init

examples/gritlm/gritlm.cpp

+#include "common.h"
+#include "llama.h"
+
+#include <string>
+#include <vector>
+
+// #define GRIT_DEBUG
+
+static std::vector<std::vector<float>> encode(llama_context * ctx, const std::vector<std::string> & sentences, const std::string & instruction) {
+    std::vector<std::vector<float>> result;
+
+    const llama_model * mdl = llama_get_model(ctx);
+
+    llama_batch batch = llama_batch_init(llama_n_batch(ctx), 0, 1);
+
+    for (uint64_t i = 0; i < sentences.size(); i++) {
+        llama_batch_clear(batch);
+
+        const std::string input_string = instruction + sentences[i];
+
+        std::vector<llama_token> inputs = llama_tokenize(mdl, input_string, true, false);
+
+        const int32_t n_toks = inputs.size();
+
+        // GritLM seems to have EOS = ""
+        // https://github.com/ContextualAI/gritlm/blob/92025b16534712b31b3c4aaaf069350e222bd5f8/gritlm/gritlm.py#L18
+        // inputs.push_back(llama_token_eos(mdl));
+
+        // we want to ignore instruction tokens for mean pooling
+        const int32_t n_inst = llama_tokenize(mdl, instruction, true, false).size();
+
+#ifdef GRIT_DEBUG
+        // debug tokens - should be matching as referenced in the GritLM sample
+        std::for_each(inputs.begin(), inputs.end(), [&ctx](llama_token t) {
+            std::printf("[%u:%s]", t, llama_token_to_piece(ctx, t).c_str());
+        });
+        std::printf("\n");
+#endif
+
+        // add input to batch (this increments n_tokens)
+        for (int32_t j = 0; j < n_toks; j++) {
+            llama_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
+        }
+
+        // clear previous kv_cache values (irrelevant for embeddings)
+        llama_kv_cache_clear(ctx);
+        llama_set_embeddings(ctx, true);
+        llama_set_causal_attn(ctx, false);
+
+        // run model
+        llama_decode(ctx, batch);
+
+        // get embedding dimensions
+        uint64_t n_embd = llama_n_embd(mdl);
+
+        // allocate embedding output
+        std::vector<float> emb_unorm(n_embd, 0.0f);
+
+        // sum up all token embeddings
+        for (int32_t k = n_inst; k < n_toks; k++) {
+            float * emb = llama_get_embeddings_ith(ctx, k);
+            for (uint64_t j = 0; j < n_embd; j++) {
+                emb_unorm[j] += emb[j];
+            }
+        }
+
+        // divide by number of tokens (mean pooling)
+        {
+            const uint64_t n_sent = n_toks - n_inst;
+
+            for (uint64_t j = 0; j < n_embd; j++) {
+                emb_unorm[j] /= n_sent;
+            }
+        }
+
+        std::vector<float> emb_norm(emb_unorm.size());
+        llama_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd);
+        result.push_back(emb_norm);
+
+#ifdef GRIT_DEBUG
+        // print out emb_norm
+        std::printf("embedding %ld: ", i);
+        for (uint64_t j = 0; j < n_embd; j++) {
+            std::printf("%.5f ", emb_norm[j]);
+        }
+        std::printf("\n\n");
+#endif
+    }
+
+    llama_batch_free(batch);
+
+    return result;
+}
+
+static std::string generate(llama_context * ctx, const std::string & prompt, bool stream) {
+    std::string result;
+
+    const llama_model * mdl = llama_get_model(ctx);
+    llama_token eos_token = llama_token_eos(mdl);
+
+    llama_kv_cache_clear(ctx);
+    llama_set_embeddings(ctx, false);
+    llama_set_causal_attn(ctx, true);
+
+    llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
+
+    std::vector<llama_token> inputs = llama_tokenize(mdl, prompt, false, true);
+    int32_t i_current_token = 0;
+
+    while (true) {
+        llama_batch_clear(bat);
+        auto n_inputs = (int32_t)inputs.size();
+        for (int32_t i = 0; i < n_inputs; i++) {
+            llama_batch_add(bat, inputs[i], i_current_token++, { 0 }, i == n_inputs - 1);
+        }
+        inputs.clear();
+
+        llama_decode(ctx, bat);
+        auto logits = llama_get_logits_ith(ctx, bat.n_tokens - 1);
+
+        auto candidates = std::vector<llama_token_data>(llama_n_vocab(mdl));
+        auto n_candidates = (int32_t)candidates.size();
+        for (int32_t token = 0; token < n_candidates; token++) {
+            candidates[token] = llama_token_data{ token, logits[token], 0.0f };
+        }
+        auto candidates_p = llama_token_data_array{ candidates.data(), candidates.size(), false };
+
+        llama_token token = llama_sample_token_greedy(ctx, &candidates_p);
+        if (token == eos_token) {
+            break;
+        }
+
+        std::string piece = llama_token_to_piece(ctx, token);
+        if (stream) {
+            std::printf("%s", piece.c_str());
+            std::fflush(stdout);
+        }
+
+        inputs.push_back(token);
+
+        result += piece;
+    }
+
+    if (stream) {
+        std::printf("\n");
+    }
+
+    llama_batch_free(bat);
+
+    return result;
+}
+
+static std::string gritlm_instruction(const std::string & instruction) {
+    return !instruction.empty() ? "<|user|>\n" + instruction + "\n<|embed|>\n" : "<|embed|>\n";
+}
+
+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;
+    }
+
+    llama_model_params mparams = llama_model_params_from_gpt_params(params);
+    llama_context_params cparams = llama_context_params_from_gpt_params(params);
+
+    llama_backend_init();
+
+    llama_model * mdl = llama_load_model_from_file(params.model.c_str(), mparams);
+
+    // create generation context
+    llama_context * ctx = llama_new_context_with_model(mdl, cparams);
+
+    // ### Embedding/Representation ###
+    // samples taken from: https://github.com/ContextualAI/gritlm#basic
+    {
+        const std::string instruction = "Given a scientific paper title, retrieve the paper's abstract";
+
+        const std::vector<std::string> queries = {
+            "Bitcoin: A Peer-to-Peer Electronic Cash System",
+            "Generative Representational Instruction Tuning",
+        };
+
+        const std::vector<std::string> documents = {
+            "A purely peer-to-peer version of electronic cash would allow online payments to be sent directly from one party to another without going through a financial institution. Digital signatures provide part of the solution, but the main benefits are lost if a trusted third party is still required to prevent double-spending. We propose a solution to the double-spending problem using a peer-to-peer network. The network timestamps transactions by hashing them into an ongoing chain of hash-based proof-of-work, forming a record that cannot be changed without redoing the proof-of-work. The longest chain not only serves as proof of the sequence of events witnessed, but proof that it came from the largest pool of CPU power. As long as a majority of CPU power is controlled by nodes that are not cooperating to attack the network, they'll generate the longest chain and outpace attackers. The network itself requires minimal structure. Messages are broadcast on a best effort basis, and nodes can leave and rejoin the network at will, accepting the longest proof-of-work chain as proof of what happened while they were gone.",
+            "All text-based language problems can be reduced to either generation or embedding. Current models only perform well at one or the other. We introduce generative representational instruction tuning (GRIT) whereby a large language model is trained to handle both generative and embedding tasks by distinguishing between them through instructions. Compared to other open models, our resulting GritLM 7B sets a new state of the art on the Massive Text Embedding Benchmark (MTEB) and outperforms all models up to its size on a range of generative tasks. By scaling up further, GritLM 8X7B outperforms all open generative language models that we tried while still being among the best embedding models. Notably, we find that GRIT matches training on only generative or embedding data, thus we can unify both at no performance loss. Among other benefits, the unification via GRIT speeds up Retrieval-Augmented Generation (RAG) by > 60% for long documents, by no longer requiring separate retrieval and generation models. Models, code, etc. are freely available at https://github.com/ContextualAI/gritlm.",
+        };
+
+        // No need to add instruction for retrieval documents
+        const std::vector<std::vector<float>> d_rep = encode(ctx, documents, gritlm_instruction(""));
+        const std::vector<std::vector<float>> q_rep = encode(ctx, queries,   gritlm_instruction(instruction));
+
+        const int n_embd = llama_n_embd(mdl);
+
+        const float cosine_sim_q0_d0 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
+        const float cosine_sim_q0_d1 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
+        const float cosine_sim_q1_d0 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[0].data(), n_embd);
+        const float cosine_sim_q1_d1 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[1].data(), n_embd);
+
+        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[0].c_str(), cosine_sim_q0_d0);
+        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[1].c_str(), cosine_sim_q0_d1);
+        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[0].c_str(), cosine_sim_q1_d0);
+        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[1].c_str(), cosine_sim_q1_d1);
+    }
+
+    // ### Generation ###
+    // GritLM models are not finetuned with system prompts, as you can just include system-like instructions together with your user instruction
+    {
+        const std::string prompt = "<|user|>\nPlease write me a poem about my recent hike of Mt. Fuji at midnight in the style of Shakespeare.\n<|assistant|>\n";
+        std::string response = generate(ctx, prompt, true);
+    }
+
+    llama_free(ctx);
+    llama_free_model(mdl);
+    llama_backend_free();
+
+    return 0;
+}

examples/imatrix/CMakeLists.txt

+set(TARGET llama-imatrix)
+add_executable(${TARGET} imatrix.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/imatrix/README.md

+# llama.cpp/examples/imatrix
+
+Compute an importance matrix for a model and given text dataset. Can be used during quantization to enchance the quality of the quantized models.
+More information is available here: https://github.com/ggerganov/llama.cpp/pull/4861
+
+## Usage
+
+```
+./llama-imatrix \
+    -m model.gguf -f some-text.txt [-o imatrix.dat] [--process-output] [--verbosity 1] \
+    [--no-ppl] [--chunk 123] [--output-frequency 10] [--save-frequency 0] \
+    [--in-file imatrix-prev-0.dat --in-file imatrix-prev-1.dat ...]
+```
+
+Here `-m` with a model name and `-f` with a file containing training data (such as e.g. `wiki.train.raw`) are mandatory.
+The parameters in square brackets are optional and have the following meaning:
+* `-o` (or `--output-file`) specifies the name of the file where the computed data will be stored. If missing `imatrix.dat` is used.
+* `--verbosity` specifies the verbosity level. If set to `0`, no output other than the perplexity of the processed chunks will be generated. If set to `1`, each time the results are saved a message is written to `stderr`. If `>=2`, a message is output each time data is collected for any tensor. Default verbosity level is `1`.
+* `--output-frequency` specifies how often the so far computed result is saved to disk. Default is 10 (i.e., every 10 chunks)
+* `--save-frequency` specifies how often to save a copy of the imatrix in a separate file. Default is 0 (i.e., never)
+* `--process-output` specifies if data will be collected for the `output.weight` tensor. My experience is that it is better to not utilize the importance matrix when quantizing `output.weight`, so this is set to `false` by default.
+
+For faster computation, make sure to use GPU offloading via the `-ngl` argument
+
+## Example
+
+```bash
+GGML_CUDA=1 make -j
+
+# generate importance matrix (imatrix.dat)
+./llama-imatrix -m ggml-model-f16.gguf -f train-data.txt -ngl 99
+
+# use the imatrix to perform a Q4_K_M quantization
+./llama-quantize --imatrix imatrix.dat ggml-model-f16.gguf ./ggml-model-q4_k_m.gguf q4_k_m
+```

examples/imatrix/imatrix.cpp

+#include "common.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <sstream>
+#include <thread>
+#include <mutex>
+#include <vector>
+#include <fstream>
+#include <unordered_map>
+#include <algorithm>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+static void print_usage(int argc, char ** argv, const gpt_params & params) {
+    gpt_params_print_usage(argc, argv, params);
+
+    LOG_TEE("\nexample usage:\n");
+    LOG_TEE("\n    %s \\\n"
+            "       -m model.gguf -f some-text.txt [-o imatrix.dat] [--process-output] [--verbosity 1] \\\n"
+            "       [--no-ppl] [--chunk 123] [--output-frequency 10] [--save-frequency 0] \\\n"
+            "       [--in-file imatrix-prev-0.dat --in-file imatrix-prev-1.dat ...]\n" , argv[0]);
+    LOG_TEE("\n");
+}
+
+struct Stats {
+    std::vector<float> values;
+    std::vector<int> counts;
+    int ncall = 0;
+};
+
+class IMatrixCollector {
+public:
+    IMatrixCollector() = default;
+    void set_params(gpt_params params) { m_params = std::move(params); }
+    bool collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data);
+    void save_imatrix(int ncall = -1) const;
+    bool load_imatrix(const char * file_name);
+private:
+    std::unordered_map<std::string, Stats> m_stats;
+    gpt_params                             m_params;
+    std::mutex                             m_mutex;
+    int                                    m_last_call = 0;
+    std::vector<float>                     m_src1_data;
+    std::vector<char>                      m_ids; // the expert ids from ggml_mul_mat_id
+};
+
+// remove any prefix and suffixes from the name
+// CUDA0#blk.0.attn_k.weight#0 => blk.0.attn_k.weight
+static std::string filter_tensor_name(const char * name) {
+    std::string wname;
+    const char * p = strchr(name, '#');
+    if (p != NULL) {
+        p = p + 1;
+        const char * q = strchr(p, '#');
+        if (q != NULL) {
+            wname = std::string(p, q - p);
+        } else {
+            wname = p;
+        }
+    } else {
+        wname = name;
+    }
+    return wname;
+}
+
+bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {
+    GGML_UNUSED(user_data);
+
+    const struct ggml_tensor * src0 = t->src[0];
+    const struct ggml_tensor * src1 = t->src[1];
+    std::string wname = filter_tensor_name(src0->name);
+
+    // when ask is true, the scheduler wants to know if we are interested in data from this tensor
+    // if we return true, a follow-up call will be made with ask=false in which we can do the actual collection
+    if (ask) {
+        if (t->op == GGML_OP_MUL_MAT_ID) return true; // collect all indirect matrix multiplications
+        if (t->op != GGML_OP_MUL_MAT) return false;
+        // why are small batches ignored (<16 tokens)?
+        if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return false;
+        if (!(wname.substr(0, 4) == "blk." || (m_params.process_output && wname == "output.weight"))) return false;
+        return true;
+    }
+
+    std::lock_guard<std::mutex> lock(m_mutex);
+
+    // copy the data from the GPU memory if needed
+    const bool is_host = ggml_backend_buffer_is_host(src1->buffer);
+
+    if (!is_host) {
+        m_src1_data.resize(ggml_nelements(src1));
+        ggml_backend_tensor_get(src1, m_src1_data.data(), 0, ggml_nbytes(src1));
+    }
+
+    const float * data = is_host ? (const float *) src1->data : m_src1_data.data();
+
+    // this has been adapted to the new format of storing merged experts in a single 3d tensor
+    // ref: https://github.com/ggerganov/llama.cpp/pull/6387
+    if (t->op == GGML_OP_MUL_MAT_ID) {
+        //   ids  -> [n_experts_used, n_tokens]
+        //   src1 -> [cols, n_expert_used, n_tokens]
+        const ggml_tensor * ids = t->src[2];
+        const int n_as = src0->ne[2];
+        const int n_ids = ids->ne[0];
+
+        // the top-k selected expert ids are stored in the ids tensor
+        // for simplicity, always copy ids to host, because it is small
+        // take into account that ids is not contiguous!
+
+        GGML_ASSERT(ids->ne[1] == src1->ne[2]);
+
+        m_ids.resize(ggml_nbytes(ids));
+        ggml_backend_tensor_get(ids, m_ids.data(), 0, ggml_nbytes(ids));
+
+        auto & e = m_stats[wname];
+
+        ++e.ncall;
+
+        if (e.values.empty()) {
+            e.values.resize(src1->ne[0]*n_as, 0);
+            e.counts.resize(src1->ne[0]*n_as, 0);
+        }
+        else if (e.values.size() != (size_t)src1->ne[0]*n_as) {
+            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
+            exit(1); //GGML_ABORT("fatal error");
+        }
+        if (m_params.verbosity > 1) {
+            printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[2], (int)src1->type);
+        }
+        // loop over all possible experts, regardless if they are used or not in the batch
+        for (int ex = 0; ex < n_as; ++ex) {
+            size_t e_start = ex*src1->ne[0];
+
+            for (int idx = 0; idx < n_ids; ++idx) {
+                for (int row = 0; row < (int)src1->ne[2]; ++row) {
+                    const int excur = *(const int32_t *) (m_ids.data() + row*ids->nb[1] + idx*ids->nb[0]);
+
+                    GGML_ASSERT(excur >= 0 && excur < n_as); // sanity check
+
+                    if (excur != ex) continue;
+
+                    const int64_t i11 = idx % src1->ne[1];
+                    const int64_t i12 = row;
+                    const float * x = (const float *)((const char *)data + i11*src1->nb[1] + i12*src1->nb[2]);
+
+                    for (int j = 0; j < (int)src1->ne[0]; ++j) {
+                        e.values[e_start + j] += x[j]*x[j];
+                        e.counts[e_start + j]++;
+                        if (!std::isfinite(e.values[e_start + j])) {
+                            fprintf(stderr, "%f detected in %s\n", e.values[e_start + j], wname.c_str());
+                            exit(1);
+                        }
+                    }
+                }
+            }
+            if (e.ncall > m_last_call) {
+                m_last_call = e.ncall;
+                if (m_last_call % m_params.n_out_freq == 0) {
+                    save_imatrix();
+                }
+                if (m_params.n_save_freq > 0 && m_last_call%m_params.n_save_freq == 0) {
+                    save_imatrix(m_last_call);
+                }
+            }
+        }
+    } else {
+        auto & e = m_stats[wname];
+        if (e.values.empty()) {
+            e.values.resize(src1->ne[0], 0);
+            e.counts.resize(src1->ne[0], 0);
+        }
+        else if (e.values.size() != (size_t)src1->ne[0]) {
+            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
+            exit(1); //GGML_ABORT("fatal error");
+        }
+        ++e.ncall;
+        if (m_params.verbosity > 1) {
+            printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
+        }
+        for (int row = 0; row < (int)src1->ne[1]; ++row) {
+            const float * x = data + row * src1->ne[0];
+            for (int j = 0; j < (int)src1->ne[0]; ++j) {
+                e.values[j] += x[j]*x[j];
+                e.counts[j]++;
+                if (!std::isfinite(e.values[j])) {
+                    fprintf(stderr, "%f detected in %s\n", e.values[j], wname.c_str());
+                    exit(1);
+                }
+            }
+        }
+        if (e.ncall > m_last_call) {
+            m_last_call = e.ncall;
+            if (m_last_call % m_params.n_out_freq == 0) {
+                save_imatrix();
+            }
+            if (m_params.n_save_freq > 0 && m_last_call%m_params.n_save_freq == 0) {
+                save_imatrix(m_last_call);
+            }
+        }
+    }
+
+    return true;
+}
+
+void IMatrixCollector::save_imatrix(int ncall) const {
+    auto fname = m_params.out_file;
+    if (fname.empty()) {
+        fname = "imatrix.dat";
+    }
+
+    if (ncall > 0) {
+        fname += ".at_";
+        fname += std::to_string(ncall);
+    }
+
+    // avoid writing imatrix entries that do not have full data
+    // this can happen with MoE models where some of the experts end up not being exercised by the provided training data
+
+    int n_entries = 0;
+    std::vector<std::string> to_store;
+
+    bool is_first = true; // for printing
+    for (const auto & kv : m_stats) {
+        const int n_all = kv.second.counts.size();
+
+        if (n_all == 0) {
+            continue;
+        }
+
+        int n_zeros = 0;
+        for (const int c : kv.second.counts) {
+            if (c == 0) {
+                n_zeros++;
+            }
+        }
+
+        if (n_zeros != 0 && is_first) {
+            fprintf(stderr, "\n");
+            is_first = false;
+        }
+
+        if (n_zeros == n_all) {
+            fprintf(stderr, "%s: entry '%40s' has no data - skipping\n", __func__, kv.first.c_str());
+            continue;
+        }
+
+        if (n_zeros > 0) {
+            fprintf(stderr, "%s: entry '%40s' has partial data (%.2f%%) - skipping\n", __func__, kv.first.c_str(), 100.0f * (n_all - n_zeros) / n_all);
+            continue;
+        }
+
+        n_entries++;
+        to_store.push_back(kv.first);
+    }
+
+    if (to_store.size() < m_stats.size()) {
+        fprintf(stderr, "%s: warning: storing only %zu out of %zu entries\n", __func__, to_store.size(), m_stats.size());
+    }
+
+    std::ofstream out(fname, std::ios::binary);
+    out.write((const char *) &n_entries, sizeof(n_entries));
+    for (const auto & name : to_store) {
+        const auto & stat = m_stats.at(name);
+        int len = name.size();
+        out.write((const char *) &len, sizeof(len));
+        out.write(name.c_str(), len);
+        out.write((const char *) &stat.ncall, sizeof(stat.ncall));
+        int nval = stat.values.size();
+        out.write((const char *) &nval, sizeof(nval));
+        if (nval > 0) {
+            std::vector<float> tmp(nval);
+            for (int i = 0; i < nval; i++) {
+                tmp[i] = (stat.values[i] / static_cast<float>(stat.counts[i])) * static_cast<float>(stat.ncall);
+            }
+            out.write((const char*)tmp.data(), nval*sizeof(float));
+        }
+    }
+
+    // Write the number of call the matrix was computed with
+    out.write((const char *) &m_last_call, sizeof(m_last_call));
+
+    // Write the input filename at the end of the file to later on specify it in quantize
+    {
+        int len = m_params.prompt_file.size();
+        out.write((const char *) &len, sizeof(len));
+        out.write(m_params.prompt_file.c_str(), len);
+    }
+
+    if (m_params.verbosity > 0) {
+        fprintf(stderr, "\n%s: stored collected data after %d chunks in %s\n", __func__, m_last_call, fname.c_str());
+    }
+}
+
+bool IMatrixCollector::load_imatrix(const char * fname) {
+    std::ifstream in(fname, std::ios::binary);
+    if (!in) {
+        printf("%s: failed to open %s\n",__func__, fname);
+        return false;
+    }
+    int n_entries;
+    in.read((char*)&n_entries, sizeof(n_entries));
+    if (in.fail() || n_entries < 1) {
+        printf("%s: no data in file %s\n", __func__, fname);
+        return false;
+    }
+    for (int i = 0; i < n_entries; ++i) {
+        int len; in.read((char *)&len, sizeof(len));
+        std::vector<char> name_as_vec(len+1);
+        in.read((char *)name_as_vec.data(), len);
+        if (in.fail()) {
+            printf("%s: failed reading name for entry %d from %s\n",__func__,i+1, fname);
+            return false;
+        }
+        name_as_vec[len] = 0;
+        std::string name{name_as_vec.data()};
+        auto & e = m_stats[std::move(name)];
+        int ncall;
+        in.read((char*)&ncall, sizeof(ncall));
+        int nval;
+        in.read((char *)&nval, sizeof(nval));
+        if (in.fail() || nval < 1) {
+            printf("%s: failed reading number of values for entry %d\n",__func__,i);
+            m_stats = {};
+            return false;
+        }
+
+        if (e.values.empty()) {
+            e.values.resize(nval, 0);
+            e.counts.resize(nval, 0);
+        }
+
+        std::vector<float> tmp(nval);
+        in.read((char*)tmp.data(), nval*sizeof(float));
+        if (in.fail()) {
+            printf("%s: failed reading data for entry %d\n",__func__,i);
+            m_stats = {};
+            return false;
+        }
+
+        // Recreate the state as expected by save_imatrix(), and corerct for weighted sum.
+        for (int i = 0; i < nval; i++) {
+            e.values[i] += tmp[i];
+            e.counts[i] += ncall;
+        }
+        e.ncall += ncall;
+
+    }
+    return true;
+}
+
+static IMatrixCollector g_collector;
+
+static bool ik_collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {
+    return g_collector.collect_imatrix(t, ask, user_data);
+}
+
+
+struct results_log_softmax {
+    double log_softmax;
+    float  logit;
+    float  prob;
+};
+
+static std::vector<float> softmax(const std::vector<float> & logits) {
+    std::vector<float> probs(logits.size());
+    float max_logit = logits[0];
+    for (float v : logits) {
+        max_logit = std::max(max_logit, v);
+    }
+    double sum_exp = 0.0;
+    for (size_t i = 0; i < logits.size(); i++) {
+        // Subtract the maximum logit value from the current logit value for numerical stability
+        const float logit = logits[i] - max_logit;
+        const float exp_logit = expf(logit);
+        sum_exp += exp_logit;
+        probs[i] = exp_logit;
+    }
+    for (size_t i = 0; i < probs.size(); i++) {
+        probs[i] /= sum_exp;
+    }
+    return probs;
+}
+
+static results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) {
+    float max_logit = logits[0];
+    for (int i = 1; i < n_vocab; ++i) {
+        max_logit = std::max(max_logit, logits[i]);
+    }
+    double sum_exp = 0.0;
+    for (int i = 0; i < n_vocab; ++i) {
+        sum_exp += expf(logits[i] - max_logit);
+    }
+    return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp};
+}
+
+static void process_logits(
+    int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread> & workers,
+    double & nll, double & nll2, float * logit_history, float * prob_history) {
+    std::mutex mutex;
+    int counter = 0;
+    auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () {
+        double local_nll  = 0;
+        double local_nll2 = 0;
+        while (true) {
+            std::unique_lock<std::mutex> lock(mutex);
+            int i = counter++;
+            if (i >= n_token) {
+                nll += local_nll; nll2 += local_nll2;
+                break;
+            }
+            lock.unlock();
+            const results_log_softmax results = log_softmax(n_vocab, logits + i*n_vocab, tokens[i+1]);
+            const double v = -results.log_softmax;
+            local_nll += v;
+            local_nll2 += v*v;
+
+            logit_history[i] = results.logit;
+            prob_history[i]  = results.prob;
+        }
+    };
+    for (auto & w : workers) {
+        w = std::thread(compute);
+    }
+    compute();
+    for (auto & w : workers) {
+        w.join();
+    }
+}
+
+static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
+    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
+    const int n_ctx = llama_n_ctx(ctx);
+
+    auto tim1 = std::chrono::high_resolution_clock::now();
+    fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
+
+    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, true);
+
+    auto tim2 = std::chrono::high_resolution_clock::now();
+    fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
+
+    if (params.i_chunk > 0) {
+        if (size_t((params.i_chunk + 2)*n_ctx) >= tokens.size()) {
+            fprintf(stderr, "%s: there will be not enough tokens left after removing %d chunks\n", __func__, params.i_chunk);
+            return false;
+        }
+        fprintf(stderr, "%s: removing initial %d chunks (%d tokens)\n", __func__, params.i_chunk, params.i_chunk*n_ctx);
+        tokens.erase(tokens.begin(), tokens.begin() + params.i_chunk*n_ctx);
+    }
+
+    if (int(tokens.size()) < 2*n_ctx) {
+        fprintf(stderr, "%s: you need at least %d tokens for a context of %d tokens\n",__func__,2*n_ctx,
+                n_ctx);
+        fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
+        return false;
+    }
+
+    std::vector<float> logit_history;
+    std::vector<float> prob_history;
+
+    if (params.compute_ppl) {
+        logit_history.resize(tokens.size());
+        prob_history.resize(tokens.size());
+    }
+
+    const int n_chunk_max = tokens.size() / n_ctx;
+
+    const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+    const int n_batch = params.n_batch;
+
+    int count = 0;
+    double nll = 0.0;
+    double nll2 = 0.0;
+
+    fprintf(stderr, "%s: computing over %d chunks with batch_size %d\n", __func__, n_chunk, n_batch);
+
+    std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
+
+    const int num_batches = (n_ctx + n_batch - 1) / n_batch;
+
+    std::vector<float> logits;
+    if (params.compute_ppl && num_batches > 1) {
+        logits.reserve((size_t)n_ctx * n_vocab);
+    }
+
+    for (int i = 0; i < n_chunk; ++i) {
+        const int start =     i * n_ctx;
+        const int end   = start + n_ctx;
+
+        std::vector<float> logits;
+
+        const auto t_start = std::chrono::high_resolution_clock::now();
+
+        // clear the KV cache
+        llama_kv_cache_clear(ctx);
+
+        for (int j = 0; j < num_batches; ++j) {
+            const int batch_start = start + j * n_batch;
+            const int batch_size  = std::min(end - batch_start, n_batch);
+
+            // save original token and restore it after eval
+            const auto token_org = tokens[batch_start];
+
+            // add BOS token for the first batch of each chunk
+            if (add_bos && j == 0) {
+                tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
+            }
+
+            // TODO: use batch.logits to save computations instead of relying on logits_all == true
+            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
+                fprintf(stderr, "%s : failed to eval\n", __func__);
+                return false;
+            }
+
+            // restore the original token in case it was set to BOS
+            tokens[batch_start] = token_org;
+
+            if (params.compute_ppl && num_batches > 1) {
+                const auto * batch_logits = llama_get_logits(ctx);
+                logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
+            }
+        }
+
+        const auto t_end = std::chrono::high_resolution_clock::now();
+
+        if (i == 0) {
+            const float t_total = std::chrono::duration<float>(t_end - t_start).count();
+            fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
+            int total_seconds = (int)(t_total * n_chunk);
+            if (total_seconds >= 60*60) {
+                fprintf(stderr, "%d hours ", total_seconds / (60*60));
+                total_seconds = total_seconds % (60*60);
+            }
+            fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
+        }
+
+        if (params.compute_ppl) {
+            const int first = n_ctx/2;
+            const auto all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
+            process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
+                    workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
+            count += n_ctx - first - 1;
+
+            printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
+            fflush(stdout);
+
+            logits.clear();
+        }
+    }
+    printf("\n");
+
+    if (params.compute_ppl) {
+        nll2 /= count;
+        nll /= count;
+        const double ppl = exp(nll);
+        nll2 -= nll * nll;
+        if (nll2 > 0) {
+            nll2 = sqrt(nll2/(count-1));
+            printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
+        } else {
+            printf("Unexpected negative standard deviation of log(prob)\n");
+        }
+    }
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+
+    params.n_ctx = 512;
+    params.logits_all = true;
+    params.verbosity = 1;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        print_usage(argc, argv, params);
+        return 1;
+    }
+
+    params.n_batch = std::min(params.n_batch, params.n_ctx);
+
+    g_collector.set_params(params);
+
+    for (const auto & in_file : params.in_files) {
+        printf("%s : loading imatrix from '%s'\n", __func__, in_file.c_str());
+        if (!g_collector.load_imatrix(in_file.c_str())) {
+            fprintf(stderr, "%s : failed to load %s\n", __func__, in_file.c_str());
+            return 1;
+        }
+    }
+
+    if (params.in_files.size() > 1) {
+        printf("%s : saving combined imatrix to '%s'\n", __func__, params.out_file.c_str());
+        g_collector.save_imatrix();
+    }
+
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    // pass the callback to the backend scheduler
+    // it will be executed for each node during the graph computation
+    params.cb_eval = ik_collect_imatrix;
+    params.cb_eval_user_data = NULL;
+    params.warmup = false;
+
+    // init
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
+    if (model == nullptr || ctx == nullptr) {
+        fprintf(stderr, "%s : failed to init\n", __func__);
+        return 1;
+    }
+
+    const int n_ctx_train = llama_n_ctx_train(model);
+    if (params.n_ctx > n_ctx_train) {
+        fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
+                __func__, n_ctx_train, params.n_ctx);
+    }
+
+    // print system information
+    {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
+    }
+
+    if (!compute_imatrix(ctx, params)) {
+        return 1;
+    }
+
+    g_collector.save_imatrix();
+
+    llama_print_timings(ctx);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    return 0;
+}

examples/infill/CMakeLists.txt

+set(TARGET llama-infill)
+add_executable(${TARGET} infill.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/infill/README.md

+# llama.cpp/example/infill
+
+This example shows how to use the infill mode with Code Llama models supporting infill mode.
+Currently the 7B and 13B models support infill mode.
+
+Infill supports most of the options available in the main example.
+
+For further information have a look at the main README.md in llama.cpp/example/main/README.md
+
+## Common Options
+
+In this section, we cover the most commonly used options for running the `infill` program with the LLaMA models:
+
+-   `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
+-   `-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.
+-   `--spm-infill`: Use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this.
+
+## Input Prompts
+
+The `infill` program provides several ways to interact with the LLaMA models using input prompts:
+
+-   `--in-prefix PROMPT_BEFORE_CURSOR`: Provide the prefix directly as a command-line option.
+-   `--in-suffix PROMPT_AFTER_CURSOR`: Provide the suffix directly as a command-line option.
+-   `--interactive-first`: Run the program in interactive mode and wait for input right away. (More on this below.)
+
+## Interaction
+
+The `infill` program offers a seamless way to interact with LLaMA models, allowing users to receive real-time infill suggestions. The interactive mode can be triggered using `--interactive`, and `--interactive-first`
+
+### Interaction Options
+
+-   `-i, --interactive`: Run the program in interactive mode, allowing users to get real time code suggestions from model.
+-   `--interactive-first`: Run the program in interactive mode and immediately wait for user input before starting the text generation.
+-   `--color`: Enable colorized output to differentiate visually distinguishing between prompts, user input, and generated text.
+
+### Example
+
+Download a model that supports infill, for example CodeLlama:
+```console
+scripts/hf.sh --repo TheBloke/CodeLlama-13B-GGUF --file codellama-13b.Q5_K_S.gguf --outdir models
+```
+
+```bash
+./llama-infill -t 10 -ngl 0 -m models/codellama-13b.Q5_K_S.gguf -c 4096 --temp 0.7 --repeat_penalty 1.1 -n 20 --in-prefix "def helloworld():\n    print(\"hell" --in-suffix "\n   print(\"goodbye world\")\n    "
+```

examples/infill/infill.cpp

+#include "common.h"
+
+#include "console.h"
+#include "llama.h"
+#include "grammar-parser.h"
+
+#include <cassert>
+#include <cinttypes>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <iostream>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+#include <signal.h>
+#include <unistd.h>
+#elif defined (_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#include <signal.h>
+#endif
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+static llama_context           ** g_ctx;
+static llama_model             ** g_model;
+static gpt_params               * g_params;
+static std::vector<llama_token> * g_input_tokens;
+static std::ostringstream       * g_output_ss;
+static std::vector<llama_token> * g_output_tokens;
+
+static bool is_interacting = false;
+
+static void write_logfile(
+    const llama_context * ctx, const gpt_params & params, const llama_model * model,
+    const std::vector<llama_token> & input_tokens, const std::string & output,
+    const std::vector<llama_token> & output_tokens
+) {
+    if (params.logdir.empty()) {
+        return;
+    }
+
+    const std::string timestamp = string_get_sortable_timestamp();
+
+    const bool success = fs_create_directory_with_parents(params.logdir);
+    if (!success) {
+        fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
+                __func__, params.logdir.c_str());
+        return;
+    }
+
+    const std::string logfile_path = params.logdir + timestamp + ".yml";
+    FILE * logfile = fopen(logfile_path.c_str(), "w");
+
+    if (logfile == NULL) {
+        fprintf(stderr, "%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
+        return;
+    }
+
+    fprintf(logfile, "binary: infill\n");
+    char model_desc[128];
+    llama_model_desc(model, model_desc, sizeof(model_desc));
+    yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);
+
+    fprintf(logfile, "\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "# Generation Results #\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "\n");
+
+    yaml_dump_string_multiline(logfile, "output", output.c_str());
+    yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
+
+    llama_dump_timing_info_yaml(logfile, ctx);
+    fclose(logfile);
+}
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
+static void sigint_handler(int signo) {
+    if (signo == SIGINT) {
+        if (!is_interacting) {
+            is_interacting = true;
+        } else {
+            console::cleanup();
+            printf("\n");
+            llama_print_timings(*g_ctx);
+            write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
+            _exit(130);
+        }
+    }
+}
+#endif
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+    llama_sampling_params & sparams = params.sparams;
+    g_params = &params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
+        return 1;
+    }
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("infill", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
+
+    console::init(params.simple_io, params.use_color);
+    atexit([]() { console::cleanup(); });
+
+    if (params.logits_all) {
+        printf("\n************\n");
+        printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+
+    if (params.embedding) {
+        printf("\n************\n");
+        printf("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+
+    if (params.n_ctx != 0 && params.n_ctx < 8) {
+        LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
+        params.n_ctx = 8;
+    }
+    if (!params.interactive_first && (params.input_prefix.empty() && params.input_suffix.empty())) {
+        printf("\n************\n");
+        printf("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+
+    if (params.rope_freq_base != 0.0) {
+        LOG_TEE("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
+    }
+
+    if (params.rope_freq_scale != 0.0) {
+        LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
+    }
+
+    LOG_TEE("%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
+    LOG_TEE("%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);
+
+    if (params.seed == LLAMA_DEFAULT_SEED) {
+        params.seed = time(NULL);
+    }
+
+    LOG_TEE("%s: seed  = %u\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+
+    LOG("%s: llama backend init\n", __func__);
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    llama_model * model;
+    llama_context * ctx;
+
+    g_model = &model;
+    g_ctx = &ctx;
+
+    // load the model and apply lora adapter, if any
+    LOG("%s: load the model and apply lora adapter, if any\n", __func__);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    model = llama_init.model;
+    ctx = llama_init.context;
+
+    if (model == NULL) {
+        LOG_TEE("%s: error: unable to load model\n", __func__);
+        return 1;
+    }
+
+    const int n_ctx_train = llama_n_ctx_train(model);
+    const int n_ctx = llama_n_ctx(ctx);
+    LOG("n_ctx: %d\n", n_ctx);
+
+    if (n_ctx > n_ctx_train) {
+        LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
+                __func__, n_ctx_train, n_ctx);
+    }
+
+    // print system information
+    {
+        LOG_TEE("\n");
+        LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
+    }
+    const bool add_bos = llama_should_add_bos_token(model);
+    GGML_ASSERT(llama_add_eos_token(model) != 1);
+    LOG("add_bos: %d\n", add_bos);
+
+    std::vector<llama_token> embd_inp;
+    std::vector<llama_token> embd_end;
+    std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false);
+    std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false);
+
+    GGML_ASSERT(llama_token_prefix(model) >= 0);
+    GGML_ASSERT(llama_token_suffix(model) >= 0);
+
+    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
+    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
+
+    embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
+    embd_end = params.spm_infill ? inp_pfx : inp_sfx;
+    if (add_bos) {
+        embd_inp.insert(embd_inp.begin(), llama_token_bos(model));
+    }
+    embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
+
+    const llama_token middle_token = llama_token_middle(model);
+    if (middle_token >= 0) {
+        embd_inp.push_back(middle_token);
+    }
+
+    LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
+    LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
+    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
+
+    // Should not run without any tokens
+    if (embd_inp.empty()) {
+        embd_inp.push_back(llama_token_bos(model));
+        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
+    }
+
+    if ((int) embd_inp.size() > n_ctx - 4) {
+        LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
+        return 1;
+    }
+
+    // number of tokens to keep when resetting context
+    if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size()) {
+        params.n_keep = (int)embd_inp.size();
+    }
+
+    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx).c_str());
+    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx).c_str());
+
+
+    // enable interactive mode if interactive start is specified
+    if (params.interactive_first) {
+        params.interactive = true;
+    }
+
+    if (params.verbose_prompt) {
+        LOG_TEE("\n");
+        LOG_TEE("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+        LOG_TEE("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+        for (int i = 0; i < (int) embd_inp.size(); i++) {
+            LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
+        }
+
+        if (params.n_keep > 0) {
+        LOG_TEE("%s: static prompt based on n_keep: '", __func__);
+            for (int i = 0; i < params.n_keep; i++) {
+                LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
+            }
+            LOG_TEE("'\n");
+        }
+        LOG_TEE("\n");
+    }
+
+    if (params.interactive) {
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+        struct sigaction sigint_action;
+        sigint_action.sa_handler = sigint_handler;
+        sigemptyset (&sigint_action.sa_mask);
+        sigint_action.sa_flags = 0;
+        sigaction(SIGINT, &sigint_action, NULL);
+#elif defined (_WIN32)
+        auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
+            return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
+        };
+        SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
+#endif
+
+        LOG_TEE("%s: interactive mode on.\n", __func__);
+
+        if (params.input_prefix_bos) {
+            LOG_TEE("Input prefix with BOS\n");
+        }
+
+        if (!params.input_prefix.empty()) {
+            LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
+        }
+
+        if (!params.input_suffix.empty()) {
+            LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
+        }
+    }
+    LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
+    LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
+    LOG_TEE("\n\n");
+
+    LOG_TEE("\n#####  Infill mode  #####\n\n");
+    if (params.infill) {
+        printf("\n************\n");
+        printf("no need to specify '--infill', always running infill\n");
+        printf("************\n\n");
+    }
+    if (params.interactive) {
+        const char *control_message;
+        if (params.multiline_input) {
+            control_message = " - To return control to LLaMA, end your input with '\\'.\n"
+                              " - To return control without starting a new line, end your input with '/'.\n";
+        } else {
+            control_message = " - Press Return to return control to LLaMA.\n"
+                              " - To return control without starting a new line, end your input with '/'.\n"
+                              " - If you want to submit another line, end your input with '\\'.\n";
+        }
+        LOG_TEE("== Running in interactive mode. ==\n");
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
+        LOG_TEE(       " - Press Ctrl+C to interject at any time.\n");
+#endif
+        LOG_TEE(       "%s\n", control_message);
+
+        is_interacting = params.interactive_first;
+    }
+
+    bool input_echo = true;
+
+    int n_past     = 0;
+    int n_remain   = params.n_predict;
+    int n_consumed = 0;
+
+    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;
+
+    // the first thing we will do is to output the prompt, so set color accordingly
+    console::set_display(console::prompt);
+
+    std::vector<llama_token> embd;
+
+    struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
+
+    while (n_remain != 0 || params.interactive) {
+        // predict
+        if (!embd.empty()) {
+            // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via
+            // --prompt or --file which uses the same value.
+            int max_embd_size = n_ctx - 4;
+
+            // Ensure the input doesn't exceed the context size by truncating embd if necessary.
+            if ((int) embd.size() > max_embd_size) {
+                const int skipped_tokens = (int) embd.size() - max_embd_size;
+                embd.resize(max_embd_size);
+
+                console::set_display(console::error);
+                printf("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
+                console::set_display(console::reset);
+                fflush(stdout);
+            }
+
+            // infinite text generation via context swapping
+            // if we run out of context:
+            // - take the n_keep first tokens from the original prompt (via n_past)
+            // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
+            if (n_past + (int) embd.size() > n_ctx) {
+                if (params.n_predict == -2) {
+                    LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
+                    break;
+                }
+
+                const int n_left    = n_past - params.n_keep - 1;
+                const int n_discard = n_left/2;
+
+                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
+                    n_past, n_left, n_ctx, params.n_keep, n_discard);
+
+                llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_cache_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+
+                n_past -= n_discard;
+
+                LOG("after swap: n_past = %d\n", n_past);
+
+                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
+
+            }
+
+            // evaluate tokens in batches
+            // embd is typically prepared beforehand to fit within a batch, but not always
+            for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
+                int n_eval = (int) embd.size() - i;
+                if (n_eval > params.n_batch) {
+                    n_eval = params.n_batch;
+                }
+
+                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
+
+                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
+                    LOG_TEE("%s : failed to eval\n", __func__);
+                    return 1;
+                }
+
+                n_past += n_eval;
+
+                LOG("n_past = %d\n", n_past);
+            }
+
+        }
+
+        embd.clear();
+
+        if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
+            const llama_token id = llama_sampling_sample(ctx_sampling, ctx, nullptr);
+
+            llama_sampling_accept(ctx_sampling, ctx, id, true);
+
+            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
+
+            embd.push_back(id);
+
+            // echo this to console
+            input_echo = true;
+
+            // decrement remaining sampling budget
+            --n_remain;
+
+            LOG("n_remain: %d\n", n_remain);
+        } else {
+            // some user input remains from prompt or interaction, forward it to processing
+            LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
+            while ((int) embd_inp.size() > n_consumed) {
+                embd.push_back(embd_inp[n_consumed]);
+
+                // push the prompt in the sampling context in order to apply repetition penalties later
+                // for the prompt, we don't apply grammar rules
+                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], false);
+
+                ++n_consumed;
+                if ((int) embd.size() >= params.n_batch) {
+                    break;
+                }
+            }
+        }
+
+        // display text
+        if (input_echo) {
+            for (auto id : embd) {
+                const std::string token_str = llama_token_to_piece(ctx, id);
+                printf("%s", token_str.c_str());
+
+                if (embd.size() > 1) {
+                    input_tokens.push_back(id);
+                } else {
+                    output_tokens.push_back(id);
+                    output_ss << token_str;
+                }
+            }
+            fflush(stdout);
+        }
+        // reset color to default if we there is no pending user input
+        if (input_echo && (int) embd_inp.size() == n_consumed) {
+            console::set_display(console::reset);
+        }
+
+        // if not currently processing queued inputs;
+        if ((int) embd_inp.size() <= n_consumed) {
+            // deal with eot token in infill mode
+            if ((llama_sampling_last(ctx_sampling) == llama_token_eot(model) || is_interacting) && params.interactive){
+                if (is_interacting && !params.interactive_first) {
+                    // print an eot token
+                    printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
+                }
+                fflush(stdout);
+                printf("\n");
+                console::set_display(console::user_input);
+                std::string buffer;
+                std::string line;
+                bool another_line=true;
+                // set a new prefix via stdin
+                do {
+                    another_line = console::readline(line, params.multiline_input);
+                    buffer += line;
+                } while (another_line);
+                // check if we got an empty line, if so we use the old input
+                if (!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
+                    params.input_prefix = buffer;
+                }
+                buffer.clear();
+                // set a new suffix via stdin
+                do {
+                    another_line = console::readline(line, params.multiline_input);
+                    buffer += line;
+                } while (another_line);
+                // check if we got an empty line
+                if (!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
+                    params.input_suffix = buffer;
+                }
+                buffer.clear();
+                // done taking input, reset color
+                console::set_display(console::reset);
+
+                if (params.escape) {
+                    //process escape sequences, for the initial prompt this is done in common.cpp when we load the params, but for the interactive mode we need to do it here
+                    string_process_escapes(params.input_prefix);
+                    string_process_escapes(params.input_suffix);
+                }
+
+                // tokenize new prefix and suffix
+                std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false);
+                std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false);
+
+                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
+                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
+
+                embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
+                embd_end = params.spm_infill ? inp_pfx : inp_sfx;
+                if (add_bos) {
+                    embd_inp.insert(embd_inp.begin(), llama_token_bos(model));
+                }
+                embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
+
+                if (middle_token >= 0) {
+                    embd_inp.push_back(middle_token);
+                }
+
+                embd.clear();
+                n_remain = params.n_predict;
+                n_past = 0;
+                n_consumed = 0;
+                // LOG_TEE("took new input\n");
+                is_interacting = false;
+            }
+            // deal with end of generation tokens in interactive mode
+            else if (llama_token_is_eog(model, llama_sampling_last(ctx_sampling))) {
+                LOG("found EOS token\n");
+
+                if (params.interactive) {
+
+                    is_interacting = true;
+                    printf("\n");
+                    console::set_display(console::user_input);
+                    fflush(stdout);
+               }
+            }
+
+            if (n_past > 0 && is_interacting && !params.interactive) {
+                LOG("waiting for user input\n");
+
+                if (params.input_prefix_bos) {
+                    LOG("adding input prefix BOS token\n");
+                    embd_inp.push_back(llama_token_bos(model));
+                }
+
+                std::string buffer;
+                if (!params.input_prefix.empty()) {
+                    LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
+                    buffer += params.input_prefix;
+                    printf("%s", buffer.c_str());
+                }
+
+                std::string line;
+                bool another_line = true;
+                do {
+                    another_line = console::readline(line, params.multiline_input);
+                    buffer += line;
+                } while (another_line);
+
+                // done taking input, reset color
+                console::set_display(console::reset);
+
+                // Add tokens to embd only if the input buffer is non-empty
+                // Entering a empty line lets the user pass control back
+                if (buffer.length() > 1) {
+                    // append input suffix if any
+                    if (!params.input_suffix.empty()) {
+                        LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
+                        buffer += params.input_suffix;
+                        printf("%s", params.input_suffix.c_str());
+                    }
+
+                    LOG("buffer: '%s'\n", buffer.c_str());
+
+                    const size_t original_size = embd_inp.size();
+
+                    const auto line_inp = ::llama_tokenize(ctx, buffer, false);
+                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
+
+                    embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
+
+                    for (size_t i = original_size; i < embd_inp.size(); ++i) {
+                        const llama_token token = embd_inp[i];
+                        output_tokens.push_back(token);
+                        output_ss << llama_token_to_piece(ctx, token);
+                    }
+
+                    n_remain -= line_inp.size();
+                    LOG("n_remain: %d\n", n_remain);
+                } else {
+                    LOG("empty line, passing control back\n");
+                }
+
+                input_echo = false; // do not echo this again
+            }
+
+            if (n_past > 0) {
+                if (is_interacting) {
+                    llama_sampling_reset(ctx_sampling);
+                }
+                is_interacting = false;
+            }
+        }
+
+        // end of generation
+        if (!embd.empty() && llama_token_is_eog(model, embd.back()) && !params.interactive) {
+            break;
+        }
+
+        // In interactive mode, respect the maximum number of tokens and drop back to user input when reached.
+        // We skip this logic when n_predict == -1 (infinite) or -2 (stop at context size).
+        if (params.interactive && n_remain <= 0 && params.n_predict >= 0) {
+            n_remain = params.n_predict;
+            is_interacting = true;
+        }
+    }
+    if (!params.interactive && n_remain <= 0) {
+        printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
+        fflush(stdout);
+    }
+
+    llama_print_timings(ctx);
+    write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_sampling_free(ctx_sampling);
+    llama_backend_free();
+
+#ifndef LOG_DISABLE_LOGS
+    LOG_TEE("Log end\n");
+#endif // LOG_DISABLE_LOGS
+
+    return 0;
+}

examples/jeopardy/README.md

+# llama.cpp/example/jeopardy
+
+This is pretty much just a straight port of aigoopy/llm-jeopardy/ with an added graph viewer.
+
+The jeopardy test can be used to compare the fact knowledge of different models and compare them to each other. This is in contrast to some other tests, which test logical deduction, creativity, writing skills, etc.
+
+
+Step 1: Open jeopardy.sh and modify the following:
+```
+MODEL=(path to your model)
+MODEL_NAME=(name of your model)
+prefix=(basically, if you use vicuna it's Human: , if you use something else it might be User: , etc)
+opts=(add -instruct here if needed for your model, or anything else you want to test out)
+```
+Step 2: Run `jeopardy.sh` from the llama.cpp folder
+
+Step 3: Repeat steps 1 and 2 until you have all the results you need.
+
+Step 4: Run `graph.py`, and follow the instructions. At the end, it will generate your final graph.
+
+Note: The Human bar is based off of the full, original 100 sample questions. If you modify the question count or questions, it will not be valid.

examples/jeopardy/graph.py

+#!/usr/bin/env python3
+import matplotlib.pyplot as plt
+import os
+import csv
+
+labels = []
+numbers = []
+numEntries = 1
+
+rows = []
+
+
+def bar_chart(numbers, labels, pos):
+    plt.bar(pos, numbers, color='blue')
+    plt.xticks(ticks=pos, labels=labels)
+    plt.title("Jeopardy Results by Model")
+    plt.xlabel("Model")
+    plt.ylabel("Questions Correct")
+    plt.show()
+
+
+def calculatecorrect():
+    directory = os.fsencode("./examples/jeopardy/results/")
+    csv_reader = csv.reader(open("./examples/jeopardy/qasheet.csv", 'rt'), delimiter=',')
+    for row in csv_reader:
+        global rows
+        rows.append(row)
+    for listing in os.listdir(directory):
+        filename = os.fsdecode(listing)
+        if filename.endswith(".txt"):
+            file = open("./examples/jeopardy/results/" + filename, "rt")
+            global labels
+            global numEntries
+            global numbers
+            labels.append(filename[:-4])
+            numEntries += 1
+            i = 1
+            totalcorrect = 0
+            for line in file.readlines():
+                if line.strip() != "------":
+                    print(line)
+                else:
+                    print("Correct answer: " + rows[i][2] + "\n")
+                    i += 1
+                    print("Did the AI get the question right? (y/n)")
+                    if input() == "y":
+                        totalcorrect += 1
+            numbers.append(totalcorrect)
+
+
+if __name__ == '__main__':
+    calculatecorrect()
+    pos = list(range(numEntries))
+    labels.append("Human")
+    numbers.append(48.11)
+    bar_chart(numbers, labels, pos)
+    print(labels)
+    print(numbers)

examples/jeopardy/jeopardy.sh

+#!/bin/bash
+set -e
+
+MODEL=./models/ggml-vicuna-13b-1.1-q4_0.bin
+MODEL_NAME=Vicuna
+
+# exec options
+prefix="Human: " # Ex. Vicuna uses "Human: "
+opts="--temp 0 -n 80" # additional flags
+nl='
+'
+introduction="You will be playing a game of Jeopardy. Simply answer the question in the correct format (Ex. What is Paris, or Who is George Washington)."
+
+# file options
+question_file=./examples/jeopardy/questions.txt
+touch ./examples/jeopardy/results/$MODEL_NAME.txt
+output_file=./examples/jeopardy/results/$MODEL_NAME.txt
+
+counter=1
+
+echo 'Running'
+while IFS= read -r question
+do
+  exe_cmd="./llama-cli -p "\"$prefix$introduction$nl$prefix$question\"" "$opts" -m ""\"$MODEL\""" >> ""\"$output_file\""
+  echo $counter
+  echo "Current Question: $question"
+  eval "$exe_cmd"
+  echo -e "\n------" >> $output_file
+  counter=$((counter+1))
+done < "$question_file"

examples/jeopardy/qasheet.csv

+Index,Original Category,Original Correct Question,Model Prompt
+1,The Oscars,Who is John Williams?,Which actor Born in 1932 was the son of a percussionist in the CBS radio orchestra has been nominated for 53 Oscars?
+2,English Literature,What is Paradise Lost?,"What work in English Literature says: 'The mind is its own place, & in itself can make a heaven of hell, a hell of heaven. What matter where, if I be still the same'?"
+3,Writers’ Lesser-Known Works,Who is Niccolò Machiavelli?,"Known for more philosophical works, he wrote the play 'La Mandragola', in which Florentines are rewarded for immoral actions?"
+4,Exploration,What is Easter Island (Rapa Nui)?,"James Cook's account of a 1774 visit where records an object 'near 27 feet long, and upwards of 8 feet over the breast or shoulders'?"
+5,The Bill of Rights,What is the Eighth Amendment?,England's 'Bloody Assizes' & a 1685 life sentence for perjury were 2 main origins of which amendment to the U.S. Constitution?
+6,Nobel Peace Prize Winners,Who are Nelson Mandela & Desmond Tutu?,"Which nobel peace price winners each lived at times on Vilakazi St. in Soweto , so it claims to be the world's only street home to 2 Nobel Peace Prize winners?"
+7,Famous Names,Who is Walt Disney?,"In 1966, the year of who's death did he share plans for an experimental prototype community in Florida?"
+8,Geography,What is Colombia?,"Of the 13 nations through which the Equator passes, what is the only one whose coastline borders the Caribbean Sea?"
+9,Fashion History,What are rhinestones?,"Which decorative items in fashion history get their name from their origin in the port city of Strasbourg, on the border of France & Germany?"
+10,Movies of the ’80s,What is Driving Miss Daisy?,What 1980's movie is based on an off-Broadway play with just 3 characters and won the Best Picture Oscar & the actors in all 3 roles were nominated?
+11,Novelists,Who is John Grisham?,"A 2012 book review for which novelist noted subjects that 'sparked his ire': capital punishment, big tobacco & 'the plight of the unjustly convicted'?"
+12,20th Century Eponyms,What is the Maginot Line?,"A 1940 headline about what 20th Century Eponym included 'failure', 'liability when it came to offense' & 'stout hearts no match for tanks'?"
+13,City History,What is Stockholm?,"Over 700 years after its traditional 1252 founding date, what port city became associated with a psychological response?"
+14,Brand Names,What is Jacuzzi?,"The success of what brand has its roots with a hydrotherapy pump its cofounder created for his son, who had arthritis?"
+15,American Authors,Who is Washington Irving?,"In a periodical in 1807, what American Author called New York City 'Gotham, Gotham! Most enlightened of cities'?"
+16,Symbols,What is “less than”?,What symbol is a rotated V in math and a feeling of some marginalized or underrepresented people in society?
+17,Movie Theme Songs,Who is James Bond?,"Monty Norman, the composer of what character's theme, said the staccato riff conveyed sexiness, mystery & ruthlessness?"
+18,American Novelists,Who is Joseph Heller?,"What American Novelist served with an airman named Yohannan in World War II & despite what readers might think, he said he enjoyed his service?"
+19,Medieval Places,"What is Canterbury, England? (Canterbury Cathedral)","In what Medieval place did one of the participants in an 1170 event say, 'Let us away, knights; he will rise no more'?"
+20,Countries of Africa,What is Morocco?,"At one time a province of the Roman Empire, what African country kingdom is known to Arabic scholars as Al-Maghrib Al-Aqsa, 'the far west'?"
+21,Statehood,What is Wyoming?,Congress relented in 1890 after what prospective state said it would wait 100 years rather than come in without the women?
+22,1980s Movies,What is Raiders of the Lost Ark?,"A writer & producer of what movie said he wanted it to be like a Western or James Bond film, 'only it takes place in the 30s'?"
+23,Art Exhibitions,Who is Rembrandt?,In 1898 what's been called the first blockbuster art show was devoted to which artist & put on for Queen Wilhelmina's coronation?
+24,Countries of the World,What is Mongolia?,"Part of the largest contiguous land empire during the 1200s & 1300s, today what is the world's second-largest landlocked country?"
+25,Literature,What is “Howl”?,A 2006 book was titled 'The Poem That Changed America:' What 'Fifty Years Later'?
+26,Invasions,Who is William of Orange?,"Backed by 14,000 troops, who invaded England to restore, in his words, its 'religion, laws, and liberties'?"
+27,Landmarks,What is the Eiffel Tower?,"After its completion in the late 19th c., what was landmark was called 'a truly tragic street lamp' & a 'high & skinny pyramid of iron ladders'?"
+28,Geographic Name’s the Same,What is Dover?,"The busiest passenger port in the U.K., what shares its name with a capital of one of the original 13 states?"
+29,Names in the Bookstore,Who is Peter Mark Roget?,"This man made lists, perhaps to cope with depression; a set of lists he published in 1852 made whose name synonymous with a type of book?"
+30,U.S. History,Who is Dr. Samuel Mudd?,"An 1869 presidential pardon was granted to which man, due in part to a plea by the Medical Society of Harford County, Maryland?"
+31,American Literature,What is The Things They Carried?,"Letters, pocket knives, C rations & steel helmets are among the tangible items referred to in the title of what American literature modern war classic?"
+32,Nonfiction,What is The Communist Manifesto,"What nonfiction book has the line, 'The discovery of America…opened up fresh ground for the rising bourgeoisie'?"
+33, a new version was passed 81 years later,Laws in U.S. History,What is the Civil Rights Act?,,,,,,,,,,,,,,,,,,0, 2/3
+34,Names of Myth,Who is Helen of Troy?,"Whose brothers, Castor & Pollux, saved her after Theseus stole her away as a kid; a larger force would seek her later in life?"
+35,African Countries,What is Sudan?,"Once Africa's largest country in area, what African Country dropped to third in 2011 when a portion of it declared independence?"
+36,The Ancient World,What is Alexandria?,"The ancient writer Galen said books on ships arriving to what city's port were seized, originals kept & copies returned?"
+37,Famous Names,Who is Andy Warhol?,"For a special 1970s cookbook, who provided one simple recipe–a can of Campbell's tomato soup & 2 cans of milk?"
+38,People & Places,What is Guam?,"Thought to descend from people of Southeast Asia, the Chamorro make up what U.S. territory’s largest ethnic group?"
+39,Current World Leaders,What is the Philippines?,"In office from 2022, the president of what country has taken so many foreign trips a play on his name is 'Ferdinand Magellan Jr.'?"
+40,Writers & The South,Who is Tennessee Williams?,In 1939 which writer lived on Toulouse Street in the French Quarter & chose the professional name that bonded him to the South?
+41,National Parks,What is Yellowstone?,"What National Park is named for a river indigenous people called Mi tse a-da-zi, translated by French-speaking trappers as 'Pierre Jaune'?"
+42,Sports,Who are the Harlem Globetrotters?,"In 2010 who introduced the 4-point shot, 35 feet from the basket?"
+43,The U.S. Military,What is “Top Gun”?,Losses over Asia in the 1960s led to the establishment of the program known as what at a San Diego naval base in 1969?
+44,Art & Science,What is Halley’s Comet?,"A craft that visited what was named for Giotto, based on the story that 680 years earlier, the painter depicted it as the Star of Bethlehem?"
+45,Words From World War I,What is “tank”?,"In World War I, 'Cistern' & 'reservoir' were suggested names for what secret invention, but the British preferred this less clumsy monosyllable?"
+46,European History,What is Holy Roman Emperor?,"Until 1806, some German nobles included among their honors the title of 'Elector' for their role in selecting this personage?"
+47,Theater History,Who is Peter Pan?,"In 1904, wearing a harness, actress Nina Boucicault became the first to play what character onstage?"
+48,European Cities,What is Aachen?,"Alphabetically the first German city in encyclopedias, what was also the first one taken by the Allies in World War II?"
+49,Word Origins,What is mantra?,This Sanskrit word referring to a spoken word or phrase comes from a word for 'to think'?
+50,Inventions,What is barbed wire?,1917's 'Elements of Trench Warfare' said what Old West invention was 'difficult to destroy' & 'difficult to get through'?
+51,World War II,What is Schindler’s list?,"Mimi Reinhard, who never learned to type using more than 2 fingers, produced what in World War II with 1,100 names, including hers?"
+52, their offspring was the source of this mythical object,Mythology,What is the Golden Fleece?
+53,Literature,What is Pride and Prejudice?,"Published in 2011, P.D. James' final novel, 'Death Comes to Pemberley', was a sequel to what novel from 200 years earlier?"
+54, only these 2 west of the Mississippi River border each other,U.S. State Names,What are Oregon & Nevada?
+55,Word Origins,What is passion?,"Originally relating to a story of suffering, what word now more commonly refers to strong emotion of any kind?"
+56,World Cinema,What is La Vie en Rose?,"The 2007 biopic called 'La Môme' in France, meaning 'The Kid', was released in the U.S. under what other French title?"
+57,History,What is Santa Maria?,"Returning home in 1493, Columbus stopped in the Azores at an island with what name, also something he'd lost off the Haiti coast?"
+58,Landmarks,What is a kremlin?,Pskov & Nizhny Novgorod are 2 of the cities that have a fortress called what?
+59,Foreign-Born Authors,Who is Vladimir Nabokov?,In the 1950s the New York Times said what author 'is writing about all lust' & his lecherous narrator 'is all of us'?
+60,Astronomy & Geography,What is Capricorn?,"At the winter solstice, the sun is in Sagittarius; it once appeared in what constellation, giving a geographic feature its name?"
+61,Television,What is Law & Order?,"Mike Post combined the sound of a slamming jail door, an anvil & 100 men stomping on a floor for what television series that debuted in 1990?"
+62,British Landmarks,What is the Tower of London?,"Like Sir Thomas More, 3 16th century English queens are buried at what British location?"
+63,Early American History,What are witches?,"In 1692 Increase Mather wrote, 'It were better that ten suspected' of these who 'escape, than that one innocent person … be condemned'?"
+64,Geography Mnemonics,What are Arkansas and Louisiana?,"The Geography Mnemonic Mimal, sometimes said to be the silhouette of a chef or elf, stands for Minnesota, Iowa, Missouri, and what other 2 states?"
+65,Business Milestones,What is the Ford Model T?,"What was first sold in 1908, at a price equivalent to about $27,000 today?"
+66,In The Bookstore,Who is Tom Clancy?,The name of what author dead since 2013 now appears on books written by a former U.S. marshal & a former Apache helicopter pilot?
+67,Historic Art,What is the Bayeux Tapestry?,The artwork once known in France as 'la tapisserie de la Reine Mathilde' is better known as what?
+68,Pop Stars,Who is Madonna?,In 2022 which pop star became the first woman to have a Billboard Top 10 album in 5 decades starting with the 1980s?
+69,Classic Tale Characters,Who is Scheherazade?,"In one 19th century translation, what female classic tale character 'perceived the dawn of day and ceased' speaking nearly 1,000 times?"
+70,USA,What is Jack Daniel’s?,"Ironically, though what company founded in the 1860s is Moore County, Tennessee's largest employer, Moore is a dry county?"
+71,Historic People,Who was William Bligh?,"After a 1789 event, who wrote, 'My first determination was to seek a supply of…water at Tofoa, & afterwards to sail for Tongataboo'?"
+72,The Movies,What is The Godfather?,Laurence Olivier & Ernest Borgnine were considered for the lead role & Sergio Leone to direct for what film that turned 50 in 2022?
+73,Continental Geography,What is Colombia?,"Until a 1903 secession, what country's contiguous territory spanned 2 continents?"
+74,Foreign-Born Authors,Who is Isabel Allende?,"Early in her career which foreign-born author translated romance novels into Spanish, often changing the dialogue to make the heroines smarter?"
+75,Historic Crimes,What is the Mona Lisa?,"Saying it was stolen by Napoleon, self-styled Italian patriot Vincenzo Peruggia took what in 1911?"
+76,U.S. Bodies of Water,What is Lake Mead?,"Continuing a downward trend, in July 2022 what US body of water was at 27% capacity, its lowest level since 1937 when it was first being filled?"
+77,Gods & Goddesses,Who is Aurora (or Eos)?,"Each morning which goddess began her ride in her chariot across the sky ahead of her brother Sol, or Helios?"
+78,America At War,What is the Battle of New Orleans?,"Until the Civil War, the Jan. 8 date of what American battle of dubious military importance but big morale value was a national holiday?"
+79,Children’s Books,What is The Velveteen Rabbit?,"Which children's book title character is told 'By the time you are real, most of your hair has been loved off your eyes drop out & you get shabby'?"
+80,TV Finales,What is Grace and Frankie?,"In a TV reunion over 40 years in the making, Dolly Parton appeared as an angel named Agnes in the final episode of what comedy in 2022?"
+81,American Poems,Who is Evangeline?,"In an 1847 American poem what character sees her town of Grand-Pré burned, but finally reunites with her beau for a kiss before his death?"
+82,Famous Names,Who is Banksy?,"In 2001 who published a book called 'Banging Your Head Against a Brick Wall'; in 2002, 'Existencilism'?"
+83,Children’s Lit,What is Charlotte’s Web?,The title object of what childrens book 'never looked more beautiful each strand held dozens of bright drops of early morning dew'?
+84,Classic Songs,What is “Here Comes Santa Claus”?,The shouts of excited children at a 1946 holiday parade are said to have inspired what perennial classic song favorite?
+85,Brand Names,What are Milk Duds?,"Unable to make what candies perfectly round, the confectioner embraced this flawed name for the product?"
+86,Countries of the World,What is Italy?,"What country is home to 58 UNESCO World Heritage Sites, more than any other country; the sites include a volcano & a lagoon?"
+87,Action Movies,What is Die Hard?,"What action movie's last line is 'If this is their idea of Christmas, I gotta be here for New Years'?"
+88,Presidential Facts,Who is Woodrow Wilson?,Only 3 presidents have married while in office— John Tyler was the first & which one was the last?
+89,19th Century Americans,Who is Frederick Douglass?,"Demonstrating the dignity & humanity of Black Americans, who sat for 160 known photographs, the most of any American in the 19th century?"
+90,Latin Phrases,What is “quid pro quo”?,"Originally, which Latin 3-word phrase referred to when a doctor or apothecary substituted one medicine for another?"
+91,1970s Movies,What is Monty Python and the Holy Grail?,The 1975 premiere of what movie comedy advertised free coconuts for the first thousand in the audience?
+92,Name’s The Same,What is Manhattan?,"A cocktail, an island & a WWII venture originally called 'Development of Substitute Materials' all bear what name?"
+93,U.S. Presidents,Who is Calvin Coolidge?,"Which US President was sworn in twice as President within 2 years, first by his father & then later by a former U.S. President?"
+94,Plays,What is The Tempest?,A 1609 story in which an exiled king of Bulgaria creates a sea palace with his magic may have inspired the plot of what play?
+95,Landmarks,What is the Berlin Wall?,"In 2009, during a 20th anniversary celebration, what landmark was called 'an edifice of fear. On Nov. 9, it became a place of joy'?"
+96,World Capitals,"What is Vienna, Austria?","Among what world capital's nicknames are the 'City of Classical Music' &, possibly in honor of a famous resident from 1860 to 1938, the 'City of Dreams'?"
+97,Language & Its Meanings,What is a night owl?,"Now meaning someone with nocturnal habits, what catches a sleeping dove in Shakespeare's 'Lucrece'?"
+98,Flags of Our Hemisphere,What is Brazil?,"The stars on what country's flag represent states, 26 of them; unlike the USA's, its 'federal district' gets its own 27th star?"
+99,Names in U.S. History,Who is Oliver Brown?,What father was the only man among the 13 plaintiffs in a US class-action case filed in 1951?
+100,Children’s Authors,"Who is Sarah? (from Sarah, Plain and Tall)","Reversing the story of what heroine she created, childrens author Patricia Maclachlan was born on the prairie but spent much of her life in New England?"
+,,,
+TOTALS,,,

examples/jeopardy/questions.txt

+Which man born in 1932 was the son of a percussionist in the CBS radio orchestra has been nominated for 53 Oscars?
+What work in English Literature says: 'The mind is its own place, & in itself can make a heaven of hell, a hell of heaven. What matter where, if I be still the same'?
+Known for more philosophical works, he wrote the play 'La Mandragola', in which Florentines are rewarded for immoral actions?
+James Cook's account of a 1774 visit where records an object 'near 27 feet long, and upwards of 8 feet over the breast or shoulders'?
+England's 'Bloody Assizes' & a 1685 life sentence for perjury were 2 main origins of which amendment to the U.S. Constitution?
+Which nobel peace price winners each lived at times on Vilakazi St. in Soweto , so it claims to be the world's only street home to 2 Nobel Peace Prize winners?
+In 1966, the year of who's death did he share plans for an experimental prototype community in Florida?
+Of the 13 nations through which the Equator passes, what is the only one whose coastline borders the Caribbean Sea?
+Which decorative items in fashion history get their name from their origin in the port city of Strasbourg, on the border of France & Germany?
+What 1980's movie is based on an off-Broadway play with just 3 characters and won the Best Picture Oscar & the actors in all 3 roles were nominated?
+A 2012 book review for which novelist noted subjects that 'sparked his ire': capital punishment, big tobacco & 'the plight of the unjustly convicted'?
+A 1940 headline about what 20th Century Eponym included 'failure', 'liability when it came to offense' & 'stout hearts no match for tanks'?
+Over 700 years after its traditional 1252 founding date, what port city became associated with a psychological response?
+The success of what brand has its roots with a hydrotherapy pump its cofounder created for his son, who had arthritis?
+In a periodical in 1807, what American Author called New York City 'Gotham, Gotham! Most enlightened of cities'?
+What symbol is a rotated V in math and a feeling of some marginalized or underrepresented people in society?
+Monty Norman, the composer of what character's theme, said the staccato riff conveyed sexiness, mystery & ruthlessness?
+What American Novelist served with an airman named Yohannan in World War II & despite what readers might think, he said he enjoyed his service?
+In what Medieval place did one of the participants in an 1170 event say, 'Let us away, knights; he will rise no more'?
+At one time a province of the Roman Empire, what African country kingdom is known to Arabic scholars as Al-Maghrib Al-Aqsa, 'the far west'?
+Congress relented in 1890 after what prospective state said it would wait 100 years rather than come in without the women?
+A writer & producer of what movie said he wanted it to be like a Western or James Bond film, 'only it takes place in the 30s'?
+In 1898 what's been called the first blockbuster art show was devoted to which artist & put on for Queen Wilhelmina's coronation?
+Part of the largest contiguous land empire during the 1200s & 1300s, today what is the world's second-largest landlocked country?
+A 2006 book was titled 'The Poem That Changed America:' What 'Fifty Years Later'?
+Backed by 14,000 troops, who invaded England to restore, in his words, its 'religion, laws, and liberties'?
+After its completion in the late 19th c., what was landmark was called 'a truly tragic street lamp' & a 'high & skinny pyramid of iron ladders'?
+The busiest passenger port in the U.K., what shares its name with a capital of one of the original 13 states?
+This man made lists, perhaps to cope with depression; a set of lists he published in 1852 made whose name synonymous with a type of book?
+An 1869 presidential pardon was granted to which man, due in part to a plea by the Medical Society of Harford County, Maryland?
+Letters, pocket knives, C rations & steel helmets are among the tangible items referred to in the title of what American literature modern war classic?
+What nonfiction book has the line, 'The discovery of America…opened up fresh ground for the rising bourgeoisie'?
+A radical Republican championed what 1875 act but the Supreme Court struck it down in 1883; a new version was passed 81 years later?
+Whose brothers, Castor & Pollux, saved her after Theseus stole her away as a kid; a larger force would seek her later in life?
+Once Africa's largest country in area, what African Country dropped to third in 2011 when a portion of it declared independence?
+The ancient writer Galen said books on ships arriving to what city's port were seized, originals kept & copies returned?
+For a special 1970s cookbook, who provided one simple recipe–a can of Campbell's tomato soup & 2 cans of milk?
+Thought to descend from people of Southeast Asia, the Chamorro make up what U.S. territory’s largest ethnic group?
+In office from 2022, the president of what country has taken so many foreign trips a play on his name is 'Ferdinand Magellan Jr.'?
+In 1939 which writer lived on Toulouse Street in the French Quarter & chose the professional name that bonded him to the South?
+What National Park is named for a river indigenous people called Mi tse a-da-zi, translated by French-speaking trappers as 'Pierre Jaune'?
+In 2010 who introduced the 4-point shot, 35 feet from the basket?
+Losses over Asia in the 1960s led to the establishment of the program known as what at a San Diego naval base in 1969?
+A craft that visited what was named for Giotto, based on the story that 680 years earlier, the painter depicted it as the Star of Bethlehem?
+In World War I, 'Cistern' & 'reservoir' were suggested names for what secret invention, but the British preferred this less clumsy monosyllable?
+Until 1806, some German nobles included among their honors the title of 'Elector' for their role in selecting this personage?
+In 1904, wearing a harness, actress Nina Boucicault became the first to play what character onstage?
+Alphabetically the first German city in encyclopedias, what was also the first one taken by the Allies in World War II?
+This Sanskrit word referring to a spoken word or phrase comes from a word for 'to think'?
+1917's 'Elements of Trench Warfare' said what Old West invention was 'difficult to destroy' & 'difficult to get through'?
+Mimi Reinhard, who never learned to type using more than 2 fingers, produced what in World War II with 1,100 names, including hers?
+Poseidon carried off the maiden Theophane & turned her into a ewe; their offspring was the source of what mythical object?
+Published in 2011, P.D. James' final novel, 'Death Comes to Pemberley', was a sequel to what novel from 200 years earlier?
+5 U.S. states have 6-letter names; only which 2 west of the Mississippi River border each other?
+Originally relating to a story of suffering, what word now more commonly refers to strong emotion of any kind?
+The 2007 biopic called 'La Môme' in France, meaning 'The Kid', was released in the U.S. under what other French title?
+Returning home in 1493, Columbus stopped in the Azores at an island with what name, also something he'd lost off the Haiti coast?
+Pskov & Nizhny Novgorod are 2 of the cities that have a fortress called what?
+In the 1950s the New York Times said what author 'is writing about all lust' & his lecherous narrator 'is all of us'?
+At the winter solstice, the sun is in Sagittarius; it once appeared in what constellation, giving a geographic feature its name?
+Mike Post combined the sound of a slamming jail door, an anvil & 100 men stomping on a floor for what television series that debuted in 1990?
+Like Sir Thomas More, 3 16th century English queens are buried at what British location?
+In 1692 Increase Mather wrote, 'It were better that ten suspected' of these who 'escape, than that one innocent person be condemned'?
+The Geography Mnemonic Mimal, sometimes said to be the silhouette of a chef or elf, stands for Minnesota, Iowa, Missouri, and what other 2 states?
+What was first sold in 1908, at a price equivalent to about $27,000 today?
+The name of what author dead since 2013 now appears on books written by a former U.S. marshal & a former Apache helicopter pilot?
+The artwork once known in France as 'la tapisserie de la Reine Mathilde' is better known as what?
+In 2022 which pop star became the first woman to have a Billboard Top 10 album in 5 decades starting with the 1980s?
+In one 19th century translation, what female classic tale character 'perceived the dawn of day and ceased' speaking nearly 1,000 times?
+Ironically, though what company founded in the 1860s is Moore County, Tennessee's largest employer, Moore is a dry county?
+After a 1789 event, who wrote, 'My first determination was to seek a supply of…water at Tofoa, & afterwards to sail for Tongataboo'?
+Laurence Olivier & Ernest Borgnine were considered for the lead role & Sergio Leone to direct for what film that turned 50 in 2022?
+Until a 1903 secession, what country's contiguous territory spanned 2 continents?
+Early in her career which foreign-born author translated romance novels into Spanish, often changing the dialogue to make the heroines smarter?
+Saying it was stolen by Napoleon, self-styled Italian patriot Vincenzo Peruggia took what in 1911?
+Continuing a downward trend, in July 2022 what US body of water was at 27% capacity, its lowest level since 1937 when it was first being filled?
+Each morning which goddess began her ride in her chariot across the sky ahead of her brother Sol, or Helios?
+Until the Civil War, the Jan. 8 date of what American battle of dubious military importance but big morale value was a national holiday?
+Which children's book title character is told 'By the time you are real, most of your hair has been loved off your eyes drop out & you get shabby'?
+In a TV reunion over 40 years in the making, Dolly Parton appeared as an angel named Agnes in the final episode of what comedy in 2022?
+In an 1847 American poem what character sees her town of Grand-Pré burned, but finally reunites with her beau for a kiss before his death?
+In 2001 who published a book called 'Banging Your Head Against a Brick Wall'; in 2002, 'Existencilism'?
+The title object of what childrens book 'never looked more beautiful each strand held dozens of bright drops of early morning dew'?
+The shouts of excited children at a 1946 holiday parade are said to have inspired what perennial classic song favorite?
+Unable to make what candies perfectly round, the confectioner embraced this flawed name for the product?
+What country is home to 58 UNESCO World Heritage Sites, more than any other country; the sites include a volcano & a lagoon?
+What action movie's last line is 'If this is their idea of Christmas, I gotta be here for New Years'?
+Only 3 presidents have married while in office— John Tyler was the first & which one was the last?
+Demonstrating the dignity & humanity of Black Americans, who sat for 160 known photographs, the most of any American in the 19th century?
+Originally, which Latin 3-word phrase referred to when a doctor or apothecary substituted one medicine for another?
+The 1975 premiere of what movie comedy advertised free coconuts for the first thousand in the audience?
+A cocktail, an island & a WWII venture originally called 'Development of Substitute Materials' all bear what name?
+Which US President was sworn in twice as President within 2 years, first by his father & then later by a former U.S. President?
+A 1609 story in which an exiled king of Bulgaria creates a sea palace with his magic may have inspired the plot of what play?
+In 2009, during a 20th anniversary celebration, what landmark was called 'an edifice of fear. On Nov. 9, it became a place of joy'?
+Among what world capital's nicknames are the 'City of Classical Music' &, possibly in honor of a famous resident from 1860 to 1938, the 'City of Dreams'?
+Now meaning someone with nocturnal habits, what catches a sleeping dove in Shakespeare's 'Lucrece'?
+The stars on what country's flag represent states, 26 of them; unlike the USA's, its 'federal district' gets its own 27th star?
+What father was the only man among the 13 plaintiffs in a US class-action case filed in 1951?
+Reversing the story of what heroine she created, childrens author Patricia Maclachlan was born on the prairie but spent much of her life in New England?

examples/json_schema_pydantic_example.py

+# Usage:
+#! ./llama-server -m some-model.gguf &
+#! pip install pydantic
+#! python json_schema_pydantic_example.py
+
+from pydantic import BaseModel, Field, TypeAdapter
+from annotated_types import MinLen
+from typing import Annotated, List, Optional
+import json, requests
+
+if True:
+
+    def create_completion(*, response_model=None, endpoint="http://localhost:8080/v1/chat/completions", messages, **kwargs):
+        '''
+        Creates a chat completion using an OpenAI-compatible endpoint w/ JSON schema support
+        (llama.cpp server, llama-cpp-python, Anyscale / Together...)
+
+        The response_model param takes a type (+ supports Pydantic) and behaves just as w/ Instructor (see below)
+        '''
+        response_format = None
+        type_adapter = None
+
+        if response_model:
+            type_adapter = TypeAdapter(response_model)
+            schema = type_adapter.json_schema()
+            messages = [{
+                "role": "system",
+                "content": f"You respond in JSON format with the following schema: {json.dumps(schema, indent=2)}"
+            }] + messages
+            response_format={"type": "json_object", "schema": schema}
+
+        data = requests.post(endpoint, headers={"Content-Type": "application/json"},
+                             json=dict(messages=messages, response_format=response_format, **kwargs)).json()
+        if 'error' in data:
+            raise Exception(data['error']['message'])
+
+        content = data["choices"][0]["message"]["content"]
+        return type_adapter.validate_json(content) if type_adapter else content
+
+else:
+
+    # This alternative branch uses Instructor + OpenAI client lib.
+    # Instructor support streamed iterable responses, retry & more.
+    # (see https://python.useinstructor.com/)
+    #! pip install instructor openai
+    import instructor, openai
+    client = instructor.patch(
+        openai.OpenAI(api_key="123", base_url="http://localhost:8080"),
+        mode=instructor.Mode.JSON_SCHEMA)
+    create_completion = client.chat.completions.create
+
+
+if __name__ == '__main__':
+
+    class QAPair(BaseModel):
+        class Config:
+            extra = 'forbid'  # triggers additionalProperties: false in the JSON schema
+        question: str
+        concise_answer: str
+        justification: str
+        stars: Annotated[int, Field(ge=1, le=5)]
+
+    class PyramidalSummary(BaseModel):
+        class Config:
+            extra = 'forbid'  # triggers additionalProperties: false in the JSON schema
+        title: str
+        summary: str
+        question_answers: Annotated[List[QAPair], MinLen(2)]
+        sub_sections: Optional[Annotated[List['PyramidalSummary'], MinLen(2)]]
+
+    print("# Summary\n", create_completion(
+        model="...",
+        response_model=PyramidalSummary,
+        messages=[{
+            "role": "user",
+            "content": f"""
+                You are a highly efficient corporate document summarizer.
+                Create a pyramidal summary of an imaginary internal document about our company processes
+                (starting high-level, going down to each sub sections).
+                Keep questions short, and answers even shorter (trivia / quizz style).
+            """
+        }]))

examples/json_schema_to_grammar.py

+#!/usr/bin/env python3
+from __future__ import annotations
+
+import argparse
+import itertools
+import json
+import re
+import sys
+from typing import Any, List, Optional, Set, Tuple, Union
+
+def _build_repetition(item_rule, min_items, max_items, separator_rule=None):
+
+    if min_items == 0 and max_items == 1:
+        return f'{item_rule}?'
+
+    if not separator_rule:
+        if min_items == 1 and max_items is None:
+            return f'{item_rule}+'
+        elif min_items == 0 and max_items is None:
+            return f'{item_rule}*'
+        else:
+            return f'{item_rule}{{{min_items},{max_items if max_items is not None else ""}}}'
+
+    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 = None):
+        self.content = content
+        self.deps = deps or []
+
+# Constraining spaces to prevent model "running away".
+SPACE_RULE = '| " " | "\\n" [ \\t]{0,20}'
+
+PRIMITIVE_RULES = {
+    'boolean'      : BuiltinRule('("true" | "false") space', []),
+    'decimal-part' : BuiltinRule('[0-9]{1,16}', []),
+    'integral-part': BuiltinRule('[0] | [1-9] [0-9]{0,15}', []),
+    'number'       : BuiltinRule('("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space', ['integral-part', 'decimal-part']),
+    'integer'      : BuiltinRule('("-"? integral-part) space', ['integral-part']),
+    'value'        : BuiltinRule('object | array | string | number | boolean | null', ['object', 'array', 'string', 'number', 'boolean', 'null']),
+    'object'       : BuiltinRule('"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', ['string', 'value']),
+    'array'        : BuiltinRule('"[" space ( value ("," space value)* )? "]" space', ['value']),
+    'uuid'         : BuiltinRule(r'"\"" [0-9a-fA-F]{8} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{12} "\"" space', []),
+    'char'         : BuiltinRule(r'[^"\\\x7F\x00-\x1F] | [\\] (["\\bfnrt] | "u" [0-9a-fA-F]{4})', []),
+    'string'       : BuiltinRule(r'"\"" char* "\"" space', ['char']),
+    'null'         : BuiltinRule('"null" space', []),
+}
+
+# TODO: support "uri", "email" string formats
+STRING_FORMAT_RULES = {
+    'date'            : BuiltinRule('[0-9]{4} "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', []),
+    'time'            : BuiltinRule('([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9]{3} )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', []),
+    'date-time'       : BuiltinRule('date "T" time', ['date', 'time']),
+    'date-string'     : BuiltinRule('"\\"" date "\\"" space', ['date']),
+    'time-string'     : BuiltinRule('"\\"" time "\\"" space', ['time']),
+    'date-time-string': BuiltinRule('"\\"" date-time "\\"" space', ['date-time']),
+}
+
+DOTALL = '[\\U00000000-\\U0010FFFF]'
+DOT = '[^\\x0A\\x0D]'
+
+RESERVED_NAMES = set(["root", "dot", *PRIMITIVE_RULES.keys(), *STRING_FORMAT_RULES.keys()])
+
+INVALID_RULE_CHARS_RE = re.compile(r'[^a-zA-Z0-9-]+')
+GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"]')
+GRAMMAR_RANGE_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"\]\-\\]')
+GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']': '\\]'}
+
+NON_LITERAL_SET = set('|.()[]{}*+?')
+ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('^$.[]()|{}*+?')
+
+
+class SchemaConverter:
+    def __init__(self, *, prop_order, allow_fetch, dotall, raw_pattern):
+        self._prop_order = prop_order
+        self._allow_fetch = allow_fetch
+        self._dotall = dotall
+        self._raw_pattern = raw_pattern
+        self._rules = {
+            'space': SPACE_RULE,
+        }
+        self._refs = {}
+        self._refs_being_resolved = set()
+
+    def _format_literal(self, literal):
+        escaped = GRAMMAR_LITERAL_ESCAPE_RE.sub(
+            lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)) or m.group(0), literal
+        )
+        return f'"{escaped}"'
+
+    def not_literal(self, literal: str, dotall: bool = True, maybe_escaped_underscores = False) -> str:
+        '''
+            not_literal('a') -> '[^a]'
+            not_literal('abc') -> '([^a] | "a" ([^b] | "b" ([^c])?)?)?'
+        '''
+        assert len(literal) > 0, 'Empty literal not supported'
+        def recurse(i: int):
+            c = literal[i]
+            if maybe_escaped_underscores and c == '_':
+                yield f'[^{c}\\\\]'
+                yield ' | '
+                yield f'"\\\\"? "{c}"'
+            else:
+                yield f'[^{c}]'
+            if i < len(literal) - 1:
+                yield ' | '
+                yield self._format_literal(c)
+                yield ' ('
+                yield from recurse(i + 1)
+                yield ')?'
+
+        return ''.join(('(', *recurse(0), ')'))
+
+    def _not_strings(self, strings):
+        class TrieNode:
+            def __init__(self):
+                self.children = {}
+                self.is_end_of_string = False
+
+            def insert(self, string):
+                node = self
+                for c in string:
+                    node = node.children.setdefault(c, TrieNode())
+                node.is_end_of_string = True
+
+        trie = TrieNode()
+        for s in strings:
+            trie.insert(s)
+
+        char_rule = self._add_primitive('char', PRIMITIVE_RULES['char'])
+        out = ['["] ( ']
+
+        def visit(node):
+            rejects = []
+            first = True
+            for c in sorted(node.children.keys()):
+                child = node.children[c]
+                rejects.append(c)
+                if first:
+                    first = False
+                else:
+                    out.append(' | ')
+                out.append(f'[{c}]')
+                if child.children:
+                    out.append(f' (')
+                    visit(child)
+                    out.append(')')
+                elif child.is_end_of_string:
+                    out.append(f' {char_rule}+')
+            if node.children:
+                if not first:
+                    out.append(' | ')
+                out.append(f'[^"{"".join(rejects)}] {char_rule}*')
+        visit(trie)
+
+        out.append(f' ){"" if trie.is_end_of_string else "?"} ["] space')
+        return ''.join(out)
+
+    def _add_rule(self, name, rule):
+        esc_name = INVALID_RULE_CHARS_RE.sub('-', name)
+        if esc_name not in self._rules or self._rules[esc_name] == rule:
+            key = esc_name
+        else:
+            i = 0
+            while f'{esc_name}{i}' in self._rules and self._rules[f'{esc_name}{i}'] != rule:
+                i += 1
+            key = f'{esc_name}{i}'
+        self._rules[key] = rule
+        return key
+
+    def resolve_refs(self, schema: dict, url: str):
+        '''
+            Resolves all $ref fields in the given schema, fetching any remote schemas,
+            replacing $ref with absolute reference URL and populating self._refs with the
+            respective referenced (sub)schema dictionaries.
+        '''
+        def visit(n: dict):
+            if isinstance(n, list):
+                return [visit(x) for x in n]
+            elif isinstance(n, dict):
+                ref = n.get('$ref')
+                if ref is not None and ref not in self._refs:
+                    if ref.startswith('https://'):
+                        assert self._allow_fetch, 'Fetching remote schemas is not allowed (use --allow-fetch for force)'
+                        import requests
+
+                        frag_split = ref.split('#')
+                        base_url = frag_split[0]
+
+                        target = self._refs.get(base_url)
+                        if target is None:
+                            target = self.resolve_refs(requests.get(ref).json(), base_url)
+                            self._refs[base_url] = target
+
+                        if len(frag_split) == 1 or frag_split[-1] == '':
+                            return target
+                    elif ref.startswith('#/'):
+                        target = schema
+                        ref = f'{url}{ref}'
+                        n['$ref'] = ref
+                    else:
+                        raise ValueError(f'Unsupported ref {ref}')
+
+                    for sel in ref.split('#')[-1].split('/')[1:]:
+                        assert target is not None and sel in target, f'Error resolving ref {ref}: {sel} not in {target}'
+                        target = target[sel]
+
+                    self._refs[ref] = target
+                else:
+                    for v in n.values():
+                        visit(v)
+
+            return n
+        return visit(schema)
+
+    def _generate_union_rule(self, name, alt_schemas):
+        return ' | '.join((
+            self.visit(alt_schema, f'{name}{"-" if name else "alternative-"}{i}')
+            for i, alt_schema in enumerate(alt_schemas)
+        ))
+
+    def _visit_pattern(self, pattern, name):
+        '''
+            Transforms a regular expression pattern into a GBNF rule.
+
+            Input: https://json-schema.org/understanding-json-schema/reference/regular_expressions
+            Output: https://github.com/ggerganov/llama.cpp/blob/master/grammars/README.md
+
+            Unsupported features: negative/positive lookaheads, greedy/non-greedy modifiers.
+
+            Mostly a 1:1 translation, except for {x} / {x,} / {x,y} quantifiers for which
+            we define sub-rules to keep the output lean.
+        '''
+
+        assert pattern.startswith('^') and pattern.endswith('$'), 'Pattern must start with "^" and end with "$"'
+        pattern = pattern[1:-1]
+        sub_rule_ids = {}
+
+        i = 0
+        length = len(pattern)
+
+        def to_rule(s: tuple[str, bool]) -> str:
+            (txt, is_literal) = s
+            return "\"" + txt + "\"" if is_literal else txt
+
+        def transform() -> tuple[str, bool]:
+            '''
+                Parse a unit at index i (advancing it), and return its string representation + whether it's a literal.
+            '''
+            nonlocal i
+            nonlocal pattern
+            nonlocal sub_rule_ids
+
+            start = i
+            # For each component of this sequence, store its string representation and whether it's a literal.
+            # We only need a flat structure here to apply repetition operators to the last item, and
+            # to merge literals at the and (we're parsing grouped ( sequences ) recursively and don't treat '|' specially
+            # (GBNF's syntax is luckily very close to regular expressions!)
+            seq: list[tuple[str, bool]] = []
+
+            def get_dot():
+                if self._dotall:
+                    rule = DOTALL
+                else:
+                    # Accept any character... except \n and \r line break chars (\x0A and \xOD)
+                    rule = DOT
+                return self._add_rule(f'dot', rule)
+
+            def join_seq():
+                nonlocal seq
+                ret = []
+                for is_literal, g in itertools.groupby(seq, lambda x: x[1]):
+                    if is_literal:
+                        ret.append((''.join(x[0] for x in g), True))
+                    else:
+                        ret.extend(g)
+                if len(ret) == 1:
+                    return ret[0]
+                return (' '.join(to_rule(x) for x in seq), False)
+
+            while i < length:
+                c = pattern[i]
+                if c == '.':
+                    seq.append((get_dot(), False))
+                    i += 1
+                elif c == '(':
+                    i += 1
+                    if i < length:
+                        assert pattern[i] != '?', f'Unsupported pattern syntax "{pattern[i]}" at index {i} of /{pattern}/'
+                    seq.append((f'({to_rule(transform())})', False))
+                elif c == ')':
+                    i += 1
+                    assert start > 0 and pattern[start-1] == '(', f'Unbalanced parentheses; start = {start}, i = {i}, pattern = {pattern}'
+                    return join_seq()
+                elif c == '[':
+                    square_brackets = c
+                    i += 1
+                    while i < length and pattern[i] != ']':
+                        if pattern[i] == '\\':
+                            square_brackets += pattern[i:i+2]
+                            i += 2
+                        else:
+                            square_brackets += pattern[i]
+                            i += 1
+                    assert i < length, f'Unbalanced square brackets; start = {start}, i = {i}, pattern = {pattern}'
+                    square_brackets += ']'
+                    i += 1
+                    seq.append((square_brackets, False))
+                elif c == '|':
+                    seq.append(('|', False))
+                    i += 1
+                elif c in ('*', '+', '?'):
+                    seq[-1] = (to_rule(seq[-1]) + c, False)
+                    i += 1
+                elif c == '{':
+                    curly_brackets = c
+                    i += 1
+                    while i < length and pattern[i] != '}':
+                        curly_brackets += pattern[i]
+                        i += 1
+                    assert i < length, f'Unbalanced curly brackets; start = {start}, i = {i}, pattern = {pattern}'
+                    curly_brackets += '}'
+                    i += 1
+                    nums = [s.strip() for s in curly_brackets[1:-1].split(',')]
+                    min_times = 0
+                    max_times = None
+                    try:
+                        if len(nums) == 1:
+                            min_times = int(nums[0])
+                            max_times = min_times
+                        else:
+                            assert len(nums) == 2
+                            min_times = int(nums[0]) if nums[0] else 0
+                            max_times = int(nums[1]) if nums[1] else None
+                    except ValueError:
+                        raise ValueError(f'Invalid quantifier {curly_brackets} in /{pattern}/')
+
+                    (sub, sub_is_literal) = seq[-1]
+
+                    if not sub_is_literal:
+                        id = sub_rule_ids.get(sub)
+                        if id is None:
+                            id = self._add_rule(f'{name}-{len(sub_rule_ids) + 1}', sub)
+                            sub_rule_ids[sub] = id
+                        sub = id
+
+                    seq[-1] = (_build_repetition(f'"{sub}"' if sub_is_literal else sub, min_times, max_times), False)
+                else:
+                    literal = ''
+                    while i < length:
+                        if pattern[i] == '\\' and i < length - 1:
+                            next = pattern[i + 1]
+                            if next in ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS:
+                                i += 1
+                                literal += pattern[i]
+                                i += 1
+                            else:
+                                literal += pattern[i:i+2]
+                                i += 2
+                        elif pattern[i] == '"' and not self._raw_pattern:
+                            literal += '\\"'
+                            i += 1
+                        elif pattern[i] not in NON_LITERAL_SET and \
+                                (i == length - 1 or literal == '' or pattern[i+1] == '.' or pattern[i+1] not in NON_LITERAL_SET):
+                            literal += pattern[i]
+                            i += 1
+                        else:
+                            break
+                    if literal:
+                        seq.append((literal, True))
+
+            return join_seq()
+
+        return self._add_rule(
+            name,
+            to_rule(transform()) if self._raw_pattern \
+                else "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space")
+
+
+    def _resolve_ref(self, ref):
+        ref_name = ref.split('/')[-1]
+        if ref_name not in self._rules and ref not in self._refs_being_resolved:
+            self._refs_being_resolved.add(ref)
+            resolved = self._refs[ref]
+            ref_name = self.visit(resolved, ref_name)
+            self._refs_being_resolved.remove(ref)
+        return ref_name
+
+    def _generate_constant_rule(self, value):
+        return self._format_literal(json.dumps(value))
+
+    def visit(self, schema, name):
+        schema_type = schema.get('type')
+        schema_format = schema.get('format')
+        rule_name = name + '-' if name in RESERVED_NAMES else name or 'root'
+
+        if (ref := schema.get('$ref')) is not None:
+            return self._add_rule(rule_name, self._resolve_ref(ref))
+
+        elif 'oneOf' in schema or 'anyOf' in schema:
+            return self._add_rule(rule_name, self._generate_union_rule(name, schema.get('oneOf') or schema['anyOf']))
+
+        elif isinstance(schema_type, list):
+            return self._add_rule(rule_name, self._generate_union_rule(name, [{**schema, 'type': t} for t in schema_type]))
+
+        elif 'const' in schema:
+            return self._add_rule(rule_name, self._generate_constant_rule(schema['const']) + ' space')
+
+        elif 'enum' in schema:
+            rule = '(' + ' | '.join((self._generate_constant_rule(v) for v in schema['enum'])) + ') space'
+            return self._add_rule(rule_name, rule)
+
+        elif schema_type in (None, 'object') and \
+             ('properties' in schema or \
+              ('additionalProperties' in schema and schema['additionalProperties'] is not True)):
+            required = set(schema.get('required', []))
+            properties = list(schema.get('properties', {}).items())
+            return self._add_rule(rule_name, self._build_object_rule(properties, required, name, schema.get('additionalProperties')))
+
+        elif schema_type in (None, 'object') and 'allOf' in schema:
+            required = set()
+            properties = []
+            hybrid_name = name
+            def add_component(comp_schema, is_required):
+                if (ref := comp_schema.get('$ref')) is not None:
+                    comp_schema = self._refs[ref]
+
+                if 'properties' in comp_schema:
+                    for prop_name, prop_schema in comp_schema['properties'].items():
+                        properties.append((prop_name, prop_schema))
+                        if is_required:
+                            required.add(prop_name)
+
+            for t in schema['allOf']:
+                if 'anyOf' in t:
+                    for tt in t['anyOf']:
+                        add_component(tt, is_required=False)
+                else:
+                    add_component(t, is_required=True)
+
+            return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=None))
+
+        elif schema_type in (None, 'array') and ('items' in schema or 'prefixItems' in schema):
+            items = schema.get('items') or schema['prefixItems']
+            if isinstance(items, list):
+                return self._add_rule(
+                    rule_name,
+                    '"[" space ' +
+                    ' "," space '.join(
+                        self.visit(item, f'{name}{"-" if name else ""}tuple-{i}')
+                        for i, item in enumerate(items)) +
+                    ' "]" space')
+            else:
+                item_rule_name = self.visit(items, f'{name}{"-" if name else ""}item')
+                min_items = schema.get("minItems", 0)
+                max_items = schema.get("maxItems")
+                return self._add_rule(rule_name, '"[" space ' + _build_repetition(item_rule_name, min_items, max_items, separator_rule='"," space') + ' "]" space')
+
+        elif schema_type in (None, 'string') and 'pattern' in schema:
+            return self._visit_pattern(schema['pattern'], rule_name)
+
+        elif schema_type in (None, 'string') and re.match(r'^uuid[1-5]?$', schema_format or ''):
+            return self._add_primitive(
+                'root' if rule_name == 'root' else schema_format,
+                PRIMITIVE_RULES['uuid']
+            )
+
+        elif schema_type in (None, 'string') and f'{schema_format}-string' in STRING_FORMAT_RULES:
+            prim_name = f'{schema_format}-string'
+            return self._add_rule(rule_name, self._add_primitive(prim_name, STRING_FORMAT_RULES[prim_name]))
+
+        elif schema_type == 'string' and ('minLength' in schema or 'maxLength' in schema):
+            char_rule = self._add_primitive('char', PRIMITIVE_RULES['char'])
+            min_len = schema.get('minLength', 0)
+            max_len = schema.get('maxLength')
+
+            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']))
+
+        else:
+            assert schema_type in PRIMITIVE_RULES, f'Unrecognized schema: {schema}'
+            # TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero
+            return self._add_primitive('root' if rule_name == 'root' else schema_type, PRIMITIVE_RULES[schema_type])
+
+    def _add_primitive(self, name: str, rule: BuiltinRule):
+        n = self._add_rule(name, rule.content)
+
+        for dep in rule.deps:
+            dep_rule = PRIMITIVE_RULES.get(dep) or STRING_FORMAT_RULES.get(dep)
+            assert dep_rule, f'Rule {dep} not known'
+            if dep not in self._rules:
+                self._add_primitive(dep, dep_rule)
+        return n
+
+    def _build_object_rule(self, properties: List[Tuple[str, Any]], required: Set[str], name: str, additional_properties: Optional[Union[bool, Any]]):
+        prop_order = self._prop_order
+        # sort by position in prop_order (if specified) then by original order
+        sorted_props = [kv[0] for _, kv in sorted(enumerate(properties), key=lambda ikv: (prop_order.get(ikv[1][0], len(prop_order)), ikv[0]))]
+
+        prop_kv_rule_names = {}
+        for prop_name, prop_schema in properties:
+            prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}')
+            prop_kv_rule_names[prop_name] = self._add_rule(
+                f'{name}{"-" if name else ""}{prop_name}-kv',
+                fr'{self._format_literal(json.dumps(prop_name))} space ":" space {prop_rule_name}'
+            )
+        required_props = [k for k in sorted_props if k in required]
+        optional_props = [k for k in sorted_props if k not in required]
+
+        if additional_properties is not None and additional_properties != False:
+            sub_name = f'{name}{"-" if name else ""}additional'
+            value_rule = self.visit(additional_properties, f'{sub_name}-value') if isinstance(additional_properties, dict) else \
+                self._add_primitive('value', PRIMITIVE_RULES['value'])
+            key_rule = self._add_primitive('string', PRIMITIVE_RULES['string']) if not sorted_props \
+                else self._add_rule(f'{sub_name}-k', self._not_strings(sorted_props))
+
+            prop_kv_rule_names["*"] = self._add_rule(
+                f'{sub_name}-kv',
+                f'{key_rule} ":" space {value_rule}'
+            )
+            optional_props.append("*")
+
+        rule = '"{" space '
+        rule += ' "," space '.join(prop_kv_rule_names[k] for k in required_props)
+
+        if optional_props:
+            rule += ' ('
+            if required_props:
+                rule += ' "," space ( '
+
+            def get_recursive_refs(ks, first_is_optional):
+                [k, *rest] = ks
+                kv_rule_name = prop_kv_rule_names[k]
+                comma_ref = f'( "," space {kv_rule_name} )'
+                if first_is_optional:
+                    res = comma_ref + ('*' if k == '*' else '?')
+                else:
+                    res = kv_rule_name + (' ' + comma_ref + "*" if k == '*' else '')
+                if len(rest) > 0:
+                    res += ' ' + self._add_rule(
+                        f'{name}{"-" if name else ""}{k}-rest',
+                        get_recursive_refs(rest, first_is_optional=True)
+                    )
+                return res
+
+            rule += ' | '.join(
+                get_recursive_refs(optional_props[i:], first_is_optional=False)
+                for i in range(len(optional_props))
+            )
+            if required_props:
+                rule += ' )'
+            rule += ' )?'
+
+        rule += ' "}" space'
+
+        return rule
+
+    def format_grammar(self):
+        return '\n'.join(
+            f'{name} ::= {rule}'
+            for name, rule in sorted(self._rules.items(), key=lambda kv: kv[0])
+        )
+
+
+def main(args_in = None):
+    parser = argparse.ArgumentParser(
+        description='''
+            Generates a grammar (suitable for use in ./llama-cli) that produces JSON conforming to a
+            given JSON schema. Only a subset of JSON schema features are supported; more may be
+            added in the future.
+        ''',
+    )
+    parser.add_argument(
+        '--prop-order',
+        default=[],
+        type=lambda s: s.split(','),
+        help='''
+            comma-separated property names defining the order of precedence for object properties;
+            properties not specified here are given lower precedence than those that are, and
+            are kept in their original order from the schema. Required properties are always
+            given precedence over optional properties.
+        '''
+    )
+    parser.add_argument(
+        '--allow-fetch',
+        action='store_true',
+        default=False,
+        help='Whether to allow fetching referenced schemas over HTTPS')
+    parser.add_argument(
+        '--dotall',
+        action='store_true',
+        default=False,
+        help='Whether to treat dot (".") as matching all chars including line breaks in regular expression patterns')
+    parser.add_argument(
+        '--raw-pattern',
+        action='store_true',
+        default=False,
+        help='Treats string patterns as raw patterns w/o quotes (or quote escapes)')
+
+    parser.add_argument('schema', help='file containing JSON schema ("-" for stdin)')
+    args = parser.parse_args(args_in)
+
+    if args.schema.startswith('https://'):
+        url = args.schema
+        import requests
+        schema = requests.get(url).json()
+    elif args.schema == '-':
+        url = 'stdin'
+        schema = json.load(sys.stdin)
+    else:
+        url = f'file://{args.schema}'
+        with open(args.schema) as f:
+            schema = json.load(f)
+    converter = SchemaConverter(
+        prop_order={name: idx for idx, name in enumerate(args.prop_order)},
+        allow_fetch=args.allow_fetch,
+        dotall=args.dotall,
+        raw_pattern=args.raw_pattern)
+    schema = converter.resolve_refs(schema, url)
+    converter.visit(schema, '')
+    print(converter.format_grammar())
+
+
+if __name__ == '__main__':
+    main()

examples/llama-bench/CMakeLists.txt

+set(TARGET llama-bench)
+add_executable(${TARGET} llama-bench.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/llama-bench/README.md

+# llama.cpp/examples/llama-bench
+
+Performance testing tool for llama.cpp.
+
+## Table of contents
+
+1. [Syntax](#syntax)
+2. [Examples](#examples)
+    1. [Text generation with different models](#text-generation-with-different-models)
+    2. [Prompt processing with different batch sizes](#prompt-processing-with-different-batch-sizes)
+    3. [Different numbers of threads](#different-numbers-of-threads)
+    4. [Different numbers of layers offloaded to the GPU](#different-numbers-of-layers-offloaded-to-the-gpu)
+3. [Output formats](#output-formats)
+    1. [Markdown](#markdown)
+    2. [CSV](#csv)
+    3. [JSON](#json)
+    4. [SQL](#sql)
+
+## Syntax
+
+```
+usage: ./llama-bench [options]
+
+options:
+  -h, --help
+  -m, --model <filename>              (default: models/7B/ggml-model-q4_0.gguf)
+  -p, --n-prompt <n>                  (default: 512)
+  -n, --n-gen <n>                     (default: 128)
+  -pg <pp,tg>                         (default: 512,128)
+  -b, --batch-size <n>                (default: 2048)
+  -ub, --ubatch-size <n>              (default: 512)
+  -ctk, --cache-type-k <t>            (default: f16)
+  -ctv, --cache-type-v <t>            (default: f16)
+  -t, --threads <n>                   (default: 16)
+  -ngl, --n-gpu-layers <n>            (default: 99)
+  -sm, --split-mode <none|layer|row>  (default: layer)
+  -mg, --main-gpu <i>                 (default: 0)
+  -nkvo, --no-kv-offload <0|1>        (default: 0)
+  -fa, --flash-attn <0|1>             (default: 0)
+  -mmp, --mmap <0|1>                  (default: 1)
+  --numa <distribute|isolate|numactl> (default: disabled)
+  -embd, --embeddings <0|1>           (default: 0)
+  -ts, --tensor-split <ts0/ts1/..>    (default: 0)
+  -r, --repetitions <n>               (default: 5)
+  -o, --output <csv|json|md|sql>      (default: md)
+  -v, --verbose                       (default: 0)
+
+Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.
+```
+
+llama-bench can perform three types of tests:
+
+- Prompt processing (pp): processing a prompt in batches (`-p`)
+- Text generation (tg): generating a sequence of tokens (`-n`)
+- Prompt processing + text generation (pg): processing a prompt followed by generating a sequence of tokens (`-pg`)
+
+With the exception of `-r`, `-o` and `-v`, all options can be specified multiple times to run multiple tests. Each pp and tg test is run with all combinations of the specified options. To specify multiple values for an option, the values can be separated by commas (e.g. `-n 16,32`), or the option can be specified multiple times (e.g. `-n 16 -n 32`).
+
+Each test is repeated the number of times given by `-r`, and the results are averaged. The results are given in average tokens per second (t/s) and standard deviation. Some output formats (e.g. json) also include the individual results of each repetition.
+
+For a description of the other options, see the [main example](../main/README.md).
+
+Note:
+
+- When using SYCL backend, there would be hang issue in some cases. Please set `--mmp 0`.
+
+## Examples
+
+### Text generation with different models
+
+```sh
+$ ./llama-bench -m models/7B/ggml-model-q4_0.gguf -m models/13B/ggml-model-q4_0.gguf -p 0 -n 128,256,512
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 128     |    132.19 ± 0.55 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 256     |    129.37 ± 0.54 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 512     |    123.83 ± 0.25 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 128     |     82.17 ± 0.31 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 256     |     80.74 ± 0.23 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 512     |     78.08 ± 0.07 |
+
+### Prompt processing with different batch sizes
+
+```sh
+$ ./llama-bench -n 0 -p 1024 -b 128,256,512,1024
+```
+
+| model                          |       size |     params | backend    | ngl |    n_batch | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        128 | pp 1024    |   1436.51 ± 3.66 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        256 | pp 1024    |  1932.43 ± 23.48 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        512 | pp 1024    |  2254.45 ± 15.59 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |       1024 | pp 1024    |  2498.61 ± 13.58 |
+
+### Different numbers of threads
+
+```sh
+$ ./llama-bench -n 0 -n 16 -p 64 -t 1,2,4,8,16,32
+```
+
+| model                          |       size |     params | backend    |    threads | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ---------: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          1 | pp 64      |      6.17 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          1 | tg 16      |      4.05 ± 0.02 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          2 | pp 64      |     12.31 ± 0.13 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          2 | tg 16      |      7.80 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          4 | pp 64      |     23.18 ± 0.06 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          4 | tg 16      |     12.22 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          8 | pp 64      |     32.29 ± 1.21 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          8 | tg 16      |     16.71 ± 0.66 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         16 | pp 64      |     33.52 ± 0.03 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         16 | tg 16      |     15.32 ± 0.05 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         32 | pp 64      |     59.00 ± 1.11 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         32 | tg 16      |     16.41 ± 0.79 ||
+
+### Different numbers of layers offloaded to the GPU
+
+```sh
+$ ./llama-bench -ngl 10,20,30,31,32,33,34,35
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  10 | pp 512     |    373.36 ± 2.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  10 | tg 128     |     13.45 ± 0.93 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  20 | pp 512     |    472.65 ± 1.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  20 | tg 128     |     21.36 ± 1.94 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  30 | pp 512     |   631.87 ± 11.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  30 | tg 128     |     40.04 ± 1.82 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  31 | pp 512     |    657.89 ± 5.08 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  31 | tg 128     |     48.19 ± 0.81 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  32 | pp 512     |    688.26 ± 3.29 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  32 | tg 128     |     54.78 ± 0.65 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  33 | pp 512     |    704.27 ± 2.24 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  33 | tg 128     |     60.62 ± 1.76 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  34 | pp 512     |    881.34 ± 5.40 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  34 | tg 128     |     71.76 ± 0.23 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | pp 512     |   2400.01 ± 7.72 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | tg 128     |    131.66 ± 0.49 |
+
+## Output formats
+
+By default, llama-bench outputs the results in markdown format. The results can be output in other formats by using the `-o` option.
+
+### Markdown
+
+```sh
+$ ./llama-bench -o md
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | pp 512     |  2368.80 ± 93.24 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 128     |    131.42 ± 0.59 |
+
+### CSV
+
+```sh
+$ ./llama-bench -o csv
+```
+
+```csv
+build_commit,build_number,cuda,metal,gpu_blas,blas,cpu_info,gpu_info,model_filename,model_type,model_size,model_n_params,n_batch,n_threads,f16_kv,n_gpu_layers,main_gpu,mul_mat_q,tensor_split,n_prompt,n_gen,test_time,avg_ns,stddev_ns,avg_ts,stddev_ts
+"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","512","0","2023-09-23T12:09:01Z","212155977","732372","2413.341687","8.305961"
+"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","0","128","2023-09-23T12:09:02Z","969320879","2728399","132.052051","0.371342"
+```
+
+### JSON
+
+```sh
+$ ./llama-bench -o json
+```
+
+```json
+[
+  {
+    "build_commit": "3469684",
+    "build_number": 1275,
+    "cuda": true,
+    "metal": false,
+    "gpu_blas": true,
+    "blas": true,
+    "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K",
+    "gpu_info": "NVIDIA GeForce RTX 3090 Ti",
+    "model_filename": "models/7B/ggml-model-q4_0.gguf",
+    "model_type": "llama 7B mostly Q4_0",
+    "model_size": 3825065984,
+    "model_n_params": 6738415616,
+    "n_batch": 512,
+    "n_threads": 16,
+    "f16_kv": true,
+    "n_gpu_layers": 99,
+    "main_gpu": 0,
+    "mul_mat_q": true,
+    "tensor_split": "0.00",
+    "n_prompt": 512,
+    "n_gen": 0,
+    "test_time": "2023-09-23T12:09:57Z",
+    "avg_ns": 212365953,
+    "stddev_ns": 985423,
+    "avg_ts": 2410.974041,
+    "stddev_ts": 11.163766,
+    "samples_ns": [ 213837238, 211635853, 212328053, 211329715, 212698907 ],
+    "samples_ts": [ 2394.34, 2419.25, 2411.36, 2422.75, 2407.16 ]
+  },
+  {
+    "build_commit": "3469684",
+    "build_number": 1275,
+    "cuda": true,
+    "metal": false,
+    "gpu_blas": true,
+    "blas": true,
+    "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K",
+    "gpu_info": "NVIDIA GeForce RTX 3090 Ti",
+    "model_filename": "models/7B/ggml-model-q4_0.gguf",
+    "model_type": "llama 7B mostly Q4_0",
+    "model_size": 3825065984,
+    "model_n_params": 6738415616,
+    "n_batch": 512,
+    "n_threads": 16,
+    "f16_kv": true,
+    "n_gpu_layers": 99,
+    "main_gpu": 0,
+    "mul_mat_q": true,
+    "tensor_split": "0.00",
+    "n_prompt": 0,
+    "n_gen": 128,
+    "test_time": "2023-09-23T12:09:59Z",
+    "avg_ns": 977425219,
+    "stddev_ns": 9268593,
+    "avg_ts": 130.965708,
+    "stddev_ts": 1.238924,
+    "samples_ns": [ 984472709, 974901233, 989474741, 970729355, 967548060 ],
+    "samples_ts": [ 130.019, 131.295, 129.362, 131.86, 132.293 ]
+  }
+]
+```
+
+### SQL
+
+SQL output is suitable for importing into a SQLite database. The output can be piped into the `sqlite3` command line tool to add the results to a database.
+
+```sh
+$ ./llama-bench -o sql
+```
+
+```sql
+CREATE TABLE IF NOT EXISTS test (
+  build_commit TEXT,
+  build_number INTEGER,
+  cuda INTEGER,
+  metal INTEGER,
+  gpu_blas INTEGER,
+  blas INTEGER,
+  cpu_info TEXT,
+  gpu_info TEXT,
+  model_filename TEXT,
+  model_type TEXT,
+  model_size INTEGER,
+  model_n_params INTEGER,
+  n_batch INTEGER,
+  n_threads INTEGER,
+  f16_kv INTEGER,
+  n_gpu_layers INTEGER,
+  main_gpu INTEGER,
+  mul_mat_q INTEGER,
+  tensor_split TEXT,
+  n_prompt INTEGER,
+  n_gen INTEGER,
+  test_time TEXT,
+  avg_ns INTEGER,
+  stddev_ns INTEGER,
+  avg_ts REAL,
+  stddev_ts REAL
+);
+
+INSERT INTO test (build_commit, build_number, cuda, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '512', '0', '2023-09-23T12:10:30Z', '212693772', '743623', '2407.240204', '8.409634');
+INSERT INTO test (build_commit, build_number, cuda, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '0', '128', '2023-09-23T12:10:31Z', '977925003', '4037361', '130.891159', '0.537692');
+```

examples/llama-bench/llama-bench.cpp

+#include <algorithm>
+#include <array>
+#include <cassert>
+#include <chrono>
+#include <cinttypes>
+#include <clocale>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <cstdlib>
+#include <iterator>
+#include <map>
+#include <numeric>
+#include <regex>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#include "ggml.h"
+#include "llama.h"
+#include "common.h"
+#include "ggml-cuda.h"
+#include "ggml-sycl.h"
+
+#ifdef GGML_USE_CANN
+#include "ggml-cann.h"
+#endif
+
+// utils
+static uint64_t get_time_ns() {
+    using clock = std::chrono::high_resolution_clock;
+    return std::chrono::nanoseconds(clock::now().time_since_epoch()).count();
+}
+
+template<class T>
+static std::string join(const std::vector<T> & values, const std::string & delim) {
+    std::ostringstream str;
+    for (size_t i = 0; i < values.size(); i++) {
+        str << values[i];
+        if (i < values.size() - 1) {
+            str << delim;
+        }
+    }
+    return str.str();
+}
+
+template<typename T, typename F>
+static std::vector<std::string> transform_to_str(const std::vector<T> & values, F f) {
+    std::vector<std::string> str_values;
+    std::transform(values.begin(), values.end(), std::back_inserter(str_values), f);
+    return str_values;
+}
+
+template<typename T>
+static T avg(const std::vector<T> & v) {
+    if (v.empty()) {
+        return 0;
+    }
+    T sum = std::accumulate(v.begin(), v.end(), T(0));
+    return sum / (T)v.size();
+}
+
+template<typename T>
+static T stdev(const std::vector<T> & v) {
+    if (v.size() <= 1) {
+        return 0;
+    }
+    T mean = avg(v);
+    T sq_sum = std::inner_product(v.begin(), v.end(), v.begin(), T(0));
+    T stdev = std::sqrt(sq_sum / (T)(v.size() - 1) - mean * mean * (T)v.size() / (T)(v.size() - 1));
+    return stdev;
+}
+
+static std::string get_cpu_info() {
+    std::string id;
+#ifdef __linux__
+    FILE * f = fopen("/proc/cpuinfo", "r");
+    if (f) {
+        char buf[1024];
+        while (fgets(buf, sizeof(buf), f)) {
+            if (strncmp(buf, "model name", 10) == 0) {
+                char * p = strchr(buf, ':');
+                if (p) {
+                    p++;
+                    while (std::isspace(*p)) {
+                        p++;
+                    }
+                    while (std::isspace(p[strlen(p) - 1])) {
+                        p[strlen(p) - 1] = '\0';
+                    }
+                    id = p;
+                    break;
+                }
+            }
+        }
+        fclose(f);
+    }
+#endif
+    // TODO: other platforms
+    return id;
+}
+
+static std::string get_gpu_info() {
+    std::string id;
+#ifdef GGML_USE_CUDA
+    int count = ggml_backend_cuda_get_device_count();
+    for (int i = 0; i < count; i++) {
+        char buf[128];
+        ggml_backend_cuda_get_device_description(i, buf, sizeof(buf));
+        id += buf;
+        if (i < count - 1) {
+            id += "/";
+        }
+    }
+#endif
+#ifdef GGML_USE_SYCL
+    int count = ggml_backend_sycl_get_device_count();
+    for (int i = 0; i < count; i++) {
+        char buf[128];
+        ggml_sycl_get_device_description(i, buf, sizeof(buf));
+        id += buf;
+        if (i < count - 1) {
+            id += "/";
+        }
+    }
+#endif
+#ifdef GGML_USE_CANN
+    uint32_t count = ggml_backend_cann_get_device_count();
+    for (uint32_t i = 0; i < count; i++) {
+        char buf[128];
+        ggml_backend_cann_get_device_description(i, buf, sizeof(buf));
+        id += buf;
+        if (i < count - 1) {
+            id += "/";
+        }
+    }
+#endif
+    // TODO: other backends
+    return id;
+}
+
+// command line params
+enum output_formats {NONE, CSV, JSON, MARKDOWN, SQL};
+
+static const char * output_format_str(output_formats format) {
+    switch (format) {
+        case NONE:     return "none";
+        case CSV:      return "csv";
+        case JSON:     return "json";
+        case MARKDOWN: return "md";
+        case SQL:      return "sql";
+        default: GGML_ABORT("invalid output format");
+    }
+}
+
+static bool output_format_from_str(const std::string & s, output_formats & format) {
+    if (s == "none") {
+        format = NONE;
+    } else if (s == "csv") {
+        format = CSV;
+    } else if (s == "json") {
+        format = JSON;
+    } else if (s == "md") {
+        format = MARKDOWN;
+    } else if (s == "sql") {
+        format = SQL;
+    } else {
+        return false;
+    }
+    return true;
+}
+
+static const char * split_mode_str(llama_split_mode mode) {
+    switch (mode) {
+        case LLAMA_SPLIT_MODE_NONE:  return "none";
+        case LLAMA_SPLIT_MODE_LAYER: return "layer";
+        case LLAMA_SPLIT_MODE_ROW:   return "row";
+        default: GGML_ABORT("invalid split mode");
+    }
+}
+
+static std::string pair_str(const std::pair<int, int> & p) {
+    static char buf[32];
+    snprintf(buf, sizeof(buf), "%d,%d", p.first, p.second);
+    return buf;
+}
+
+struct cmd_params {
+    std::vector<std::string> model;
+    std::vector<int> n_prompt;
+    std::vector<int> n_gen;
+    std::vector<std::pair<int, int>> n_pg;
+    std::vector<int> n_batch;
+    std::vector<int> n_ubatch;
+    std::vector<ggml_type> type_k;
+    std::vector<ggml_type> type_v;
+    std::vector<int> n_threads;
+    std::vector<int> n_gpu_layers;
+    std::vector<std::string> rpc_servers;
+    std::vector<llama_split_mode> split_mode;
+    std::vector<int> main_gpu;
+    std::vector<bool> no_kv_offload;
+    std::vector<bool> flash_attn;
+    std::vector<std::vector<float>> tensor_split;
+    std::vector<bool> use_mmap;
+    std::vector<bool> embeddings;
+    ggml_numa_strategy numa;
+    int reps;
+    bool verbose;
+    output_formats output_format;
+    output_formats output_format_stderr;
+};
+
+static const cmd_params cmd_params_defaults = {
+    /* model                */ {"models/7B/ggml-model-q4_0.gguf"},
+    /* n_prompt             */ {512},
+    /* n_gen                */ {128},
+    /* n_pg                 */ {},
+    /* n_batch              */ {2048},
+    /* n_ubatch             */ {512},
+    /* type_k               */ {GGML_TYPE_F16},
+    /* type_v               */ {GGML_TYPE_F16},
+    /* n_threads            */ {cpu_get_num_math()},
+    /* n_gpu_layers         */ {99},
+    /* rpc_servers          */ {""},
+    /* split_mode           */ {LLAMA_SPLIT_MODE_LAYER},
+    /* main_gpu             */ {0},
+    /* no_kv_offload        */ {false},
+    /* flash_attn           */ {false},
+    /* tensor_split         */ {std::vector<float>(llama_max_devices(), 0.0f)},
+    /* use_mmap             */ {true},
+    /* embeddings           */ {false},
+    /* numa                 */ GGML_NUMA_STRATEGY_DISABLED,
+    /* reps                 */ 5,
+    /* verbose              */ false,
+    /* output_format        */ MARKDOWN,
+    /* output_format_stderr */ NONE,
+};
+
+static void print_usage(int /* argc */, char ** argv) {
+    printf("usage: %s [options]\n", argv[0]);
+    printf("\n");
+    printf("options:\n");
+    printf("  -h, --help\n");
+    printf("  -m, --model <filename>              (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
+    printf("  -p, --n-prompt <n>                  (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
+    printf("  -n, --n-gen <n>                     (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
+    printf("  -pg <pp,tg>                         (default: %s)\n", join(transform_to_str(cmd_params_defaults.n_pg, pair_str), ",").c_str());
+    printf("  -b, --batch-size <n>                (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
+    printf("  -ub, --ubatch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_ubatch, ",").c_str());
+    printf("  -ctk, --cache-type-k <t>            (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
+    printf("  -ctv, --cache-type-v <t>            (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
+    printf("  -t, --threads <n>                   (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
+    printf("  -ngl, --n-gpu-layers <n>            (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
+    printf("  -rpc, --rpc <rpc_servers>           (default: %s)\n", join(cmd_params_defaults.rpc_servers, ",").c_str());
+    printf("  -sm, --split-mode <none|layer|row>  (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
+    printf("  -mg, --main-gpu <i>                 (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
+    printf("  -nkvo, --no-kv-offload <0|1>        (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
+    printf("  -fa, --flash-attn <0|1>             (default: %s)\n", join(cmd_params_defaults.flash_attn, ",").c_str());
+    printf("  -mmp, --mmap <0|1>                  (default: %s)\n", join(cmd_params_defaults.use_mmap, ",").c_str());
+    printf("  --numa <distribute|isolate|numactl> (default: disabled)\n");
+    printf("  -embd, --embeddings <0|1>           (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str());
+    printf("  -ts, --tensor-split <ts0/ts1/..>    (default: 0)\n");
+    printf("  -r, --repetitions <n>               (default: %d)\n", cmd_params_defaults.reps);
+    printf("  -o, --output <csv|json|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
+    printf("  -oe, --output-err <csv|json|md|sql> (default: %s)\n", output_format_str(cmd_params_defaults.output_format_stderr));
+    printf("  -v, --verbose                       (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
+    printf("\n");
+    printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
+}
+
+static ggml_type ggml_type_from_name(const std::string & s) {
+    if (s == "f16") {
+        return GGML_TYPE_F16;
+    }
+    if (s == "q8_0") {
+        return GGML_TYPE_Q8_0;
+    }
+    if (s == "q4_0") {
+        return GGML_TYPE_Q4_0;
+    }
+    if (s == "q4_1") {
+        return GGML_TYPE_Q4_1;
+    }
+    if (s == "q5_0") {
+        return GGML_TYPE_Q5_0;
+    }
+    if (s == "q5_1") {
+        return GGML_TYPE_Q5_1;
+    }
+    if (s == "iq4_nl") {
+        return GGML_TYPE_IQ4_NL;
+    }
+
+    return GGML_TYPE_COUNT;
+}
+
+
+static cmd_params parse_cmd_params(int argc, char ** argv) {
+    cmd_params params;
+    std::string arg;
+    bool invalid_param = false;
+    const std::string arg_prefix = "--";
+    const char split_delim = ',';
+
+    params.verbose = cmd_params_defaults.verbose;
+    params.output_format = cmd_params_defaults.output_format;
+    params.output_format_stderr = cmd_params_defaults.output_format_stderr;
+    params.reps = cmd_params_defaults.reps;
+    params.numa = cmd_params_defaults.numa;
+
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+            std::replace(arg.begin(), arg.end(), '_', '-');
+        }
+
+        if (arg == "-h" || arg == "--help") {
+            print_usage(argc, argv);
+            exit(0);
+        } else if (arg == "-m" || arg == "--model") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<std::string>(argv[i], split_delim);
+            params.model.insert(params.model.end(), p.begin(), p.end());
+        } else if (arg == "-p" || arg == "--n-prompt") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<int>(argv[i], split_delim);
+            params.n_prompt.insert(params.n_prompt.end(), p.begin(), p.end());
+        } else if (arg == "-n" || arg == "--n-gen") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<int>(argv[i], split_delim);
+            params.n_gen.insert(params.n_gen.end(), p.begin(), p.end());
+        } else if (arg == "-pg") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<std::string>(argv[i], ',');
+            if (p.size() != 2) {
+                invalid_param = true;
+                break;
+            }
+            params.n_pg.push_back({std::stoi(p[0]), std::stoi(p[1])});
+        } else if (arg == "-b" || arg == "--batch-size") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<int>(argv[i], split_delim);
+            params.n_batch.insert(params.n_batch.end(), p.begin(), p.end());
+        } else if (arg == "-ub" || arg == "--ubatch-size") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<int>(argv[i], split_delim);
+            params.n_ubatch.insert(params.n_ubatch.end(), p.begin(), p.end());
+        } else if (arg == "-ctk" || arg == "--cache-type-k") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<std::string>(argv[i], split_delim);
+            std::vector<ggml_type> types;
+            for (const auto & t : p) {
+                ggml_type gt = ggml_type_from_name(t);
+                if (gt == GGML_TYPE_COUNT) {
+                    invalid_param = true;
+                    break;
+                }
+                types.push_back(gt);
+            }
+            params.type_k.insert(params.type_k.end(), types.begin(), types.end());
+        } else if (arg == "-ctv" || arg == "--cache-type-v") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<std::string>(argv[i], split_delim);
+            std::vector<ggml_type> types;
+            for (const auto & t : p) {
+                ggml_type gt = ggml_type_from_name(t);
+                if (gt == GGML_TYPE_COUNT) {
+                    invalid_param = true;
+                    break;
+                }
+                types.push_back(gt);
+            }
+            params.type_v.insert(params.type_v.end(), types.begin(), types.end());
+        } else if (arg == "-t" || arg == "--threads") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<int>(argv[i], split_delim);
+            params.n_threads.insert(params.n_threads.end(), p.begin(), p.end());
+        } else if (arg == "-ngl" || arg == "--n-gpu-layers") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<int>(argv[i], split_delim);
+            params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end());
+        } else if (arg == "-rpc" || arg == "--rpc") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.rpc_servers.push_back(argv[i]);
+        } else if (arg == "-sm" || arg == "--split-mode") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<std::string>(argv[i], split_delim);
+            std::vector<llama_split_mode> modes;
+            for (const auto & m : p) {
+                llama_split_mode mode;
+                if (m == "none") {
+                    mode = LLAMA_SPLIT_MODE_NONE;
+                } else if (m == "layer") {
+                    mode = LLAMA_SPLIT_MODE_LAYER;
+                } else if (m == "row") {
+                    mode = LLAMA_SPLIT_MODE_ROW;
+                } else {
+                    invalid_param = true;
+                    break;
+                }
+                modes.push_back(mode);
+            }
+            params.split_mode.insert(params.split_mode.end(), modes.begin(), modes.end());
+        } else if (arg == "-mg" || arg == "--main-gpu") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.main_gpu = string_split<int>(argv[i], split_delim);
+        } else if (arg == "-nkvo" || arg == "--no-kv-offload") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<bool>(argv[i], split_delim);
+            params.no_kv_offload.insert(params.no_kv_offload.end(), p.begin(), p.end());
+        } else if (arg == "--numa") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            } else {
+                std::string value(argv[i]);
+                /**/ if (value == "distribute" || value == "" ) { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; }
+                else if (value == "isolate")                    { params.numa = GGML_NUMA_STRATEGY_ISOLATE; }
+                else if (value == "numactl")                    { params.numa = GGML_NUMA_STRATEGY_NUMACTL; }
+                else { invalid_param = true; break; }
+            }
+        } else if (arg == "-fa" || arg == "--flash-attn") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<bool>(argv[i], split_delim);
+            params.flash_attn.insert(params.flash_attn.end(), p.begin(), p.end());
+        } else if (arg == "-mmp" || arg == "--mmap") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<bool>(argv[i], split_delim);
+            params.use_mmap.insert(params.use_mmap.end(), p.begin(), p.end());
+        } else if (arg == "-embd" || arg == "--embeddings") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<bool>(argv[i], split_delim);
+            params.embeddings.insert(params.embeddings.end(), p.begin(), p.end());
+        } else if (arg == "-ts" || arg == "--tensor-split") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            for (auto ts : string_split<std::string>(argv[i], split_delim)) {
+                // split string by ; and /
+                const std::regex regex{R"([;/]+)"};
+                std::sregex_token_iterator it{ts.begin(), ts.end(), regex, -1};
+                std::vector<std::string> split_arg{it, {}};
+                GGML_ASSERT(split_arg.size() <= llama_max_devices());
+
+                std::vector<float> tensor_split(llama_max_devices());
+                for (size_t i = 0; i < llama_max_devices(); ++i) {
+                    if (i < split_arg.size()) {
+                        tensor_split[i] = std::stof(split_arg[i]);
+                    } else {
+                        tensor_split[i] = 0.0f;
+                    }
+                }
+                params.tensor_split.push_back(tensor_split);
+            }
+        } else if (arg == "-r" || arg == "--repetitions") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.reps = std::stoi(argv[i]);
+        } else if (arg == "-o" || arg == "--output") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            invalid_param = !output_format_from_str(argv[i], params.output_format);
+        } else if (arg == "-oe" || arg == "--output-err") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            invalid_param = !output_format_from_str(argv[i], params.output_format_stderr);
+        } else if (arg == "-v" || arg == "--verbose") {
+            params.verbose = true;
+        } else {
+            invalid_param = true;
+            break;
+        }
+    }
+    if (invalid_param) {
+        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+        print_usage(argc, argv);
+        exit(1);
+    }
+
+    // set defaults
+    if (params.model.empty())        { params.model = cmd_params_defaults.model; }
+    if (params.n_prompt.empty())     { params.n_prompt = cmd_params_defaults.n_prompt; }
+    if (params.n_gen.empty())        { params.n_gen = cmd_params_defaults.n_gen; }
+    if (params.n_pg.empty())         { params.n_pg = cmd_params_defaults.n_pg; }
+    if (params.n_batch.empty())      { params.n_batch = cmd_params_defaults.n_batch; }
+    if (params.n_ubatch.empty())     { params.n_ubatch = cmd_params_defaults.n_ubatch; }
+    if (params.type_k.empty())       { params.type_k = cmd_params_defaults.type_k; }
+    if (params.type_v.empty())       { params.type_v = cmd_params_defaults.type_v; }
+    if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
+    if (params.rpc_servers.empty())  { params.rpc_servers = cmd_params_defaults.rpc_servers; }
+    if (params.split_mode.empty())   { params.split_mode = cmd_params_defaults.split_mode; }
+    if (params.main_gpu.empty())     { params.main_gpu = cmd_params_defaults.main_gpu; }
+    if (params.no_kv_offload.empty()){ params.no_kv_offload = cmd_params_defaults.no_kv_offload; }
+    if (params.flash_attn.empty())   { params.flash_attn = cmd_params_defaults.flash_attn; }
+    if (params.tensor_split.empty()) { params.tensor_split = cmd_params_defaults.tensor_split; }
+    if (params.use_mmap.empty())     { params.use_mmap = cmd_params_defaults.use_mmap; }
+    if (params.embeddings.empty())   { params.embeddings = cmd_params_defaults.embeddings; }
+    if (params.n_threads.empty())    { params.n_threads = cmd_params_defaults.n_threads; }
+
+    return params;
+}
+
+struct cmd_params_instance {
+    std::string model;
+    int n_prompt;
+    int n_gen;
+    int n_batch;
+    int n_ubatch;
+    ggml_type type_k;
+    ggml_type type_v;
+    int n_threads;
+    int n_gpu_layers;
+    std::string rpc_servers;
+    llama_split_mode split_mode;
+    int main_gpu;
+    bool no_kv_offload;
+    bool flash_attn;
+    std::vector<float> tensor_split;
+    bool use_mmap;
+    bool embeddings;
+
+    llama_model_params to_llama_mparams() const {
+        llama_model_params mparams = llama_model_default_params();
+
+        mparams.n_gpu_layers = n_gpu_layers;
+        if (!rpc_servers.empty()) {
+            mparams.rpc_servers = rpc_servers.c_str();
+        }
+        mparams.split_mode = split_mode;
+        mparams.main_gpu = main_gpu;
+        mparams.tensor_split = tensor_split.data();
+        mparams.use_mmap = use_mmap;
+
+        return mparams;
+    }
+
+    bool equal_mparams(const cmd_params_instance & other) const {
+        return model == other.model &&
+               n_gpu_layers == other.n_gpu_layers &&
+               rpc_servers == other.rpc_servers &&
+               split_mode == other.split_mode &&
+               main_gpu == other.main_gpu &&
+               use_mmap == other.use_mmap &&
+               tensor_split == other.tensor_split;
+    }
+
+    llama_context_params to_llama_cparams() const {
+        llama_context_params cparams = llama_context_default_params();
+
+        cparams.n_ctx = n_prompt + n_gen;
+        cparams.n_batch = n_batch;
+        cparams.n_ubatch = n_ubatch;
+        cparams.type_k = type_k;
+        cparams.type_v = type_v;
+        cparams.offload_kqv = !no_kv_offload;
+        cparams.flash_attn = flash_attn;
+        cparams.embeddings = embeddings;
+
+        return cparams;
+    }
+};
+
+static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_params & params) {
+    std::vector<cmd_params_instance> instances;
+
+    // this ordering minimizes the number of times that each model needs to be reloaded
+    for (const auto & m : params.model)
+    for (const auto & nl : params.n_gpu_layers)
+    for (const auto & rpc : params.rpc_servers)
+    for (const auto & sm : params.split_mode)
+    for (const auto & mg : params.main_gpu)
+    for (const auto & ts : params.tensor_split)
+    for (const auto & mmp : params.use_mmap)
+    for (const auto & embd : params.embeddings)
+    for (const auto & nb : params.n_batch)
+    for (const auto & nub : params.n_ubatch)
+    for (const auto & tk : params.type_k)
+    for (const auto & tv : params.type_v)
+    for (const auto & nkvo : params.no_kv_offload)
+    for (const auto & fa : params.flash_attn)
+    for (const auto & nt : params.n_threads) {
+        for (const auto & n_prompt : params.n_prompt) {
+            if (n_prompt == 0) {
+                continue;
+            }
+            cmd_params_instance instance = {
+                /* .model        = */ m,
+                /* .n_prompt     = */ n_prompt,
+                /* .n_gen        = */ 0,
+                /* .n_batch      = */ nb,
+                /* .n_ubatch     = */ nub,
+                /* .type_k       = */ tk,
+                /* .type_v       = */ tv,
+                /* .n_threads    = */ nt,
+                /* .n_gpu_layers = */ nl,
+                /* .rpc_servers  = */ rpc,
+                /* .split_mode   = */ sm,
+                /* .main_gpu     = */ mg,
+                /* .no_kv_offload= */ nkvo,
+                /* .flash_attn   = */ fa,
+                /* .tensor_split = */ ts,
+                /* .use_mmap     = */ mmp,
+                /* .embeddings   = */ embd,
+            };
+            instances.push_back(instance);
+        }
+
+        for (const auto & n_gen : params.n_gen) {
+            if (n_gen == 0) {
+                continue;
+            }
+            cmd_params_instance instance = {
+                /* .model        = */ m,
+                /* .n_prompt     = */ 0,
+                /* .n_gen        = */ n_gen,
+                /* .n_batch      = */ nb,
+                /* .n_ubatch     = */ nub,
+                /* .type_k       = */ tk,
+                /* .type_v       = */ tv,
+                /* .n_threads    = */ nt,
+                /* .n_gpu_layers = */ nl,
+                /* .rpc_servers  = */ rpc,
+                /* .split_mode   = */ sm,
+                /* .main_gpu     = */ mg,
+                /* .no_kv_offload= */ nkvo,
+                /* .flash_attn   = */ fa,
+                /* .tensor_split = */ ts,
+                /* .use_mmap     = */ mmp,
+                /* .embeddings   = */ embd,
+            };
+            instances.push_back(instance);
+        }
+
+        for (const auto & n_pg : params.n_pg) {
+            if (n_pg.first == 0 && n_pg.second == 0) {
+                continue;
+            }
+            cmd_params_instance instance = {
+                /* .model        = */ m,
+                /* .n_prompt     = */ n_pg.first,
+                /* .n_gen        = */ n_pg.second,
+                /* .n_batch      = */ nb,
+                /* .n_ubatch     = */ nub,
+                /* .type_k       = */ tk,
+                /* .type_v       = */ tv,
+                /* .n_threads    = */ nt,
+                /* .n_gpu_layers = */ nl,
+                /* .rpc_servers  = */ rpc,
+                /* .split_mode   = */ sm,
+                /* .main_gpu     = */ mg,
+                /* .no_kv_offload= */ nkvo,
+                /* .flash_attn   = */ fa,
+                /* .tensor_split = */ ts,
+                /* .use_mmap     = */ mmp,
+                /* .embeddings   = */ embd,
+            };
+            instances.push_back(instance);
+        }
+    }
+
+    return instances;
+}
+
+struct test {
+    static const std::string build_commit;
+    static const int build_number;
+    static const bool cuda;
+    static const bool vulkan;
+    static const bool kompute;
+    static const bool metal;
+    static const bool sycl;
+    static const bool gpu_blas;
+    static const bool blas;
+    static const std::string cpu_info;
+    static const std::string gpu_info;
+    std::string model_filename;
+    std::string model_type;
+    uint64_t model_size;
+    uint64_t model_n_params;
+    int n_batch;
+    int n_ubatch;
+    int n_threads;
+    bool has_rpc;
+    ggml_type type_k;
+    ggml_type type_v;
+    int n_gpu_layers;
+    llama_split_mode split_mode;
+    int main_gpu;
+    bool no_kv_offload;
+    bool flash_attn;
+    std::vector<float> tensor_split;
+    bool use_mmap;
+    bool embeddings;
+    int n_prompt;
+    int n_gen;
+    std::string test_time;
+    std::vector<uint64_t> samples_ns;
+
+    test(const cmd_params_instance & inst, const llama_model * lmodel, const llama_context * ctx) {
+        model_filename = inst.model;
+        char buf[128];
+        llama_model_desc(lmodel, buf, sizeof(buf));
+        model_type = buf;
+        model_size = llama_model_size(lmodel);
+        model_n_params = llama_model_n_params(lmodel);
+        n_batch = inst.n_batch;
+        n_ubatch = inst.n_ubatch;
+        n_threads = inst.n_threads;
+        has_rpc = !inst.rpc_servers.empty();
+        type_k = inst.type_k;
+        type_v = inst.type_v;
+        n_gpu_layers = inst.n_gpu_layers;
+        split_mode = inst.split_mode;
+        main_gpu = inst.main_gpu;
+        no_kv_offload = inst.no_kv_offload;
+        flash_attn = inst.flash_attn;
+        tensor_split = inst.tensor_split;
+        use_mmap = inst.use_mmap;
+        embeddings = inst.embeddings;
+        n_prompt = inst.n_prompt;
+        n_gen = inst.n_gen;
+        // RFC 3339 date-time format
+        time_t t = time(NULL);
+        std::strftime(buf, sizeof(buf), "%FT%TZ", gmtime(&t));
+        test_time = buf;
+
+        (void) ctx;
+    }
+
+    uint64_t avg_ns() const {
+        return ::avg(samples_ns);
+    }
+
+    uint64_t stdev_ns() const {
+        return ::stdev(samples_ns);
+    }
+
+    std::vector<double> get_ts() const {
+        int n_tokens = n_prompt + n_gen;
+        std::vector<double> ts;
+        std::transform(samples_ns.begin(), samples_ns.end(), std::back_inserter(ts), [n_tokens](uint64_t t) { return 1e9 * n_tokens / t; });
+        return ts;
+    }
+
+    double avg_ts() const {
+        return ::avg(get_ts());
+    }
+
+    double stdev_ts() const {
+        return ::stdev(get_ts());
+    }
+
+    static std::string get_backend() {
+        if (cuda) {
+            return GGML_CUDA_NAME;
+        }
+        if (vulkan) {
+            return "Vulkan";
+        }
+        if (kompute) {
+            return "Kompute";
+        }
+        if (metal) {
+            return "Metal";
+        }
+        if (sycl) {
+            return GGML_SYCL_NAME;
+        }
+        if (gpu_blas) {
+            return "GPU BLAS";
+        }
+        if (blas) {
+            return "BLAS";
+        }
+
+        return "CPU";
+    }
+
+    static const std::vector<std::string> & get_fields() {
+        static const std::vector<std::string> fields = {
+            "build_commit", "build_number",
+            "cuda", "vulkan", "kompute", "metal", "sycl", "rpc", "gpu_blas", "blas",
+            "cpu_info", "gpu_info",
+            "model_filename", "model_type", "model_size", "model_n_params",
+            "n_batch", "n_ubatch",
+            "n_threads", "type_k", "type_v",
+            "n_gpu_layers", "split_mode",
+            "main_gpu", "no_kv_offload", "flash_attn",
+            "tensor_split", "use_mmap", "embeddings",
+            "n_prompt", "n_gen", "test_time",
+            "avg_ns", "stddev_ns",
+            "avg_ts", "stddev_ts"
+        };
+        return fields;
+    }
+
+    enum field_type {STRING, BOOL, INT, FLOAT};
+
+    static field_type get_field_type(const std::string & field) {
+        if (field == "build_number" || field == "n_batch" || field == "n_ubatch" ||
+            field == "n_threads" ||
+            field == "model_size" || field == "model_n_params" ||
+            field == "n_gpu_layers" || field == "main_gpu" ||
+            field == "n_prompt" || field == "n_gen" ||
+            field == "avg_ns" || field == "stddev_ns") {
+            return INT;
+        }
+        if (field == "cuda" || field == "vulkan" || field == "kompute" || field == "metal" ||
+            field == "gpu_blas" || field == "blas" || field == "sycl" ||field == "f16_kv" || field == "no_kv_offload" ||
+            field == "flash_attn" || field == "use_mmap" || field == "embeddings") {
+            return BOOL;
+        }
+        if (field == "avg_ts" || field == "stddev_ts") {
+            return FLOAT;
+        }
+        return STRING;
+    }
+
+    std::vector<std::string> get_values() const {
+        std::string tensor_split_str;
+        int max_nonzero = 0;
+        for (size_t i = 0; i < llama_max_devices(); i++) {
+            if (tensor_split[i] > 0) {
+                max_nonzero = i;
+            }
+        }
+        for (int i = 0; i <= max_nonzero; i++) {
+            char buf[32];
+            snprintf(buf, sizeof(buf), "%.2f", tensor_split[i]);
+            tensor_split_str += buf;
+            if (i < max_nonzero) {
+                tensor_split_str += "/";
+            }
+        }
+        std::vector<std::string> values = {
+            build_commit, std::to_string(build_number),
+            std::to_string(cuda), std::to_string(vulkan), std::to_string(vulkan),
+            std::to_string(metal), std::to_string(sycl), std::to_string(has_rpc), std::to_string(gpu_blas), std::to_string(blas),
+            cpu_info, gpu_info,
+            model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
+            std::to_string(n_batch), std::to_string(n_ubatch),
+            std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
+            std::to_string(n_gpu_layers), split_mode_str(split_mode),
+            std::to_string(main_gpu), std::to_string(no_kv_offload), std::to_string(flash_attn),
+            tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings),
+            std::to_string(n_prompt), std::to_string(n_gen), test_time,
+            std::to_string(avg_ns()), std::to_string(stdev_ns()),
+            std::to_string(avg_ts()), std::to_string(stdev_ts())
+        };
+        return values;
+    }
+
+    std::map<std::string, std::string> get_map() const {
+        std::map<std::string, std::string> map;
+        auto fields = get_fields();
+        auto values = get_values();
+        std::transform(fields.begin(), fields.end(), values.begin(),
+                std::inserter(map, map.end()), std::make_pair<const std::string &, const std::string &>);
+        return map;
+    }
+};
+
+const std::string test::build_commit = LLAMA_COMMIT;
+const int         test::build_number = LLAMA_BUILD_NUMBER;
+const bool        test::cuda         = !!ggml_cpu_has_cuda();
+const bool        test::vulkan       = !!ggml_cpu_has_vulkan();
+const bool        test::kompute      = !!ggml_cpu_has_kompute();
+const bool        test::metal        = !!ggml_cpu_has_metal();
+const bool        test::gpu_blas     = !!ggml_cpu_has_gpublas();
+const bool        test::blas         = !!ggml_cpu_has_blas();
+const bool        test::sycl         = !!ggml_cpu_has_sycl();
+const std::string test::cpu_info     = get_cpu_info();
+const std::string test::gpu_info     = get_gpu_info();
+
+struct printer {
+    virtual ~printer() {}
+
+    FILE * fout;
+    virtual void print_header(const cmd_params & params) { (void) params; }
+    virtual void print_test(const test & t) = 0;
+    virtual void print_footer() { }
+};
+
+struct csv_printer : public printer {
+    static std::string escape_csv(const std::string & field) {
+        std::string escaped = "\"";
+        for (auto c : field) {
+            if (c == '"') {
+                escaped += "\"";
+            }
+            escaped += c;
+        }
+        escaped += "\"";
+        return escaped;
+    }
+
+    void print_header(const cmd_params & params) override  {
+        std::vector<std::string> fields = test::get_fields();
+        fprintf(fout, "%s\n", join(fields, ",").c_str());
+        (void) params;
+    }
+
+    void print_test(const test & t) override {
+        std::vector<std::string> values = t.get_values();
+        std::transform(values.begin(), values.end(), values.begin(), escape_csv);
+        fprintf(fout, "%s\n", join(values, ",").c_str());
+    }
+};
+
+struct json_printer : public printer {
+    bool first = true;
+
+    static std::string escape_json(const std::string & value) {
+        std::string escaped;
+        for (auto c : value) {
+            if (c == '"') {
+                escaped += "\\\"";
+            } else if (c == '\\') {
+                escaped += "\\\\";
+            } else  if (c <= 0x1f) {
+                char buf[8];
+                snprintf(buf, sizeof(buf), "\\u%04x", c);
+                escaped += buf;
+            } else {
+                escaped += c;
+            }
+        }
+        return escaped;
+    }
+
+    static std::string format_value(const std::string & field, const std::string & value) {
+        switch (test::get_field_type(field)) {
+            case test::STRING:
+                return "\"" + escape_json(value) + "\"";
+            case test::BOOL:
+                return value == "0" ? "false" : "true";
+            default:
+                return value;
+        }
+    }
+
+    void print_header(const cmd_params & params) override {
+        fprintf(fout, "[\n");
+        (void) params;
+    }
+
+    void print_fields(const std::vector<std::string> & fields, const std::vector<std::string> & values) {
+        assert(fields.size() == values.size());
+        for (size_t i = 0; i < fields.size(); i++) {
+            fprintf(fout, "    \"%s\": %s,\n", fields.at(i).c_str(), format_value(fields.at(i), values.at(i)).c_str());
+        }
+    }
+
+    void print_test(const test & t) override {
+        if (first) {
+            first = false;
+        } else {
+            fprintf(fout, ",\n");
+        }
+        fprintf(fout, "  {\n");
+        print_fields(test::get_fields(), t.get_values());
+        fprintf(fout, "    \"samples_ns\": [ %s ],\n", join(t.samples_ns, ", ").c_str());
+        fprintf(fout, "    \"samples_ts\": [ %s ]\n", join(t.get_ts(), ", ").c_str());
+        fprintf(fout, "  }");
+        fflush(fout);
+    }
+
+    void print_footer() override {
+        fprintf(fout, "\n]\n");
+    }
+};
+
+struct markdown_printer : public printer {
+    std::vector<std::string> fields;
+
+    static int get_field_width(const std::string & field) {
+        if (field == "model") {
+            return -30;
+        }
+        if (field == "t/s") {
+            return 16;
+        }
+        if (field == "size" || field == "params") {
+            return 10;
+        }
+        if (field == "n_gpu_layers") {
+            return 3;
+        }
+        if (field == "n_threads") {
+            return 7;
+        }
+        if (field == "n_batch") {
+            return 7;
+        }
+        if (field == "n_ubatch") {
+            return 8;
+        }
+        if (field == "type_k" || field == "type_v") {
+            return 6;
+        }
+        if (field == "split_mode") {
+            return 5;
+        }
+        if (field == "flash_attn") {
+            return 2;
+        }
+        if (field == "use_mmap") {
+            return 4;
+        }
+        if (field == "test") {
+            return 13;
+        }
+
+        int width = std::max((int)field.length(), 10);
+
+        if (test::get_field_type(field) == test::STRING) {
+            return -width;
+        }
+        return width;
+    }
+
+    static std::string get_field_display_name(const std::string & field) {
+        if (field == "n_gpu_layers") {
+            return "ngl";
+        }
+        if (field == "split_mode") {
+            return "sm";
+        }
+        if (field == "n_threads") {
+            return "threads";
+        }
+        if (field == "no_kv_offload") {
+            return "nkvo";
+        }
+        if (field == "flash_attn") {
+            return "fa";
+        }
+        if (field == "use_mmap") {
+            return "mmap";
+        }
+        if (field == "embeddings") {
+            return "embd";
+        }
+        if (field == "tensor_split") {
+            return "ts";
+        }
+        return field;
+    }
+
+    void print_header(const cmd_params & params) override {
+        // select fields to print
+        fields.emplace_back("model");
+        fields.emplace_back("size");
+        fields.emplace_back("params");
+        fields.emplace_back("backend");
+        bool is_cpu_backend = test::get_backend() == "CPU" || test::get_backend() == "BLAS";
+        if (!is_cpu_backend) {
+            fields.emplace_back("n_gpu_layers");
+        }
+        if (params.n_threads.size() > 1 || params.n_threads != cmd_params_defaults.n_threads || is_cpu_backend) {
+            fields.emplace_back("n_threads");
+        }
+        if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) {
+            fields.emplace_back("n_batch");
+        }
+        if (params.n_ubatch.size() > 1 || params.n_ubatch != cmd_params_defaults.n_ubatch) {
+            fields.emplace_back("n_ubatch");
+        }
+        if (params.type_k.size() > 1 || params.type_k != cmd_params_defaults.type_k) {
+            fields.emplace_back("type_k");
+        }
+        if (params.type_v.size() > 1 || params.type_v != cmd_params_defaults.type_v) {
+            fields.emplace_back("type_v");
+        }
+        if (params.main_gpu.size() > 1 || params.main_gpu != cmd_params_defaults.main_gpu) {
+            fields.emplace_back("main_gpu");
+        }
+        if (params.split_mode.size() > 1 || params.split_mode != cmd_params_defaults.split_mode) {
+            fields.emplace_back("split_mode");
+        }
+        if (params.no_kv_offload.size() > 1 || params.no_kv_offload != cmd_params_defaults.no_kv_offload) {
+            fields.emplace_back("no_kv_offload");
+        }
+        if (params.flash_attn.size() > 1 || params.flash_attn != cmd_params_defaults.flash_attn) {
+            fields.emplace_back("flash_attn");
+        }
+        if (params.tensor_split.size() > 1 || params.tensor_split != cmd_params_defaults.tensor_split) {
+            fields.emplace_back("tensor_split");
+        }
+        if (params.use_mmap.size() > 1 || params.use_mmap != cmd_params_defaults.use_mmap) {
+            fields.emplace_back("use_mmap");
+        }
+        if (params.embeddings.size() > 1 || params.embeddings != cmd_params_defaults.embeddings) {
+            fields.emplace_back("embeddings");
+        }
+        fields.emplace_back("test");
+        fields.emplace_back("t/s");
+
+        fprintf(fout, "|");
+        for (const auto & field : fields) {
+            fprintf(fout, " %*s |", get_field_width(field), get_field_display_name(field).c_str());
+        }
+        fprintf(fout, "\n");
+        fprintf(fout, "|");
+        for (const auto & field : fields) {
+            int width = get_field_width(field);
+            fprintf(fout, " %s%s |", std::string(std::abs(width) - 1, '-').c_str(), width > 0 ? ":" : "-");
+        }
+        fprintf(fout, "\n");
+    }
+
+    void print_test(const test & t) override {
+        std::map<std::string, std::string> vmap = t.get_map();
+
+        fprintf(fout, "|");
+        for (const auto & field : fields) {
+            std::string value;
+            char buf[128];
+            if (field == "model") {
+                value = t.model_type;
+            } else if (field == "size") {
+                if (t.model_size < 1024*1024*1024) {
+                    snprintf(buf, sizeof(buf), "%.2f MiB", t.model_size / 1024.0 / 1024.0);
+                } else {
+                    snprintf(buf, sizeof(buf), "%.2f GiB", t.model_size / 1024.0 / 1024.0 / 1024.0);
+                }
+                value = buf;
+            } else if (field == "params") {
+                if (t.model_n_params < 1000*1000*1000) {
+                    snprintf(buf, sizeof(buf), "%.2f M", t.model_n_params / 1e6);
+                } else {
+                    snprintf(buf, sizeof(buf), "%.2f B", t.model_n_params / 1e9);
+                }
+                value = buf;
+            } else if (field == "backend") {
+                value = test::get_backend();
+                if (t.has_rpc) {
+                    value += "+RPC";
+                }
+            } else if (field == "test") {
+                if (t.n_prompt > 0 && t.n_gen == 0) {
+                    snprintf(buf, sizeof(buf), "pp%d", t.n_prompt);
+                } else if (t.n_gen > 0 && t.n_prompt == 0) {
+                    snprintf(buf, sizeof(buf), "tg%d", t.n_gen);
+                } else {
+                    snprintf(buf, sizeof(buf), "pp%d+tg%d", t.n_prompt, t.n_gen);
+                }
+                value = buf;
+            } else if (field == "t/s") {
+                snprintf(buf, sizeof(buf), "%.2f ± %.2f", t.avg_ts(), t.stdev_ts());
+                value = buf;
+            } else if (vmap.find(field) != vmap.end()) {
+                value = vmap.at(field);
+            } else {
+                assert(false);
+                exit(1);
+            }
+
+            int width = get_field_width(field);
+            if (field == "t/s") {
+                // HACK: the utf-8 character is 2 bytes
+                width += 1;
+            }
+            fprintf(fout, " %*s |", width, value.c_str());
+        }
+        fprintf(fout, "\n");
+    }
+
+    void print_footer() override {
+        fprintf(fout, "\nbuild: %s (%d)\n", test::build_commit.c_str(), test::build_number);
+    }
+};
+
+struct sql_printer : public printer {
+    static std::string get_sql_field_type(const std::string & field) {
+        switch (test::get_field_type(field)) {
+            case test::STRING:
+                return "TEXT";
+            case test::BOOL:
+            case test::INT:
+                return "INTEGER";
+            case test::FLOAT:
+                return "REAL";
+            default:
+                assert(false);
+                exit(1);
+        }
+    }
+
+    void print_header(const cmd_params & params) override {
+        std::vector<std::string> fields = test::get_fields();
+        fprintf(fout, "CREATE TABLE IF NOT EXISTS test (\n");
+        for (size_t i = 0; i < fields.size(); i++) {
+            fprintf(fout, "  %s %s%s\n", fields.at(i).c_str(), get_sql_field_type(fields.at(i)).c_str(),  i < fields.size() - 1 ? "," : "");
+        }
+        fprintf(fout, ");\n");
+        fprintf(fout, "\n");
+        (void) params;
+    }
+
+    void print_test(const test & t) override {
+        fprintf(fout, "INSERT INTO test (%s) ", join(test::get_fields(), ", ").c_str());
+        fprintf(fout, "VALUES (");
+        std::vector<std::string> values = t.get_values();
+        for (size_t i = 0; i < values.size(); i++) {
+            fprintf(fout, "'%s'%s", values.at(i).c_str(), i < values.size() - 1 ? ", " : "");
+        }
+        fprintf(fout, ");\n");
+    }
+};
+
+static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
+    llama_set_n_threads(ctx, n_threads, n_threads);
+
+    const llama_model * model = llama_get_model(ctx);
+    const int32_t n_vocab = llama_n_vocab(model);
+
+    std::vector<llama_token> tokens(n_batch);
+
+    int n_processed = 0;
+
+    while (n_processed < n_prompt) {
+        int n_tokens = std::min(n_prompt - n_processed, n_batch);
+        tokens[0] = n_processed == 0 && llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
+        for (int i = 1; i < n_tokens; i++) {
+            tokens[i] = std::rand() % n_vocab;
+        }
+        llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens, n_past + n_processed, 0));
+        n_processed += n_tokens;
+    }
+
+    llama_synchronize(ctx);
+}
+
+static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
+    llama_set_n_threads(ctx, n_threads, n_threads);
+
+    const llama_model * model = llama_get_model(ctx);
+    const int32_t n_vocab = llama_n_vocab(model);
+
+    llama_token token = llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
+
+    for (int i = 0; i < n_gen; i++) {
+        llama_decode(ctx, llama_batch_get_one(&token, 1, n_past + i, 0));
+        llama_synchronize(ctx);
+        token = std::rand() % n_vocab;
+    }
+}
+
+static void llama_null_log_callback(enum ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) text;
+    (void) user_data;
+}
+
+static std::unique_ptr<printer> create_printer(output_formats format) {
+    switch (format) {
+        case NONE:
+            return nullptr;
+        case CSV:
+            return std::unique_ptr<printer>(new csv_printer());
+        case JSON:
+            return std::unique_ptr<printer>(new json_printer());
+        case MARKDOWN:
+            return std::unique_ptr<printer>(new markdown_printer());
+        case SQL:
+            return std::unique_ptr<printer>(new sql_printer());
+    }
+    GGML_ABORT("fatal error");
+}
+
+int main(int argc, char ** argv) {
+    // try to set locale for unicode characters in markdown
+    setlocale(LC_CTYPE, ".UTF-8");
+
+#if !defined(NDEBUG)
+    fprintf(stderr, "warning: asserts enabled, performance may be affected\n");
+#endif
+
+#if (defined(_MSC_VER) && defined(_DEBUG)) || (!defined(_MSC_VER) && !defined(__OPTIMIZE__))
+    fprintf(stderr, "warning: debug build, performance may be affected\n");
+#endif
+
+#if defined(__SANITIZE_ADDRESS__) || defined(__SANITIZE_THREAD__)
+    fprintf(stderr, "warning: sanitizer enabled, performance may be affected\n");
+#endif
+
+    cmd_params params = parse_cmd_params(argc, argv);
+
+    // initialize llama.cpp
+    if (!params.verbose) {
+        llama_log_set(llama_null_log_callback, NULL);
+    }
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    // initialize printer
+    std::unique_ptr<printer> p = create_printer(params.output_format);
+    std::unique_ptr<printer> p_err = create_printer(params.output_format_stderr);
+
+    if (p) {
+        p->fout = stdout;
+        p->print_header(params);
+    }
+
+    if (p_err) {
+        p_err->fout = stderr;
+        p_err->print_header(params);
+    }
+
+    std::vector<cmd_params_instance> params_instances = get_cmd_params_instances(params);
+
+    llama_model * lmodel = nullptr;
+    const cmd_params_instance * prev_inst = nullptr;
+
+    for (const auto & inst : params_instances) {
+        // keep the same model between tests when possible
+        if (!lmodel || !prev_inst || !inst.equal_mparams(*prev_inst)) {
+            if (lmodel) {
+                llama_free_model(lmodel);
+            }
+
+            lmodel = llama_load_model_from_file(inst.model.c_str(), inst.to_llama_mparams());
+            if (lmodel == NULL) {
+                fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, inst.model.c_str());
+                return 1;
+            }
+            prev_inst = &inst;
+        }
+
+        llama_context * ctx = llama_new_context_with_model(lmodel, inst.to_llama_cparams());
+        if (ctx == NULL) {
+            fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, inst.model.c_str());
+            llama_free_model(lmodel);
+            return 1;
+        }
+
+        test t(inst, lmodel, ctx);
+
+        llama_kv_cache_clear(ctx);
+
+        // warmup run
+        if (t.n_prompt > 0) {
+            //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
+            test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
+        }
+        if (t.n_gen > 0) {
+            test_gen(ctx, 1, 0, t.n_threads);
+        }
+
+        for (int i = 0; i < params.reps; i++) {
+            llama_kv_cache_clear(ctx);
+
+            uint64_t t_start = get_time_ns();
+
+            if (t.n_prompt > 0) {
+                test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
+            }
+            if (t.n_gen > 0) {
+                test_gen(ctx, t.n_gen, t.n_prompt, t.n_threads);
+            }
+
+            uint64_t t_ns = get_time_ns() - t_start;
+            t.samples_ns.push_back(t_ns);
+        }
+
+        if (p) {
+            p->print_test(t);
+            fflush(p->fout);
+        }
+
+        if (p_err) {
+            p_err->print_test(t);
+            fflush(p_err->fout);
+        }
+
+        llama_print_timings(ctx);
+
+        llama_free(ctx);
+    }
+
+    llama_free_model(lmodel);
+
+    if (p) {
+        p->print_footer();
+    }
+
+    if (p_err) {
+        p_err->print_footer();
+    }
+
+    llama_backend_free();
+
+    return 0;
+}

examples/llama.android/.gitignore

+# Gradle files
+.gradle/
+build/
+
+# Local configuration file (sdk path, etc)
+local.properties
+
+# Log/OS Files
+*.log
+
+# Android Studio generated files and folders
+captures/
+.externalNativeBuild/
+.cxx/
+*.apk
+output.json
+
+# IntelliJ
+*.iml
+.idea/
+misc.xml
+deploymentTargetDropDown.xml
+render.experimental.xml
+
+# Keystore files
+*.jks
+*.keystore
+
+# Google Services (e.g. APIs or Firebase)
+google-services.json
+
+# Android Profiling
+*.hprof

examples/llama.android/app/.gitignore

+/build