30
+        })""",
+        R"""(
+            root ::= ("-" [1-9] [0-9]{0,15} | [0-9] | ([1-2] [0-9] | [3] "0")) space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "min -5",
+        R"""({
+            "type": "integer",
+            "minimum": -5
+        })""",
+        R"""(
+            root ::= ("-" ([0-5]) | [0] | [1-9] [0-9]{0,15}) space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "min -123",
+        R"""({
+            "type": "integer",
+            "minimum": -123
+        })""",
+        R"""(
+            root ::= ("-" ([0-9] | ([1-8] [0-9] | [9] [0-9]) | "1" ([0-1] [0-9] | [2] [0-3])) | [0] | [1-9] [0-9]{0,15}) space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "max -5",
+        R"""({
+            "type": "integer",
+            "maximum": -5
+        })""",
+        R"""(
+            root ::= ("-" ([0-4] [0-9]{1,15} | [5-9] [0-9]{0,15})) space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "max 1",
+        R"""({
+            "type": "integer",
+            "maximum": 1
+        })""",
+        R"""(
+            root ::= ("-" [1-9] [0-9]{0,15} | [0-1]) space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "max 100",
+        R"""({
+            "type": "integer",
+            "maximum": 100
+        })""",
+        R"""(
+            root ::= ("-" [1-9] [0-9]{0,15} | [0-9] | ([1-8] [0-9] | [9] [0-9]) | "100") space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "min 0 max 23",
+        R"""({
+            "type": "integer",
+            "minimum": 0,
+            "maximum": 23
+        })""",
+        R"""(
+            root ::= ([0-9] | ([1] [0-9] | [2] [0-3])) space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "min 15 max 300",
+        R"""({
+            "type": "integer",
+            "minimum": 15,
+            "maximum": 300
+        })""",
+        R"""(
+            root ::= (([1] ([5-9]) | [2-9] [0-9]) | ([1-2] [0-9]{2} | [3] "00")) space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "min 5 max 30",
+        R"""({
+            "type": "integer",
+            "minimum": 5,
+            "maximum": 30
+        })""",
+        R"""(
+            root ::= ([5-9] | ([1-2] [0-9] | [3] "0")) space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "min -123 max 42",
+        R"""({
+            "type": "integer",
+            "minimum": -123,
+            "maximum": 42
+        })""",
+        R"""(
+            root ::= ("-" ([0-9] | ([1-8] [0-9] | [9] [0-9]) | "1" ([0-1] [0-9] | [2] [0-3])) | [0-9] | ([1-3] [0-9] | [4] [0-2])) space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "min -10 max 10",
+        R"""({
+            "type": "integer",
+            "minimum": -10,
+            "maximum": 10
+        })""",
+        R"""(
+            root ::= ("-" ([0-9] | "10") | [0-9] | "10") space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
     test({
         FAILURE,
         "unknown type",
@@ -422,6 +648,44 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
         )"""
     });
 
+    test({
+        SUCCESS,
+        "min + max items with min + max values across zero",
+        R"""({
+            "items": {
+                "type": "integer",
+                "minimum": -12,
+                "maximum": 207
+            },
+            "minItems": 3,
+            "maxItems": 5
+        })""",
+        R"""(
+            item ::= ("-" ([0-9] | "1" [0-2]) | [0-9] | ([1-8] [0-9] | [9] [0-9]) | ([1] [0-9]{2} | [2] "0" [0-7])) space
+            root ::= "[" space item ("," space item){2,4} "]" space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "min + max items with min + max values",
+        R"""({
+            "items": {
+                "type": "integer",
+                "minimum": 12,
+                "maximum": 207
+            },
+            "minItems": 3,
+            "maxItems": 5
+        })""",
+        R"""(
+            item ::= (([1] ([2-9]) | [2-9] [0-9]) | ([1] [0-9]{2} | [2] "0" [0-7])) space
+            root ::= "[" space item ("," space item){2,4} "]" space
+            space ::= | " " | "\n" [ \t]{0,20}
+        )"""
+    });
+
     test({
         SUCCESS,
         "simple regexp",

tests/test-llama-grammar.cpp

@@ -116,6 +116,10 @@ int main()
     std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
     grammar = llama_grammar_init(
         grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
+    if (grammar == nullptr)
+    {
+        throw std::runtime_error("Failed to initialize llama_grammar");
+    }
 
     std::vector<std::vector<llama_grammar_element>> expected_stacks = {
         {

unicode.cpp

@@ -23,7 +23,7 @@ static std::string unicode_cpts_to_utf8(const std::vector<uint32_t> & cps) {
     return result;
 }
 
-static uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset) {
+uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset) {
     assert(offset < utf8.size());
     if (!(utf8[offset + 0] & 0x80)) {
         auto result = utf8[offset + 0];

unicode.h

@@ -48,6 +48,7 @@ struct codepoint_flags {
 
 
 std::string unicode_cpt_to_utf8(uint32_t cp);
+uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset);
 std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8);
 
 std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & cpts);