tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-26 11:45:21 +00:00
Code Issues Packages Projects Releases Wiki Activity

examples : remove infill (#13283)

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2025-05-07 10:28:02 +03:00
committed by GitHub
parent 6c7fd67b64
commit 4773d7a02f
7 changed files with 6 additions and 655 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
Makefile
View File

@ -1187,11 +1187,6 @@ llama-cli: tools/main/main.cpp \
@echo '==== Run ./llama-cli -h for help. ===='
@echo
llama-infill: examples/infill/infill.cpp \
$(OBJ_ALL)
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
llama-run: tools/run/run.cpp \
$(OBJ_ALL)
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)

12
common/arg.cpp
View File

@ -1283,7 +1283,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
[](common_params & params) {
params.use_color = true;
}
).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL, LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP}));
).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP}));
add_opt(common_arg(
{"-t", "--threads"}, "N",
string_format("number of threads to use during generation (default: %d)", params.cpuparams.n_threads),
@ -1416,7 +1416,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
add_opt(common_arg(
{"-n", "--predict", "--n-predict"}, "N",
string_format(
ex == LLAMA_EXAMPLE_MAIN || ex == LLAMA_EXAMPLE_INFILL
ex == LLAMA_EXAMPLE_MAIN
? "number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)"
: "number of tokens to predict (default: %d, -1 = infinity)",
params.n_predict),
@ -1655,7 +1655,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.input_prefix = value;
params.enable_chat_template = false;
}
).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL}));
).set_examples({LLAMA_EXAMPLE_MAIN}));
add_opt(common_arg(
{"--in-suffix"}, "STRING",
"string to suffix after user inputs with (default: empty)",
@ -1663,7 +1663,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.input_suffix = value;
params.enable_chat_template = false;
}
).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL}));
).set_examples({LLAMA_EXAMPLE_MAIN}));
add_opt(common_arg(
{"--no-warmup"},
"skip warming up the model with an empty run",
@ -1680,7 +1680,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
[](common_params & params) {
params.spm_infill = true;
}
).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_INFILL}));
).set_examples({LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"--samplers"}, "SAMPLERS",
string_format("samplers that will be used for generation in the order, separated by \';\'\n(default: %s)", sampler_type_names.c_str()),
@ -2892,7 +2892,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
[](common_params & params) {
params.simple_io = true;
}
).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL}));
).set_examples({LLAMA_EXAMPLE_MAIN}));
add_opt(common_arg(
{"--positive-file"}, "FNAME",
string_format("positive prompts file, one prompt per line (default: '%s')", params.cvector_positive_file.c_str()),

1
common/common.h
View File

@ -66,7 +66,6 @@ enum llama_example {
LLAMA_EXAMPLE_COMMON,
LLAMA_EXAMPLE_SPECULATIVE,
LLAMA_EXAMPLE_MAIN,
LLAMA_EXAMPLE_INFILL,
LLAMA_EXAMPLE_EMBEDDING,
LLAMA_EXAMPLE_PERPLEXITY,
LLAMA_EXAMPLE_RETRIEVAL,

1
examples/CMakeLists.txt
View File

@ -21,7 +21,6 @@ else()
add_subdirectory(gguf-hash)
add_subdirectory(gguf)
add_subdirectory(gritlm)
add_subdirectory(infill)
add_subdirectory(lookahead)
add_subdirectory(lookup)
add_subdirectory(parallel)

5
examples/infill/CMakeLists.txt
View File

@ -1,5 +0,0 @@
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_17)

47
examples/infill/README.md
View File

@ -1,47 +0,0 @@
# 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 4096, but if a LLaMA model was built with a longer context, increasing this value 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 "
```

590
examples/infill/infill.cpp
View File

@ -1,590 +0,0 @@
#include "arg.h"
#include "common.h"
#include "console.h"
#include "sampling.h"
#include "log.h"
#include "llama.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 common_sampler ** g_smpl;
static common_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;
#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();
LOG("\n");
common_perf_print(*g_ctx, *g_smpl);
// make sure all logs are flushed
LOG("Interrupted by user\n");
common_log_pause(common_log_main());
_exit(130);
}
}
}
#endif
int main(int argc, char ** argv) {
common_params params;
g_params = &params;
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_INFILL)) {
return 1;
}
common_init();
auto & sparams = params.sampling;
console::init(params.simple_io, params.use_color);
atexit([]() { console::cleanup(); });
if (params.logits_all) {
LOG_ERR("\n************\n");
LOG_ERR("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
LOG_ERR("************\n\n");
return 0;
}
if (params.embedding) {
LOG_ERR("\n************\n");
LOG_ERR("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
LOG_ERR("************\n\n");
return 0;
}
if (params.n_ctx != 0 && params.n_ctx < 8) {
LOG_WRN("%s: 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())) {
LOG_ERR("\n************\n");
LOG_ERR("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__);
LOG_ERR("************\n\n");
return 0;
}
if (params.rope_freq_base != 0.0) {
LOG_WRN("%s: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
}
if (params.rope_freq_scale != 0.0) {
LOG_WRN("%s: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
}
LOG_INF("%s: llama backend init\n", __func__);
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model = nullptr;
llama_context * ctx = nullptr;
common_sampler * smpl = nullptr;
g_model = &model;
g_ctx = &ctx;
g_smpl = &smpl;
// load the model and apply lora adapter, if any
LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
common_init_result llama_init = common_init_from_params(params);
model = llama_init.model.get();
ctx = llama_init.context.get();
if (model == NULL) {
LOG_ERR("%s: unable to load model\n", __func__);
return 1;
}
const llama_vocab * vocab = llama_model_get_vocab(model);
const int n_ctx_train = llama_model_n_ctx_train(model);
const int n_ctx = llama_n_ctx(ctx);
LOG_DBG("n_ctx: %d\n", n_ctx);
if (n_ctx > n_ctx_train) {
LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n", __func__, n_ctx_train, n_ctx);
}
// print system information
{
LOG_INF("\n");
LOG_INF("%s\n", common_params_get_system_info(params).c_str());
}
const bool add_bos = llama_vocab_get_add_bos(vocab);
GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
std::vector<llama_token> embd_inp;
std::vector<llama_token> embd_end;
std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
GGML_ASSERT(llama_vocab_fim_pre(vocab) >= 0);
GGML_ASSERT(llama_vocab_fim_suf(vocab) >= 0);
inp_pfx.insert(inp_pfx.begin(), llama_vocab_fim_pre(vocab));
inp_sfx.insert(inp_sfx.begin(), llama_vocab_fim_suf(vocab));
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_vocab_bos(vocab));
}
embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
const llama_token middle_token = llama_vocab_fim_mid(vocab);
if (middle_token >= 0) {
embd_inp.push_back(middle_token);
}
LOG_DBG("add_bos: %d\n", add_bos);
LOG_DBG("prefix: \"%s\"\n", params.input_prefix.c_str());
LOG_DBG("suffix: \"%s\"\n", params.input_suffix.c_str());
LOG_DBG("tokens: %s\n", string_from(ctx, embd_inp).c_str());
// Should not run without any tokens
if (embd_inp.empty()) {
embd_inp.push_back(llama_vocab_bos(vocab));
LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str());
}
if ((int) embd_inp.size() > n_ctx - 4) {
LOG_ERR("%s: 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_INF("inp_pfx: %s\n", string_from(ctx, inp_pfx).c_str());
LOG_INF("inp_sfx: %s\n", string_from(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_INF("\n");
LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
for (int i = 0; i < (int) embd_inp.size(); i++) {
LOG_INF("%6d -> '%s'\n", embd_inp[i], common_token_to_piece(ctx, embd_inp[i]).c_str());
}
if (params.n_keep > 0) {
LOG_INF("%s: static prompt based on n_keep: '", __func__);
for (int i = 0; i < params.n_keep; i++) {
LOG_CNT("%s", common_token_to_piece(ctx, embd_inp[i]).c_str());
}
LOG_CNT("'\n");
}
LOG_INF("\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_INF("%s: interactive mode on.\n", __func__);
if (params.input_prefix_bos) {
LOG_INF("Input prefix with BOS\n");
}
if (!params.input_prefix.empty()) {
LOG_INF("Input prefix: '%s'\n", params.input_prefix.c_str());
}
if (!params.input_suffix.empty()) {
LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str());
}
}
smpl = common_sampler_init(model, sparams);
LOG_INF("sampler seed: %u\n", common_sampler_get_seed(smpl));
LOG_INF("sampler params: \n%s\n", sparams.print().c_str());
LOG_INF("sampler chain: %s\n", common_sampler_print(smpl).c_str());
LOG_INF("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_INF("\n");
LOG_INF("\n##### Infill mode #####\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_INF("== Running in interactive mode. ==\n");
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
LOG_INF( " - Press Ctrl+C to interject at any time.\n");
#endif
LOG_INF( "%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;
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);
LOG_WRN("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
console::set_display(console::reset);
}
// 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_DBG("\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_DBG("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_self_seq_rm (ctx, 0, params.n_keep + 1 , params.n_keep + n_discard + 1);
llama_kv_self_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
n_past -= n_discard;
LOG_DBG("after swap: n_past = %d\n", n_past);
LOG_DBG("embd: %s\n", string_from(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_DBG("eval: %s\n", string_from(ctx, embd).c_str());
if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
LOG_ERR("%s : failed to eval\n", __func__);
return 1;
}
n_past += n_eval;
LOG_DBG("n_past = %d\n", n_past);
}
}
embd.clear();
if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
const llama_token id = common_sampler_sample(smpl, ctx, -1);
common_sampler_accept(smpl, id, true);
// LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());
embd.push_back(id);
// echo this to console
input_echo = true;
// decrement remaining sampling budget
--n_remain;
LOG_DBG("n_remain: %d\n", n_remain);
} else {
// some user input remains from prompt or interaction, forward it to processing
LOG_DBG("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
common_sampler_accept(smpl, 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 = common_token_to_piece(ctx, id);
LOG("%s", token_str.c_str());
if (embd.size() > 1) {
input_tokens.push_back(id);
} else {
output_tokens.push_back(id);
output_ss << token_str;
}
}
}
// 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 ((common_sampler_last(smpl) == llama_vocab_eot(vocab) || is_interacting) && params.interactive){
if (is_interacting && !params.interactive_first) {
// print an eot token
LOG("%s", common_token_to_piece(ctx, llama_vocab_eot(vocab)).c_str());
}
LOG("\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 = common_tokenize(ctx, params.input_prefix, false);
std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
inp_pfx.insert(inp_pfx.begin(), llama_vocab_fim_pre(vocab));
inp_sfx.insert(inp_sfx.begin(), llama_vocab_fim_suf(vocab));
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_vocab_bos(vocab));
}
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;
is_interacting = false;
}
// deal with end of generation tokens in interactive mode
else if (llama_vocab_is_eog(vocab, common_sampler_last(smpl))) {
LOG_DBG("found EOS token\n");
if (params.interactive) {
is_interacting = true;
LOG("\n");
console::set_display(console::user_input);
}
}
if (n_past > 0 && is_interacting && !params.interactive) {
LOG_DBG("waiting for user input\n");
if (params.input_prefix_bos) {
LOG_DBG("adding input prefix BOS token\n");
embd_inp.push_back(llama_vocab_bos(vocab));
}
std::string buffer;
if (!params.input_prefix.empty()) {
LOG_DBG("appending input prefix: '%s'\n", params.input_prefix.c_str());
buffer += params.input_prefix;
LOG("%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_DBG("appending input suffix: '%s'\n", params.input_suffix.c_str());
buffer += params.input_suffix;
LOG("%s", params.input_suffix.c_str());
}
LOG_DBG("buffer: '%s'\n", buffer.c_str());
const size_t original_size = embd_inp.size();
const auto line_inp = common_tokenize(ctx, buffer, false);
LOG_DBG("input tokens: %s\n", string_from(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 << common_token_to_piece(ctx, token);
}
n_remain -= line_inp.size();
LOG_DBG("n_remain: %d\n", n_remain);
} else {
LOG_DBG("empty line, passing control back\n");
}
input_echo = false; // do not echo this again
}
if (n_past > 0) {
if (is_interacting) {
common_sampler_reset(smpl);
}
is_interacting = false;
}
}
// end of generation
if (!embd.empty() && llama_vocab_is_eog(vocab, 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) {
LOG("%s", common_token_to_piece(ctx, llama_vocab_eot(vocab)).c_str());
}
LOG("\n");
common_perf_print(ctx, smpl);
common_sampler_free(smpl);
llama_backend_free();
return 0;
}