init

examples/lookup/lookup-merge.cpp

+#include "ggml.h"
+#include "llama.h"
+#include "common.h"
+#include "ngram-cache.h"
+
+#include <cstdint>
+#include <cstdio>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+static void print_usage(char* argv0) {
+    fprintf(stderr, "Merges multiple lookup cache files into a single one.\n");
+    fprintf(stderr, "Usage: %s [--help] lookup_part_1.bin lookup_part_2.bin ... lookup_merged.bin\n", argv0);
+}
+
+int main(int argc, char ** argv){
+    if (argc < 3) {
+        print_usage(argv[0]);
+        exit(1);
+    }
+
+    std::vector<std::string> args;
+    args.resize(argc-1);
+    for (int i = 0; i < argc-1; ++i) {
+        args[i] = argv[i+1];
+        if (args[i] == "-h" || args[i] == "--help") {
+            print_usage(argv[0]);
+            exit(0);
+        }
+    }
+
+    fprintf(stderr, "lookup-merge: loading file %s\n", args[0].c_str());
+    llama_ngram_cache ngram_cache_merged = llama_ngram_cache_load(args[0]);
+
+    for (size_t i = 1; i < args.size()-1; ++i) {
+        fprintf(stderr, "lookup-merge: loading file %s\n", args[i].c_str());
+        llama_ngram_cache ngram_cache = llama_ngram_cache_load(args[i]);
+
+        llama_ngram_cache_merge(ngram_cache_merged, ngram_cache);
+    }
+
+    fprintf(stderr, "lookup-merge: saving file %s\n", args.back().c_str());
+    llama_ngram_cache_save(ngram_cache_merged, args.back());
+}

examples/lookup/lookup-stats.cpp

+#include "ggml.h"
+#include "common.h"
+#include "llama.h"
+#include "log.h"
+#include "ngram-cache.h"
+
+#include <cmath>
+#include <cstdint>
+#include <cstdio>
+#include <fstream>
+#include <string>
+#include <vector>
+#include <unordered_map>
+
+int main(int argc, char ** argv){
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
+        return 1;
+    }
+
+    const int n_draft = params.n_draft;
+
+    // init llama.cpp
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    // load the model
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
+
+    // tokenize the prompt
+    std::vector<llama_token> inp;
+    inp = ::llama_tokenize(ctx, params.prompt, true, true);
+
+    llama_ngram_cache ngram_cache_context;
+    llama_ngram_cache ngram_cache_dynamic;
+    llama_ngram_cache ngram_cache_static;
+    int64_t t_draft_flat_us = 0;
+    int64_t t_draft_us = 0;
+
+    {
+        const int64_t t_start_draft_us = ggml_time_us();
+
+        if (!params.lookup_cache_static.empty()) {
+            try {
+                ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static);
+            } catch (std::ifstream::failure const &) {
+                fprintf(stderr, "error: failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
+                exit(1);
+            }
+        }
+
+        if (!params.lookup_cache_dynamic.empty()) {
+            try {
+                ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic);
+            } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
+        }
+
+        t_draft_flat_us += ggml_time_us() - t_start_draft_us;
+    }
+
+    const int n_input = inp.size();
+    const int n_ctx = llama_n_ctx(ctx);
+
+    int n_drafted = 0;
+    int n_accept  = 0;
+
+    const int64_t t_start_ms = ggml_time_ms();
+
+    // Iterate over input tokens in chunks of size n_ctx.
+    // Each chunk is treated as if a sequential generation but with pre-determined tokens to ensure reproducibility.
+    for (int i_start = 0; i_start + n_ctx < n_input; i_start += n_ctx) {
+        const std::vector<llama_token> inp_slice(inp.begin() + i_start, inp.begin() + i_start + n_ctx);
+        std::vector<llama_token> pseudo_output;
+        pseudo_output.push_back(inp_slice[0]);
+
+        while ((int) pseudo_output.size() < n_ctx) {
+            // Simulate drafting and decoding from draft:
+            std::vector<llama_token> draft;
+            draft.push_back(pseudo_output.back());
+
+            {
+                const int64_t t_start_draft_us = ggml_time_us();
+                llama_ngram_cache_draft(pseudo_output, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
+                t_draft_us += ggml_time_us() - t_start_draft_us;
+            }
+
+            n_drafted += draft.size() - 1;
+
+            for (size_t j = 1; j < draft.size() && (int) pseudo_output.size() < n_ctx; ++j) {
+                const llama_token ground_truth = inp_slice[pseudo_output.size()];
+                const llama_token drafted = draft[j];
+
+                if (ground_truth != drafted) {
+                    break;
+                }
+
+                ++n_accept;
+                pseudo_output.push_back(ground_truth);
+
+                {
+                    const int64_t t_start_draft_us = ggml_time_us();
+                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
+                    t_draft_us += ggml_time_us() - t_start_draft_us;
+                }
+            }
+
+            // After each simulated batch decoding simulate the sampling of a single token:
+            if ((int) pseudo_output.size() < n_ctx) {
+                pseudo_output.push_back(inp_slice[pseudo_output.size()]);
+                {
+                    const int64_t t_start_draft_us = ggml_time_us();
+                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
+                    t_draft_us += ggml_time_us() - t_start_draft_us;
+                }
+            }
+
+            draft.erase(draft.begin());
+
+        }
+        if (i_start > 0 && i_start / 100000 != (i_start - n_ctx) / 100000) {
+            const int64_t t_now_ms = ggml_time_ms();
+            const int64_t eta_ms   = (n_input - i_start) * (t_now_ms - t_start_ms) / i_start;
+            const int64_t eta_min  = eta_ms / (60*1000);
+            const int64_t eta_s    = (eta_ms - 60*1000*eta_min) / 1000;
+
+            LOG_TEE("lookup-stats: %d/%d done, ETA: %02" PRId64 ":%02" PRId64 "\n", i_start, n_input, eta_min, eta_s);
+        }
+
+        // After each chunk, update the dynamic ngram cache with the context ngram cache:
+        llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
+        ngram_cache_context.clear();
+    }
+
+    LOG_TEE("\n");
+
+    LOG_TEE("\n");
+    LOG_TEE("n_draft      = %d\n", n_draft);
+    LOG_TEE("n_predict    = %d\n", n_input - n_input % n_ctx);
+    LOG_TEE("n_drafted    = %d\n", n_drafted);
+    LOG_TEE("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
+    LOG_TEE("t_draft      = %.2f ms, %.2f us per token, %.2f tokens per second\n",
+            t_draft_us*1e-3, 1.0f*t_draft_us/n_drafted, n_drafted/(1e-6*t_draft_us));
+    LOG_TEE("n_accept     = %d\n", n_accept);
+    LOG_TEE("accept       = %.3f%%\n", 100.0f * n_accept / n_drafted);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    fprintf(stderr, "\n\n");
+
+    return 0;
+}

examples/lookup/lookup.cpp

+#include "ggml.h"
+#include "llama.h"
+#include "common.h"
+#include "ngram-cache.h"
+
+#include <cmath>
+#include <cstdint>
+#include <cstdio>
+#include <fstream>
+#include <string>
+#include <vector>
+#include <unordered_map>
+
+int main(int argc, char ** argv){
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
+        return 1;
+    }
+
+    // max. number of additional tokens to draft if match is found
+    const int n_draft = params.n_draft;
+
+    const bool dump_kv_cache = params.dump_kv_cache;
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("lookup", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
+
+    // init llama.cpp
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    // load the model
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
+
+    // tokenize the prompt
+    std::vector<llama_token> inp;
+    inp = ::llama_tokenize(ctx, params.prompt, true, true);
+
+    llama_ngram_cache ngram_cache_context;
+    llama_ngram_cache ngram_cache_dynamic;
+    llama_ngram_cache ngram_cache_static;
+    int64_t t_draft_flat_us = 0;
+    int64_t t_draft_us = 0;
+
+    {
+        // Fill up context ngram cache with tokens from user input:
+        const int64_t t_start_draft_us = ggml_time_us();
+        llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, inp.size(), false);
+
+        if (!params.lookup_cache_static.empty()) {
+            try {
+                ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static);
+            } catch (std::ifstream::failure const &) {
+                fprintf(stderr, "error: failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
+                exit(1);
+            }
+        }
+
+        if (!params.lookup_cache_dynamic.empty()) {
+            try {
+                ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic);
+            } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
+        }
+
+        t_draft_flat_us += ggml_time_us() - t_start_draft_us;
+    }
+
+    const int max_context_size     = llama_n_ctx(ctx);
+    const int max_tokens_list_size = max_context_size - 4;
+
+    if ((int) inp.size() > max_tokens_list_size) {
+        fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
+        return 1;
+    }
+
+    fprintf(stderr, "\n\n");
+
+    for (auto id : inp) {
+        fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
+    }
+
+    fflush(stderr);
+
+    const int n_input = inp.size();
+
+    const auto t_enc_start = ggml_time_us();
+
+    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1, 0,           0));
+    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1, n_input - 1, 0));
+
+    const auto t_enc_end = ggml_time_us();
+
+    int n_predict = 0;
+    int n_drafted = 0;
+    int n_accept  = 0;
+
+    int n_past = inp.size();
+
+    bool has_eos = false;
+
+    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams);
+
+    std::vector<llama_token> draft;
+
+    llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, 1);
+
+    // debug
+    struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, 1);
+
+    const auto t_dec_start = ggml_time_us();
+
+    while (true) {
+        // debug
+        if (dump_kv_cache) {
+            llama_kv_cache_view_update(ctx, &kvc_view);
+            llama_kv_cache_dump_view_seqs(kvc_view, 40);
+        }
+
+        // print current draft sequence
+        LOG("drafted %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, draft).c_str());
+
+        int i_dft = 0;
+        while (true) {
+            // sample from the target model
+            llama_token id = llama_sampling_sample(ctx_sampling, ctx, NULL, i_dft);
+
+            llama_sampling_accept(ctx_sampling, ctx, id, true);
+
+            const std::string token_str = llama_token_to_piece(ctx, id);
+
+            if (!params.use_color) {
+                printf("%s", token_str.c_str());
+            }
+
+            if (llama_token_is_eog(model, id)) {
+                has_eos = true;
+            }
+
+            ++n_predict;
+
+            // check if the target token matches the draft
+            if (i_dft < (int) draft.size() && id == draft[i_dft]) {
+                LOG("the sampled target token matches the %dth drafted token (%d, '%s') - accepted\n", i_dft, id, token_str.c_str());
+                ++n_accept;
+                ++n_past;
+                ++i_dft;
+                inp.push_back(id);
+                {
+                    // Update context ngram cache with the newly accepted token:
+                    const int64_t t_start_draft_us = ggml_time_us();
+                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
+                    t_draft_us += ggml_time_us() - t_start_draft_us;
+                }
+
+                if (params.use_color) {
+                    // color accepted draft token
+                    printf("\033[34m%s\033[0m", token_str.c_str());
+                    fflush(stdout);
+                }
+                continue;
+            }
+
+            if (params.use_color) {
+                printf("%s", token_str.c_str());
+            }
+            fflush(stdout);
+
+
+            LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
+
+            draft.clear();
+            draft.push_back(id);
+            inp.push_back(id);
+            {
+                // Update context ngram cache with the newly accepted token:
+                const int64_t t_start_draft_us = ggml_time_us();
+                llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
+                t_draft_us += ggml_time_us() - t_start_draft_us;
+            }
+            break;
+        }
+
+        if ((params.n_predict > 0 && n_predict > params.n_predict) || has_eos) {
+            break;
+        }
+
+        // KV cache management
+        // clean the cache of draft tokens that weren't accepted
+        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+
+        llama_batch_clear(batch_tgt);
+        llama_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);
+
+        // Draft already contains a single token sampled from the model:
+        GGML_ASSERT(draft.size() == 1);
+        GGML_ASSERT(draft[0] == inp.back());
+        const int64_t t_start_draft_us = ggml_time_us();
+
+        llama_ngram_cache_draft(inp, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
+
+        for (size_t i = 1; i < draft.size(); ++i) {
+            llama_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true);
+        }
+
+        t_draft_us += ggml_time_us() - t_start_draft_us;
+        n_drafted += draft.size() - 1;
+
+        llama_decode(ctx, batch_tgt);
+        ++n_past;
+
+        draft.erase(draft.begin());
+    }
+
+    auto t_dec_end = ggml_time_us();
+
+    // Update dynamic ngram cache with context ngram cache and save it to disk:
+    llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
+    llama_ngram_cache_save(ngram_cache_dynamic, params.lookup_cache_dynamic);
+
+    LOG_TEE("\n\n");
+
+    LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input,   (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
+    LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict  / ((t_dec_end - t_dec_start) / 1e6f));
+
+    LOG_TEE("\n");
+    LOG_TEE("n_draft      = %d\n", n_draft);
+    LOG_TEE("n_predict    = %d\n", n_predict);
+    LOG_TEE("n_drafted    = %d\n", n_drafted);
+    LOG_TEE("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
+    LOG_TEE("t_draft      = %.2f ms, %.2f us per token, %.2f tokens per second\n",
+            t_draft_us*1e-3, 1.0f*t_draft_us/n_drafted, n_drafted/(1e-6*t_draft_us));
+    LOG_TEE("n_accept     = %d\n", n_accept);
+    LOG_TEE("accept       = %.3f%%\n", 100.0f * n_accept / n_drafted);
+
+    LOG_TEE("\ntarget:\n");
+    llama_print_timings(ctx);
+
+    llama_sampling_free(ctx_sampling);
+    llama_batch_free(batch_tgt);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    fprintf(stderr, "\n\n");
+
+    return 0;
+}

examples/main-cmake-pkg/.gitignore

+# Prerequisites
+*.d
+
+# Compiled Object files
+*.slo
+*.lo
+*.o
+*.obj
+
+# Precompiled Headers
+*.gch
+*.pch
+
+# Compiled Dynamic libraries
+*.so
+*.dylib
+*.dll
+
+# Fortran module files
+*.mod
+*.smod
+
+# Compiled Static libraries
+*.lai
+*.la
+*.a
+*.lib
+
+# Executables
+*.exe
+*.out
+*.app
+
+*.gguf
+
+*.log
+.DS_Store
+.build/
+.cache/
+.direnv/
+.envrc
+.swiftpm
+.venv
+.clang-tidy
+.vs/
+.vscode/
+
+build*/
+out/
+tmp/

examples/main-cmake-pkg/CMakeLists.txt

+cmake_minimum_required(VERSION 3.12)
+project("llama-cli-cmake-pkg" C CXX)
+set(TARGET llama-cli-cmake-pkg)
+
+find_package(Llama 0.0.1 REQUIRED)
+
+# Bake common functionality in with target. Because applications
+# using the relocatable Llama package should be outside of the
+# source tree, llama-cli-cmake-pkg pretends the dependencies are built-in.
+set(_common_path "${CMAKE_CURRENT_LIST_DIR}/../../common")
+add_library(common OBJECT)
+file(GLOB _common_files
+    "${_common_path}/*.h"
+    "${_common_path}/*.cpp"
+)
+target_sources(common PRIVATE ${_common_files})
+
+# If the common project was part of "llama-cli-cmake-pkg" the transient
+# defines would automatically be attached. Because the common func-
+# tionality is separate, but dependent upon the defines, it must be
+# explicitly extracted from the "llama" target.
+#
+get_target_property(_llama_transient_defines llama
+    INTERFACE_COMPILE_DEFINITIONS)
+
+target_compile_definitions(common PRIVATE "${_llama_transient_defines}")
+
+add_executable(${TARGET} ${CMAKE_CURRENT_LIST_DIR}/../main/main.cpp)
+target_include_directories(${TARGET} PRIVATE ${_common_path})
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/main-cmake-pkg/README.md

+# llama.cpp/example/main-cmake-pkg
+
+This program builds [llama-cli](../main) using a relocatable CMake package. It serves as an example of using the `find_package()` CMake command to conveniently include [llama.cpp](https://github.com/ggerganov/llama.cpp) in projects which live outside of the source tree.
+
+## Building
+
+Because this example is "outside of the source tree", it is important to first build/install llama.cpp using CMake. An example is provided here, but please see the [llama.cpp build instructions](../..) for more detailed build instructions.
+
+### Considerations
+
+When hardware acceleration libraries are used (e.g. CUDA, Metal, etc.), CMake must be able to locate the associated CMake package.
+
+### Build llama.cpp and install to C:\LlamaCPP directory
+
+```cmd
+git clone https://github.com/ggerganov/llama.cpp
+cd llama.cpp
+cmake -B build -DBUILD_SHARED_LIBS=OFF -G "Visual Studio 17 2022" -A x64
+cmake --build build --config Release
+cmake --install build --prefix C:/LlamaCPP
+```
+
+### Build llama-cli-cmake-pkg
+
+
+```cmd
+cd ..\examples\main-cmake-pkg
+cmake -B build -DBUILD_SHARED_LIBS=OFF -DCMAKE_PREFIX_PATH="C:/LlamaCPP/lib/cmake/Llama" -G "Visual Studio 17 2022" -A x64
+cmake --build build --config Release
+cmake --install build --prefix C:/MyLlamaApp
+```

examples/main/CMakeLists.txt

+set(TARGET llama-cli)
+add_executable(${TARGET} main.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/main/README.md

+# llama.cpp/examples/main
+
+This example program allows you to use various LLaMA language models easily and efficiently. It is specifically designed to work with the [llama.cpp](https://github.com/ggerganov/llama.cpp) project, which provides a plain C/C++ implementation with optional 4-bit quantization support for faster, lower memory inference, and is optimized for desktop CPUs. This program can be used to perform various inference tasks with LLaMA models, including generating text based on user-provided prompts and chat-like interactions with reverse prompts.
+
+## Table of Contents
+
+1. [Quick Start](#quick-start)
+2. [Common Options](#common-options)
+3. [Input Prompts](#input-prompts)
+4. [Interaction](#interaction)
+5. [Context Management](#context-management)
+6. [Generation Flags](#generation-flags)
+7. [Performance Tuning and Memory Options](#performance-tuning-and-memory-options)
+8. [Additional Options](#additional-options)
+
+## Quick Start
+
+To get started right away, run the following command, making sure to use the correct path for the model you have:
+
+First, we will need to download a model. In these examples, we will use the Gemma model from the ggml-org repo on Hugging Face.
+[https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true](https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true)
+
+Once downloaded, place your model in the models folder in llama.cpp.
+
+### Unix-based systems (Linux, macOS, etc.):
+
+##### Input prompt (One-and-done)
+
+```bash
+./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --prompt "Once upon a time"
+```
+##### Conversation mode (Allow for continuous interaction with the model)
+
+```bash
+./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf -cnv --chat-template gemma
+```
+
+##### Infinite text from a starting prompt (you can use `Ctrl-C` to stop it):
+```bash
+./llama-cli -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
+```
+
+### Windows:
+
+##### Input prompt (One-and-done)
+```powershell
+./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --prompt "Once upon a time"
+```
+##### Conversation mode (Allow for continuous interaction with the model)
+
+```powershell
+./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf -cnv --chat-template gemma
+```
+
+#### Infinite text from a starting prompt (you can use `Ctrl-C` to stop it):
+
+```powershell
+llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
+```
+
+## Common Options
+
+In this section, we cover the most commonly used options for running the `llama-cli` program with the LLaMA models:
+
+-   `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/gemma-1.1-7b-it.Q4_K_M.gguf`; inferred from `--model-url` if set).
+-   `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (e.g [https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true](https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true)).
+-   `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
+-   `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
+-   `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
+-   `-mli, --multiline-input`: Allows you to write or paste multiple lines without ending each in '\'
+-   `-t N, --threads N`: Set the number of threads to use during generation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has.
+-   -   `-ngl N, --n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
+
+## Input Prompts
+
+The `llama-cli` program provides several ways to interact with the LLaMA models using input prompts:
+
+-   `--prompt PROMPT`: Provide a prompt directly as a command-line option.
+-   `--file FNAME`: Provide a file containing a prompt or multiple prompts.
+-   `--interactive-first`: Run the program in interactive mode and wait for input right away. (More on this below.)
+
+## Interaction
+
+The `llama-cli` program offers a seamless way to interact with LLaMA models, allowing users to engage in real-time conversations or provide instructions for specific tasks. The interactive mode can be triggered using various options, including `--interactive` and `--interactive-first`.
+
+In interactive mode, users can participate in text generation by injecting their input during the process. Users can press `Ctrl+C` at any time to interject and type their input, followed by pressing `Return` to submit it to the LLaMA model. To submit additional lines without finalizing input, users can end the current line with a backslash (`\`) and continue typing.
+
+### Interaction Options
+
+-   `-i, --interactive`: Run the program in interactive mode, allowing users to engage in real-time conversations or provide specific instructions to the model.
+-   `--interactive-first`: Run the program in interactive mode and immediately wait for user input before starting the text generation.
+-   `-cnv,  --conversation`:  Run the program in conversation mode (does not print special tokens and suffix/prefix, use default chat template) (default: false)
+-   `--color`: Enable colorized output to differentiate visually distinguishing between prompts, user input, and generated text.
+
+By understanding and utilizing these interaction options, you can create engaging and dynamic experiences with the LLaMA models, tailoring the text generation process to your specific needs.
+
+### Reverse Prompts
+
+Reverse prompts are a powerful way to create a chat-like experience with a LLaMA model by pausing the text generation when specific text strings are encountered:
+
+-   `-r PROMPT, --reverse-prompt PROMPT`: Specify one or multiple reverse prompts to pause text generation and switch to interactive mode. For example, `-r "User:"` can be used to jump back into the conversation whenever it's the user's turn to speak. This helps create a more interactive and conversational experience. However, the reverse prompt doesn't work when it ends with a space.
+
+To overcome this limitation, you can use the `--in-prefix` flag to add a space or any other characters after the reverse prompt.
+
+### In-Prefix
+
+The `--in-prefix` flag is used to add a prefix to your input, primarily, this is used to insert a space after the reverse prompt. Here's an example of how to use the `--in-prefix` flag in conjunction with the `--reverse-prompt` flag:
+
+```sh
+./llama-cli -r "User:" --in-prefix " "
+```
+
+### In-Suffix
+
+The `--in-suffix` flag is used to add a suffix after your input. This is useful for adding an "Assistant:" prompt after the user's input. It's added after the new-line character (`\n`) that's automatically added to the end of the user's input. Here's an example of how to use the `--in-suffix` flag in conjunction with the `--reverse-prompt` flag:
+
+```sh
+./llama-cli -r "User:" --in-prefix " " --in-suffix "Assistant:"
+```
+When --in-prefix or --in-suffix options are enabled the chat template ( --chat-template ) is disabled
+
+### Chat templates
+
+ `--chat-template JINJA_TEMPLATE`: This option sets a custom jinja chat template. It accepts a string, not a file name.  Default: template taken from model's metadata. Llama.cpp only supports [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template). These include llama2, llama3, gemma, monarch, chatml, orion, vicuna, vicuna-orca, deepseek, command-r, zephyr. When --in-prefix or --in-suffix options are enabled the chat template ( --chat-template ) is disabled.
+
+ Example usage: `--chat-template gemma`
+
+## Context Management
+
+During text generation, LLaMA models have a limited context size, which means they can only consider a certain number of tokens from the input and generated text. When the context fills up, the model resets internally, potentially losing some information from the beginning of the conversation or instructions. Context management options help maintain continuity and coherence in these situations.
+
+### Context Size
+
+- `-c N, --ctx-size N`: Set the size of the prompt context (default: 0, 0 = loaded from model). The LLaMA models were built with a context of 2048-8192, which will yield the best results on longer input/inference.
+
+### Extended Context Size
+
+Some fine-tuned models have extended the context length by scaling RoPE. For example, if the original pre-trained model has a context length (max sequence length) of 4096 (4k) and the fine-tuned model has 32k. That is a scaling factor of 8, and should work by setting the above `--ctx-size` to 32768 (32k) and `--rope-scale` to 8.
+
+-   `--rope-scale N`: Where N is the linear scaling factor used by the fine-tuned model.
+
+### Keep Prompt
+
+The `--keep` option allows users to retain the original prompt when the model runs out of context, ensuring a connection to the initial instruction or conversation topic is maintained.
+
+-   `--keep N`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context. By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt.
+
+By utilizing context management options like `--ctx-size` and `--keep`, you can maintain a more coherent and consistent interaction with the LLaMA models, ensuring that the generated text remains relevant to the original prompt or conversation.
+
+## Generation Flags
+
+The following options allow you to control the text generation process and fine-tune the diversity, creativity, and quality of the generated text according to your needs. By adjusting these options and experimenting with different combinations of values, you can find the best settings for your specific use case.
+
+### Number of Tokens to Predict
+
+-   `-n N, --predict N`: Set the number of tokens to predict when generating text (default: -1, -1 = infinity, -2 = until context filled)
+
+The `--predict` option controls the number of tokens the model generates in response to the input prompt. By adjusting this value, you can influence the length of the generated text. A higher value will result in longer text, while a lower value will produce shorter text.
+
+A value of -1 will enable infinite text generation, even though we have a finite context window. When the context window is full, some of the earlier tokens (half of the tokens after `--keep`) will be discarded. The context must then be re-evaluated before generation can resume. On large models and/or large context windows, this will result in a significant pause in output.
+
+If the pause is undesirable, a value of -2 will stop generation immediately when the context is filled.
+
+It is important to note that the generated text may be shorter than the specified number of tokens if an End-of-Sequence (EOS) token or a reverse prompt is encountered. In interactive mode, text generation will pause and control will be returned to the user. In non-interactive mode, the program will end. In both cases, the text generation may stop before reaching the specified `--predict` value. If you want the model to keep going without ever producing End-of-Sequence on its own, you can use the `--ignore-eos` parameter.
+
+### Temperature
+
+-   `--temp N`: Adjust the randomness of the generated text (default: 0.8).
+
+Temperature is a hyperparameter that controls the randomness of the generated text. It affects the probability distribution of the model's output tokens. A higher temperature (e.g., 1.5) makes the output more random and creative, while a lower temperature (e.g., 0.5) makes the output more focused, deterministic, and conservative. The default value is 0.8, which provides a balance between randomness and determinism. At the extreme, a temperature of 0 will always pick the most likely next token, leading to identical outputs in each run.
+
+Example usage: `--temp 0`
+
+### Repeat Penalty
+
+-   `--repeat-penalty N`: Control the repetition of token sequences in the generated text default: 1.0, 1.0 = disabled).
+-   `--repeat-last-n N`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx-size).
+-   `--no-penalize-nl`: Disable penalization for newline tokens when applying the repeat penalty.
+
+The `repeat-penalty` option helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. The default value is 1.
+
+The `repeat-last-n` option controls the number of tokens in the history to consider for penalizing repetition. A larger value will look further back in the generated text to prevent repetitions, while a smaller value will only consider recent tokens. A value of 0 disables the penalty, and a value of -1 sets the number of tokens considered equal to the context size (`ctx-size`).
+
+Use the `--no-penalize-nl` option to disable newline penalization when applying the repeat penalty. This option is particularly useful for generating chat conversations, dialogues, code, poetry, or any text where newline tokens play a significant role in structure and formatting. Disabling newline penalization helps maintain the natural flow and intended formatting in these specific use cases.
+
+Example usage: `--repeat-penalty 1.15 --repeat-last-n 128 --no-penalize-nl`
+
+### Top-K Sampling
+
+-   `--top-k N`: Limit the next token selection to the K most probable tokens (default: 40).
+
+Top-k sampling is a text generation method that selects the next token only from the top k most likely tokens predicted by the model. It helps reduce the risk of generating low-probability or nonsensical tokens, but it may also limit the diversity of the output. A higher value for top-k (e.g., 100) will consider more tokens and lead to more diverse text, while a lower value (e.g., 10) will focus on the most probable tokens and generate more conservative text. The default value is 40.
+
+Example usage: `--top-k 30`
+
+### Top-P Sampling
+
+-   `--top-p N`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.9).
+
+Top-p sampling, also known as nucleus sampling, is another text generation method that selects the next token from a subset of tokens that together have a cumulative probability of at least p. This method provides a balance between diversity and quality by considering both the probabilities of tokens and the number of tokens to sample from. A higher value for top-p (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. The default value is 0.9.
+
+Example usage: `--top-p 0.95`
+
+### Min-P Sampling
+
+-   `--min-p N`: Sets a minimum base probability threshold for token selection (default: 0.1).
+
+The Min-P sampling method was designed as an alternative to Top-P, and aims to ensure a balance of quality and variety. The parameter *p* represents the minimum probability for a token to be considered, relative to the probability of the most likely token. For example, with *p*=0.05 and the most likely token having a probability of 0.9, logits with a value less than 0.045 are filtered out.
+
+Example usage: `--min-p 0.05`
+
+### Tail-Free Sampling (TFS)
+
+-   `--tfs N`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled).
+
+Tail-free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. Similar to Top-P it tries to determine the bulk of the most likely tokens dynamically. But TFS filters out logits based on the second derivative of their probabilities. Adding tokens is stopped after the sum of the second derivatives reaches the parameter z. In short: TFS looks at how quickly the probabilities of the tokens decrease and cuts off the tail of unlikely tokens using the parameter z. Typical values for z are in the range of 0.9 to 0.95. A value of 1.0 would include all tokens and thus disables the effect of TFS.
+
+Example usage: `--tfs 0.95`
+
+### Locally Typical Sampling
+
+-   `--typical N`: Enable locally typical sampling with parameter p (default: 1.0, 1.0 = disabled).
+
+Locally typical sampling promotes the generation of contextually coherent and diverse text by sampling tokens that are typical or expected based on the surrounding context. By setting the parameter p between 0 and 1, you can control the balance between producing text that is locally coherent and diverse. A value closer to 1 will promote more contextually coherent tokens, while a value closer to 0 will promote more diverse tokens. A value equal to 1 disables locally typical sampling.
+
+Example usage: `--typical 0.9`
+
+### Mirostat Sampling
+
+-   `--mirostat N`: Enable Mirostat sampling, controlling perplexity during text generation (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0).
+-   `--mirostat-lr N`: Set the Mirostat learning rate, parameter eta (default: 0.1).
+-   `--mirostat-ent N`: Set the Mirostat target entropy, parameter tau (default: 5.0).
+
+Mirostat is an algorithm that actively maintains the quality of generated text within a desired range during text generation. It aims to strike a balance between coherence and diversity, avoiding low-quality output caused by excessive repetition (boredom traps) or incoherence (confusion traps).
+
+The `--mirostat-lr` option sets the Mirostat learning rate (eta). The learning rate influences how quickly the algorithm responds to feedback from the generated text. A lower learning rate will result in slower adjustments, while a higher learning rate will make the algorithm more responsive. The default value is `0.1`.
+
+The `--mirostat-ent` option sets the Mirostat target entropy (tau), which represents the desired perplexity value for the generated text. Adjusting the target entropy allows you to control the balance between coherence and diversity in the generated text. A lower value will result in more focused and coherent text, while a higher value will lead to more diverse and potentially less coherent text. The default value is `5.0`.
+
+Example usage: `--mirostat 2 --mirostat-lr 0.05 --mirostat-ent 3.0`
+
+### Logit Bias
+
+-   `-l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS`: Modify the likelihood of a token appearing in the generated text completion.
+
+The logit bias option allows you to manually adjust the likelihood of specific tokens appearing in the generated text. By providing a token ID and a positive or negative bias value, you can increase or decrease the probability of that token being generated.
+
+For example, use `--logit-bias 15043+1` to increase the likelihood of the token 'Hello', or `--logit-bias 15043-1` to decrease its likelihood. Using a value of negative infinity, `--logit-bias 15043-inf` ensures that the token `Hello` is never produced.
+
+A more practical use case might be to prevent the generation of `\code{begin}` and `\code{end}` by setting the `\` token (29905) to negative infinity with `-l 29905-inf`. (This is due to the prevalence of LaTeX codes that show up in LLaMA model inference.)
+
+Example usage: `--logit-bias 29905-inf`
+
+### RNG Seed
+
+-   `-s SEED, --seed SEED`: Set the random number generator (RNG) seed (default: -1, -1 = random seed).
+
+The RNG seed is used to initialize the random number generator that influences the text generation process. By setting a specific seed value, you can obtain consistent and reproducible results across multiple runs with the same input and settings. This can be helpful for testing, debugging, or comparing the effects of different options on the generated text to see when they diverge. If the seed is set to a value less than 0, a random seed will be used, which will result in different outputs on each run.
+
+## Performance Tuning and Memory Options
+
+These options help improve the performance and memory usage of the LLaMA models. By adjusting these settings, you can fine-tune the model's behavior to better suit your system's capabilities and achieve optimal performance for your specific use case.
+
+### Number of Threads
+
+-   `-t N, --threads N`: Set the number of threads to use during generation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has (as opposed to the logical number of cores). Using the correct number of threads can greatly improve performance.
+-   `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. In some systems, it is beneficial to use a higher number of threads during batch processing than during generation. If not specified, the number of threads used for batch processing will be the same as the number of threads used for generation.
+
+### Mlock
+
+-   `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped. This can improve performance but trades away some of the advantages of memory-mapping by requiring more RAM to run and potentially slowing down load times as the model loads into RAM.
+
+### No Memory Mapping
+
+-   `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed. However, if the model is larger than your total amount of RAM or if your system is low on available memory, using mmap might increase the risk of pageouts, negatively impacting performance. Disabling mmap results in slower load times but may reduce pageouts if you're not using `--mlock`. Note that if the model is larger than the total amount of RAM, turning off mmap would prevent the model from loading at all.
+
+### NUMA support
+
+-   `--numa distribute`: Pin an equal proportion of the threads to the cores on each NUMA node. This will spread the load amongst all cores on the system, utilitizing all memory channels at the expense of potentially requiring memory to travel over the slow links between nodes.
+-   `--numa isolate`: Pin all threads to the NUMA node that the program starts on. This limits the number of cores and amount of memory that can be used, but guarantees all memory access remains local to the NUMA node.
+-   `--numa numactl`: Pin threads to the CPUMAP that is passed to the program by starting it with the numactl utility. This is the most flexible mode, and allow arbitrary core usage patterns, for example a map that uses all the cores on one NUMA nodes, and just enough cores on a second node to saturate the inter-node memory bus.
+
+ These flags attempt optimizations that help on some systems with non-uniform memory access. This currently consists of one of the above strategies, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
+
+### Memory Float 32
+
+-   `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. This doubles the context memory requirement and cached prompt file size but does not appear to increase generation quality in a measurable way. Not recommended.
+
+### Batch Size
+
+-   `-b N, --batch-size N`: Set the batch size for prompt processing (default: `2048`). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.
+
+- `-ub N`, `--ubatch-size N`: physical maximum batch size. This is for pipeline parallelization. Default: `512`.
+
+### Prompt Caching
+
+-   `--prompt-cache FNAME`: Specify a file to cache the model state after the initial prompt. This can significantly speed up the startup time when you're using longer prompts. The file is created during the first run and is reused and updated in subsequent runs. **Note**: Restoring a cached prompt does not imply restoring the exact state of the session at the point it was saved. So even when specifying a specific seed, you are not guaranteed to get the same sequence of tokens as the original generation.
+
+### Grammars & JSON schemas
+
+-   `--grammar GRAMMAR`, `--grammar-file FILE`: Specify a grammar (defined inline or in a file) to constrain model output to a specific format. For example, you could force the model to output JSON or to speak only in emojis. See the [GBNF guide](../../grammars/README.md) for details on the syntax.
+
+-   `--json-schema SCHEMA`: Specify a [JSON schema](https://json-schema.org/) to constrain model output to (e.g. `{}` for any JSON object, or `{"items": {"type": "string", "minLength": 10, "maxLength": 100}, "minItems": 10}` for a JSON array of strings with size constraints). If a schema uses external `$ref`s, you should use `--grammar "$( python examples/json_schema_to_grammar.py myschema.json )"` instead.
+
+### Quantization
+
+For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-and-quantize).
+
+## Additional Options
+
+These options provide extra functionality and customization when running the LLaMA models:
+
+-   `-h, --help`: Display a help message showing all available options and their default values. This is particularly useful for checking the latest options and default values, as they can change frequently, and the information in this document may become outdated.
+-   `--verbose-prompt`: Print the prompt before generating text.
+-   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used.
+-   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance.
+-   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
+-   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
+-   `-hfr URL --hf-repo URL`: The url to the Hugging Face model repository. Used in conjunction with `--hf-file` or `-hff`. The model is downloaded and stored in the file provided by `-m` or `--model`. If `-m` is not provided, the model is auto-stored in the path specified by the `LLAMA_CACHE` environment variable  or in an OS-specific local cache.