infill : add new example + extend server API (#3296)

* vvhg-code-infill (#1)

* infill in separate example (#2)

* reverted changes to main and added infill example

* cleanup

* naming improvement

* make : add missing blank line

* fix missing semicolon

* brought infill up to current main code

* cleanup

---------

Co-authored-by: Cebtenzzre <cebtenzzre@gmail.com>

.gitignore

@@ -40,6 +40,7 @@ models-mnt
 /embedding
 /gguf
 /gguf-llama-simple
+/infill
 /libllama.so
 /llama-bench
 /main

Makefile

@@ -1,5 +1,5 @@
 # Define the default target now so that it is always the first target
-BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml simple batched save-load-state server embd-input-test gguf llama-bench baby-llama beam-search speculative benchmark-matmult parallel finetune export-lora tests/test-c.o
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml simple batched save-load-state server embd-input-test gguf llama-bench baby-llama beam-search speculative infill benchmark-matmult parallel finetune export-lora tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama
@@ -543,6 +543,9 @@ main: examples/main/main.cpp                                  build-info.h ggml.
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
+infill: examples/infill/infill.cpp                            build-info.h ggml.o llama.o common.o console.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 simple: examples/simple/simple.cpp                            build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 

common/common.cpp

@@ -389,6 +389,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.interactive_first = true;
         } else if (arg == "-ins" || arg == "--instruct") {
             params.instruct = true;
+        } else if (arg == "--infill") {
+            params.infill = true;
         } else if (arg == "--multiline-input") {
             params.multiline_input = true;
         } else if (arg == "--simple-io") {

common/common.h

@@ -120,6 +120,7 @@ struct gpt_params {
     bool use_mlock         = false; // use mlock to keep model in memory
     bool numa              = false; // attempt optimizations that help on some NUMA systems
     bool verbose_prompt    = false; // print prompt tokens before generation
+    bool infill            = false; // use infill mode
 };
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params);

examples/infill/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(TARGET 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)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/infill/README.md

@@ -0,0 +1,41 @@
+# 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.
+
+## 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
+
+```bash
+./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

@@ -0,0 +1,769 @@
+#include "common.h"
+
+#include "console.h"
+#include "llama.h"
+#include "build-info.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 = get_sortable_timestamp();
+
+    const bool success = 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));
+    dump_non_result_info_yaml(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");
+
+    dump_string_yaml_multiline(logfile, "output", output.c_str());
+    dump_vector_int_yaml(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;
+    g_params = &params;
+
+    if (!gpt_params_parse(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.instruct) {
+        printf("\n************\n");
+        printf("%s: please use the 'main' tool for instruct mode\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+    if (!params.antiprompt.empty()) {
+        printf("\n************\n");
+        printf("%s: please use the 'main' tool for antiprompt mode\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+    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.random_prompt) {
+        printf("\n************\n");
+        printf("%s: please use the 'main' tool for random prompt mode\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+    if (!params.path_prompt_cache.empty()) {
+        printf("\n************\n");
+        printf("%s: infill does not support prompt caching\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__, BUILD_NUMBER, BUILD_COMMIT);
+    LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, 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(params.numa);
+
+    llama_model * model;
+    llama_context * ctx;
+    llama_context * ctx_guidance = NULL;
+    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__);
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (params.cfg_scale > 1.f) {
+        struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
+        ctx_guidance = llama_new_context_with_model(model, lparams);
+    }
+
+    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", get_system_info(params).c_str());
+    }
+    const bool add_bos = llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM;
+    LOG("add_bos: %d\n", add_bos);
+
+    std::vector<llama_token> embd_inp;
+    std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, add_bos);
+    std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, add_bos);
+    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
+    embd_inp = inp_pfx;
+    embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
+    embd_inp.push_back(llama_token_middle(ctx));
+
+    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));
+
+    // Should not run without any tokens
+    if (embd_inp.empty()) {
+        embd_inp.push_back(llama_token_bos(ctx));
+        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
+    }
+
+    // Tokenize negative prompt
+    std::vector<llama_token> guidance_inp;
+    int guidance_offset = 0;
+    int original_prompt_len = 0;
+    if (ctx_guidance) {
+        LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(params.cfg_negative_prompt));
+
+        guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, add_bos);
+        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp));
+
+        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
+        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp));
+
+        original_prompt_len = original_inp.size();
+        guidance_offset = (int)guidance_inp.size() - original_prompt_len;
+        LOG("original_prompt_len: %s", log_tostr(original_prompt_len));
+        LOG("guidance_offset:     %s", log_tostr(guidance_offset));
+    }
+
+    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));
+    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx));
+
+
+    // 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 (ctx_guidance) {
+            LOG_TEE("\n");
+            LOG_TEE("%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str());
+            LOG_TEE("%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
+            for (int i = 0; i < (int) guidance_inp.size(); i++) {
+                LOG_TEE("%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_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: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
+            params.repeat_last_n, params.repeat_penalty, params.presence_penalty, params.frequency_penalty, params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp, params.mirostat, params.mirostat_eta, params.mirostat_tau);
+    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");
+
+    struct llama_grammar * grammar = NULL;
+    grammar_parser::parse_state parsed_grammar;
+
+    if (!params.grammar.empty()) {
+        parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+        // will be empty (default) if there are parse errors
+        if (parsed_grammar.rules.empty()) {
+            return 1;
+        }
+        LOG_TEE("%s: grammar:\n", __func__);
+        grammar_parser::print_grammar(stderr, parsed_grammar);
+        LOG_TEE("\n");
+
+        {
+            auto it = params.logit_bias.find(llama_token_eos(ctx));
+            if (it != params.logit_bias.end() && it->second == -INFINITY) {
+                LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
+            }
+        }
+
+        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+        grammar = llama_grammar_init(
+            grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
+    }
+
+    // TODO: replace with ring-buffer
+    std::vector<llama_token> last_tokens(n_ctx);
+    std::fill(last_tokens.begin(), last_tokens.end(), 0);
+    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;
+    int n_past_guidance    = 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;
+    std::vector<llama_token> embd_guidance;
+
+    const int n_vocab = llama_n_vocab(model);
+
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
+
+    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() + std::max<int>(0, guidance_offset) > 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_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+
+                n_past -= n_discard;
+
+                if (ctx_guidance) {
+                    n_past_guidance -= n_discard;
+                }
+
+                LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
+
+                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+
+            }
+
+            // evaluate tokens in batches
+            // embd is typically prepared beforehand to fit within a batch, but not always
+
+            if (ctx_guidance) {
+                int input_size = 0;
+                llama_token * input_buf = NULL;
+
+                if (n_past_guidance < (int) guidance_inp.size()) {
+                    // Guidance context should have the same data with these modifications:
+                    //
+                    // * Replace the initial prompt
+                    // * Shift everything by guidance_offset
+                    embd_guidance = guidance_inp;
+                    if (embd.begin() + original_prompt_len < embd.end()) {
+                        embd_guidance.insert(
+                            embd_guidance.end(),
+                            embd.begin() + original_prompt_len,
+                            embd.end()
+                        );
+                    }
+
+                    input_buf  = embd_guidance.data();
+                    input_size = embd_guidance.size();
+
+                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance));
+                } else {
+                    input_buf  = embd.data();
+                    input_size = embd.size();
+                }
+
+                for (int i = 0; i < input_size; i += params.n_batch) {
+                    int n_eval = std::min(input_size - i, params.n_batch);
+                    if (llama_decode(ctx_guidance, llama_batch_get_one(input_buf + i, n_eval, n_past_guidance, 0))) {
+                        LOG_TEE("%s : failed to eval\n", __func__);
+                        return 1;
+                    }
+
+                    n_past_guidance += n_eval;
+                }
+            }
+
+            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));
+
+                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();
+        embd_guidance.clear();
+
+        if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
+
+            const llama_token id = llama_sample_token(ctx, ctx_guidance, grammar, params, last_tokens, candidates);
+
+            last_tokens.erase(last_tokens.begin());
+            last_tokens.push_back(id);
+
+            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_tokens));
+
+            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]);
+                last_tokens.erase(last_tokens.begin());
+                last_tokens.push_back(embd_inp[n_consumed]);
+                ++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 ((last_tokens.back() == llama_token_eot(ctx) || is_interacting) && params.interactive){
+                if(is_interacting && !params.interactive_first) {
+                    // print an eot token
+                    printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).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);
+                // tokenize new prefix and suffix
+                std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, add_bos);
+                std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, add_bos);
+                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
+                embd_inp = inp_pfx;
+                embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
+                embd_inp.push_back(llama_token_middle(ctx));
+                embd.clear();
+                embd_guidance.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 text token in interactive mode
+            else if (last_tokens.back() == llama_token_eos(ctx)) {
+                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(ctx));
+                }
+
+                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));
+
+                    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) {
+                    // reset grammar state if we're restarting generation
+                    if (grammar != NULL) {
+                        llama_grammar_free(grammar);
+
+                        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+                        grammar = llama_grammar_init(
+                            grammar_rules.data(), grammar_rules.size(),
+                            parsed_grammar.symbol_ids.at("root"));
+                    }
+                }
+                is_interacting = false;
+            }
+        }
+
+        // end of text token
+        if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !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(ctx)).c_str());
+        fflush(stdout);
+    }
+
+    llama_print_timings(ctx);
+    write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
+
+    if (ctx_guidance) { llama_free(ctx_guidance); }
+    llama_free(ctx);
+    llama_free_model(model);
+
+    if (grammar != NULL) {
+        llama_grammar_free(grammar);
+    }
+    llama_backend_free();
+
+#ifndef LOG_DISABLE_LOGS
+    LOG_TEE("Log end\n");
+#endif // LOG_DISABLE_LOGS
+
+    return 0;
+}
+

examples/server/README.md

@@ -176,6 +176,16 @@ node index.js
 
     `content`: Set the text to process.
 
+    **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
+
+    *Options:*
+
+    `input_prefix`: Set the prefix of the code to infill.
+
+    `input_suffix`: Set the suffix of the code to infill.
+
+    It also accepts all the options of `/completion` except `stream` and `prompt`.
+
 ## More examples
 
 ### Interactive mode

examples/server/server.cpp

@@ -342,6 +342,70 @@ struct llama_server_context
         return true;
     }
 
+    void loadInfill()
+    {
+        auto prefix_tokens = tokenize(params.input_prefix, true);  // always add BOS
+        auto suffix_tokens = tokenize(params.input_suffix, true);  // always add BOS
+        prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(ctx));
+        prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(ctx));
+        prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end());
+        prefix_tokens.push_back(llama_token_middle(ctx));
+        auto prompt_tokens = prefix_tokens;
+
+        num_prompt_tokens = prompt_tokens.size();
+
+        if (params.n_keep < 0)
+        {
+            params.n_keep = (int)num_prompt_tokens;
+        }
+        params.n_keep = std::min(params.n_ctx - 4, params.n_keep);
+
+        // if input prompt is too big, truncate like normal
+        if (num_prompt_tokens >= (size_t)params.n_ctx)
+        {
+            printf("Input prompt is too big, truncating. Can only take %d tokens but got %zu\n", params.n_ctx, num_prompt_tokens);
+            // todo we probably want to cut from both sides
+            const int n_left = (params.n_ctx - params.n_keep) / 2;
+            std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep);
+            const int erased_blocks = (num_prompt_tokens - params.n_keep - n_left - 1) / n_left;
+            new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + params.n_keep + erased_blocks * n_left, prompt_tokens.end());
+            std::copy(prompt_tokens.end() - params.n_ctx, prompt_tokens.end(), last_n_tokens.begin());
+
+            LOG_VERBOSE("input truncated", {
+                                               {"n_ctx", params.n_ctx},
+                                               {"n_keep", params.n_keep},
+                                               {"n_left", n_left},
+                                               {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
+                                           });
+
+            truncated = true;
+            prompt_tokens = new_tokens;
+        }
+        else
+        {
+            const size_t ps = num_prompt_tokens;
+            std::fill(last_n_tokens.begin(), last_n_tokens.end() - ps, 0);
+            std::copy(prompt_tokens.begin(), prompt_tokens.end(), last_n_tokens.end() - ps);
+        }
+
+        // compare the evaluated prompt with the new prompt
+        n_past = common_part(embd, prompt_tokens);
+        embd = prompt_tokens;
+        if (n_past == num_prompt_tokens)
+        {
+            // we have to evaluate at least 1 token to generate logits.
+            printf("we have to evaluate at least 1 token to generate logits\n");
+            n_past--;
+        }
+
+        LOG_VERBOSE("prompt ingested", {
+                                           {"n_past", n_past},
+                                           {"cached", tokens_to_str(ctx, embd.cbegin(), embd.cbegin() + n_past)},
+                                           {"to_eval", tokens_to_str(ctx, embd.cbegin() + n_past, embd.cend())},
+                                       });
+
+        has_next_token = true;
+    }
     void loadPrompt()
     {
         auto prompt_tokens = tokenize(prompt, true);  // always add BOS
@@ -1219,6 +1283,27 @@ static void parse_options_completion(const json &body, llama_server_context &lla
     LOG_VERBOSE("completion parameters parsed", format_generation_settings(llama));
 }
 
+static void parse_options_infill(const json &body, llama_server_context &llama)
+{
+    if (body.count("input_prefix") != 0)
+    {
+        llama.params.input_prefix = body["input_prefix"];
+    }
+    else
+    {
+        llama.params.input_prefix = "";
+    }
+    if (body.count("input_suffix") != 0)
+    {
+        llama.params.input_suffix = body["input_suffix"];
+    }
+    else
+    {
+        llama.params.input_suffix = "";
+    }
+    parse_options_completion(body, llama);
+}
+
 static void log_server_request(const Request &req, const Response &res)
 {
     LOG_INFO("request", {
@@ -1519,6 +1604,127 @@ int main(int argc, char **argv)
             res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
         } });
 
+    svr.Post("/infill", [&llama](const Request &req, Response &res)
+             {
+        auto lock = llama.lock();
+
+        llama.rewind();
+
+        llama_reset_timings(llama.ctx);
+
+        parse_options_infill(json::parse(req.body), llama);
+
+        if (!llama.loadGrammar())
+        {
+            res.status = 400;
+            return;
+        }
+        llama.loadInfill();
+        llama.beginCompletion();
+        const auto chunked_content_provider = [&](size_t, DataSink & sink) {
+            size_t sent_count = 0;
+            size_t sent_token_probs_index = 0;
+
+            while (llama.has_next_token) {
+                const completion_token_output token_with_probs = llama.doCompletion();
+                if (token_with_probs.tok == -1 || llama.multibyte_pending > 0) {
+                    continue;
+                }
+                const std::string token_text = llama_token_to_piece(llama.ctx, token_with_probs.tok);
+
+                size_t pos = std::min(sent_count, llama.generated_text.size());
+
+                const std::string str_test = llama.generated_text.substr(pos);
+                bool is_stop_full = false;
+                size_t stop_pos =
+                    llama.findStoppingStrings(str_test, token_text.size(), STOP_FULL);
+                if (stop_pos != std::string::npos) {
+                    is_stop_full = true;
+                    llama.generated_text.erase(
+                        llama.generated_text.begin() + pos + stop_pos,
+                        llama.generated_text.end());
+                    pos = std::min(sent_count, llama.generated_text.size());
+                } else {
+                    is_stop_full = false;
+                    stop_pos = llama.findStoppingStrings(str_test, token_text.size(),
+                        STOP_PARTIAL);
+                }
+
+                if (
+                    stop_pos == std::string::npos ||
+                    // Send rest of the text if we are at the end of the generation
+                    (!llama.has_next_token && !is_stop_full && stop_pos > 0)
+                ) {
+                    const std::string to_send = llama.generated_text.substr(pos, std::string::npos);
+
+                    sent_count += to_send.size();
+
+                    std::vector<completion_token_output> probs_output = {};
+
+                    if (llama.params.n_probs > 0) {
+                        const std::vector<llama_token> to_send_toks = llama_tokenize(llama.ctx, to_send, false);
+                        size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size());
+                        size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size());
+                        if (probs_pos < probs_stop_pos) {
+                            probs_output = std::vector<completion_token_output>(llama.generated_token_probs.begin() + probs_pos, llama.generated_token_probs.begin() + probs_stop_pos);
+                        }
+                        sent_token_probs_index = probs_stop_pos;
+                    }
+
+                    const json data = format_partial_response(llama, to_send, probs_output);
+
+                    const std::string str =
+                        "data: " +
+                        data.dump(-1, ' ', false, json::error_handler_t::replace) +
+                        "\n\n";
+
+                    LOG_VERBOSE("data stream", {
+                        { "to_send", str }
+                    });
+
+                    if (!sink.write(str.data(), str.size())) {
+                        LOG_VERBOSE("stream closed", {});
+                        llama_print_timings(llama.ctx);
+                        return false;
+                    }
+                }
+
+                if (!llama.has_next_token) {
+                    // Generation is done, send extra information.
+                    const json data = format_final_response(
+                        llama,
+                        "",
+                        std::vector<completion_token_output>(llama.generated_token_probs.begin(), llama.generated_token_probs.begin() + sent_token_probs_index)
+                    );
+
+                    const std::string str =
+                        "data: " +
+                        data.dump(-1, ' ', false, json::error_handler_t::replace) +
+                        "\n\n";
+
+                    LOG_VERBOSE("data stream", {
+                        { "to_send", str }
+                    });
+
+                    if (!sink.write(str.data(), str.size())) {
+                        LOG_VERBOSE("stream closed", {});
+                        llama_print_timings(llama.ctx);
+                        return false;
+                    }
+                }
+            }
+
+            llama_print_timings(llama.ctx);
+            sink.done();
+            return true;
+        };
+        const auto on_complete = [&](bool) {
+            llama.mutex.unlock();
+        };
+        lock.release();
+        res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
+        });
+
     svr.Get("/model.json", [&llama](const Request &, Response &res)
             {
         const json data = format_generation_settings(llama);

llama.cpp

@@ -1076,6 +1076,10 @@ struct llama_vocab {
     id special_pad_id = -1;
 
     id linefeed_id = 13;
+    id special_prefix_id = 32007;
+    id special_middle_id = 32009;
+    id special_suffix_id = 32008;
+    id special_eot_id = 32010;
 
     int find_bpe_rank(std::string token_left, std::string token_right) const {
         replace_all(token_left,  " ",  "\u0120");
@@ -7489,6 +7493,22 @@ llama_token llama_token_eos(const struct llama_context * ctx) {
 llama_token llama_token_nl(const struct llama_context * ctx) {
     return ctx->model.vocab.linefeed_id;
 }
+llama_token llama_token_prefix(const struct llama_context * ctx) {
+    return ctx->model.vocab.special_prefix_id;
+}
+
+llama_token llama_token_middle(const struct llama_context * ctx) {
+    return ctx->model.vocab.special_middle_id;
+}
+
+llama_token llama_token_suffix(const struct llama_context * ctx) {
+    return ctx->model.vocab.special_suffix_id;
+}
+
+llama_token llama_token_eot(const struct llama_context * ctx) {
+    return ctx->model.vocab.special_eot_id;
+}
+
 
 int llama_tokenize(
     const struct llama_model * model,

llama.h

@@ -490,6 +490,11 @@ extern "C" {
     LLAMA_API llama_token llama_token_bos(const struct llama_context * ctx);  // beginning-of-sentence
     LLAMA_API llama_token llama_token_eos(const struct llama_context * ctx);  // end-of-sentence
     LLAMA_API llama_token llama_token_nl (const struct llama_context * ctx);  // next-line
+    // codellama infill tokens
+    LLAMA_API llama_token llama_token_prefix(const struct llama_context * ctx); // Beginning of infill prefix
+    LLAMA_API llama_token llama_token_middle(const struct llama_context * ctx); // Beginning of infill middle
+    LLAMA_API llama_token llama_token_suffix(const struct llama_context * ctx); // Beginning of infill suffix
+    LLAMA_API llama_token llama_token_eot   (const struct llama_context * ctx); // End of infill middle
 
     //
     // Tokenization