vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into compilade/imatrix-batched-chunks

Browse source
This commit is contained in:
Francis Couture-Harpin 2025-02-09 12:06:15 -05:00
commit db502ddd0e
762 changed files with 149462 additions and 91773 deletions
Download patch file Download diff file Expand all files Collapse all files

563
common/common.h
View file

@ -2,22 +2,11 @@
#pragma once
#include "llama.h"
#include "llama-cpp.h"
#include "sampling.h"
#define LOG_NO_FILE_LINE_FUNCTION
#include "log.h"
#include <cmath>
#include <string>
#include <vector>
#include <random>
#include <thread>
#include <set>
#include <unordered_map>
#include <tuple>
#include <functional>
#include <sstream>
#ifdef _WIN32
#define DIRECTORY_SEPARATOR '\\'
@ -35,32 +24,41 @@
#define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
struct llama_lora_adapter_info {
struct common_adapter_lora_info {
std::string path;
float scale;
struct llama_adapter_lora * ptr;
};
struct llama_lora_adapter_container : llama_lora_adapter_info {
struct llama_lora_adapter * adapter;
};
using llama_tokens = std::vector<llama_token>;
// build info
extern int LLAMA_BUILD_NUMBER;
extern char const * LLAMA_COMMIT;
extern char const * LLAMA_COMPILER;
extern char const * LLAMA_BUILD_TARGET;
extern const char * LLAMA_COMMIT;
extern const char * LLAMA_COMPILER;
extern const char * LLAMA_BUILD_TARGET;
struct llama_control_vector_load_info;
struct common_control_vector_load_info;
//
// CPU utils
//
struct cpu_params {
int n_threads = -1;
bool cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask.
bool mask_valid = false; // Default: any CPU
enum ggml_sched_priority priority = GGML_SCHED_PRIO_NORMAL; // Scheduling prio : (0 - normal, 1 - medium, 2 - high, 3 - realtime)
bool strict_cpu = false; // Use strict CPU placement
uint32_t poll = 50; // Polling (busywait) level (0 - no polling, 100 - mostly polling)
};
int32_t cpu_get_num_physical_cores();
int32_t cpu_get_num_math();
//
// CLI argument parsing
// Common params
//
enum llama_example {
@ -78,42 +76,139 @@ enum llama_example {
LLAMA_EXAMPLE_CVECTOR_GENERATOR,
LLAMA_EXAMPLE_EXPORT_LORA,
LLAMA_EXAMPLE_LLAVA,
LLAMA_EXAMPLE_LOOKUP,
LLAMA_EXAMPLE_PARALLEL,
LLAMA_EXAMPLE_TTS,
LLAMA_EXAMPLE_COUNT,
};
enum common_sampler_type {
COMMON_SAMPLER_TYPE_NONE = 0,
COMMON_SAMPLER_TYPE_DRY = 1,
COMMON_SAMPLER_TYPE_TOP_K = 2,
COMMON_SAMPLER_TYPE_TOP_P = 3,
COMMON_SAMPLER_TYPE_MIN_P = 4,
//COMMON_SAMPLER_TYPE_TFS_Z = 5,
COMMON_SAMPLER_TYPE_TYPICAL_P = 6,
COMMON_SAMPLER_TYPE_TEMPERATURE = 7,
COMMON_SAMPLER_TYPE_XTC = 8,
COMMON_SAMPLER_TYPE_INFILL = 9,
COMMON_SAMPLER_TYPE_PENALTIES = 10,
};
// dimensionality reduction methods, used by cvector-generator
enum dimre_method {
DIMRE_METHOD_PCA,
DIMRE_METHOD_MEAN,
};
struct cpu_params {
int n_threads = -1;
bool cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask.
bool mask_valid = false; // Default: any CPU
enum ggml_sched_priority priority = GGML_SCHED_PRIO_NORMAL; // Scheduling prio : (0 - normal, 1 - medium, 2 - high, 3 - realtime)
bool strict_cpu = false; // Use strict CPU placement
uint32_t poll = 50; // Polling (busywait) level (0 - no polling, 100 - mostly polling)
enum common_conversation_mode {
COMMON_CONVERSATION_MODE_DISABLED = 0,
COMMON_CONVERSATION_MODE_ENABLED = 1,
COMMON_CONVERSATION_MODE_AUTO = 2,
};
struct gpt_params {
enum llama_example curr_ex = LLAMA_EXAMPLE_COMMON;
struct common_grammar_trigger {
std::string word;
bool at_start;
};
// sampling parameters
struct common_params_sampling {
uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
int32_t n_prev = 64; // number of previous tokens to remember
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
int32_t min_keep = 0; // 0 = disabled, otherwise samplers should return at least min_keep tokens
int32_t top_k = 40; // <= 0 to use vocab size
float top_p = 0.95f; // 1.0 = disabled
float min_p = 0.05f; // 0.0 = disabled
float xtc_probability = 0.00f; // 0.0 = disabled
float xtc_threshold = 0.10f; // > 0.5 disables XTC
float typ_p = 1.00f; // typical_p, 1.0 = disabled
float temp = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
float dynatemp_range = 0.00f; // 0.0 = disabled
float dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
int32_t penalty_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size)
float penalty_repeat = 1.00f; // 1.0 = disabled
float penalty_freq = 0.00f; // 0.0 = disabled
float penalty_present = 0.00f; // 0.0 = disabled
float dry_multiplier = 0.0f; // 0.0 = disabled; DRY repetition penalty for tokens extending repetition:
float dry_base = 1.75f; // 0.0 = disabled; multiplier * base ^ (length of sequence before token - allowed length)
int32_t dry_allowed_length = 2; // tokens extending repetitions beyond this receive penalty
int32_t dry_penalty_last_n = -1; // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
int32_t mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
float mirostat_tau = 5.00f; // target entropy
float mirostat_eta = 0.10f; // learning rate
bool ignore_eos = false;
bool no_perf = false; // disable performance metrics
bool timing_per_token = false;
std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"}; // default sequence breakers for DRY
std::vector<enum common_sampler_type> samplers = {
COMMON_SAMPLER_TYPE_PENALTIES,
COMMON_SAMPLER_TYPE_DRY,
COMMON_SAMPLER_TYPE_TOP_K,
COMMON_SAMPLER_TYPE_TYPICAL_P,
COMMON_SAMPLER_TYPE_TOP_P,
COMMON_SAMPLER_TYPE_MIN_P,
COMMON_SAMPLER_TYPE_XTC,
COMMON_SAMPLER_TYPE_TEMPERATURE,
};
std::string grammar; // optional BNF-like grammar to constrain sampling
bool grammar_lazy = false;
std::vector<common_grammar_trigger> grammar_trigger_words; // optional trigger words to trigger lazy grammar
std::vector<llama_token> grammar_trigger_tokens; // optional trigger tokens to trigger lazy grammar and print trigger special tokens.
std::vector<llama_logit_bias> logit_bias; // logit biases to apply
// print the parameters into a string
std::string print() const;
};
struct common_params_speculative {
std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
int32_t n_ctx = 0; // draft context size
int32_t n_max = 16; // maximum number of tokens to draft during speculative decoding
int32_t n_min = 5; // minimum number of draft tokens to use for speculative decoding
int32_t n_gpu_layers = -1; // number of layers to store in VRAM for the draft model (-1 - use default)
float p_split = 0.1f; // speculative decoding split probability
float p_min = 0.9f; // minimum speculative decoding probability (greedy)
struct cpu_params cpuparams;
struct cpu_params cpuparams_batch;
std::string hf_repo = ""; // HF repo // NOLINT
std::string hf_file = ""; // HF file // NOLINT
std::string model = ""; // draft model for speculative decoding // NOLINT
std::string model_url = ""; // model url to download // NOLINT
};
struct common_params_vocoder {
std::string hf_repo = ""; // HF repo // NOLINT
std::string hf_file = ""; // HF file // NOLINT
std::string model = ""; // model path // NOLINT
std::string model_url = ""; // model url to download // NOLINT
bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy // NOLINT
};
struct common_params {
int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 0; // context size
int32_t n_ctx = 4096; // context size
int32_t n_batch = 2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
int32_t n_ubatch = 512; // physical batch size for prompt processing (must be >=32 to use BLAS)
int32_t n_keep = 0; // number of tokens to keep from initial prompt
int32_t n_draft = 5; // number of tokens to draft during speculative decoding
int32_t n_chunks = -1; // max number of chunks to process (-1 = unlimited)
int32_t n_parallel = 1; // number of parallel sequences to decode
int32_t n_sequences = 1; // number of sequences to decode
float p_split = 0.1f; // speculative decoding split probability
int32_t n_gpu_layers = -1; // number of layers to store in VRAM (-1 - use default)
int32_t n_gpu_layers_draft = -1; // number of layers to store in VRAM for the draft model (-1 - use default)
int32_t main_gpu = 0; // the GPU that is used for scratch and small tensors
float tensor_split[128] = {0}; // how split tensors should be distributed across GPUs
int32_t grp_attn_n = 1; // group-attention factor
int32_t grp_attn_w = 512; // group-attention width
int32_t n_print = -1; // print token count every n tokens (-1 = disabled)
@ -124,51 +219,56 @@ struct gpt_params {
float yarn_beta_fast = 32.0f; // YaRN low correction dim
float yarn_beta_slow = 1.0f; // YaRN high correction dim
int32_t yarn_orig_ctx = 0; // YaRN original context length
float defrag_thold = -1.0f; // KV cache defragmentation threshold
float defrag_thold = 0.1f; // KV cache defragmentation threshold
// offload params
std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
int32_t n_gpu_layers = -1; // number of layers to store in VRAM (-1 - use default)
int32_t main_gpu = 0; // the GPU that is used for scratch and small tensors
float tensor_split[128] = {0}; // how split tensors should be distributed across GPUs
enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
struct cpu_params cpuparams;
struct cpu_params cpuparams_batch;
struct cpu_params draft_cpuparams;
struct cpu_params draft_cpuparams_batch;
ggml_backend_sched_eval_callback cb_eval = nullptr;
void * cb_eval_user_data = nullptr;
ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
enum llama_attention_type attention_type = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
struct gpt_sampler_params sparams;
struct common_params_sampling sampling;
struct common_params_speculative speculative;
struct common_params_vocoder vocoder;
std::string model = ""; // model path
std::string model_draft = ""; // draft model for speculative decoding
std::string model_alias = "unknown"; // model alias
std::string model_url = ""; // model url to download
std::string hf_token = ""; // HF token
std::string hf_repo = ""; // HF repo
std::string hf_file = ""; // HF file
std::string prompt = "";
std::string prompt_file = ""; // store the external prompt file name
std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state
std::string input_prefix = ""; // string to prefix user inputs with
std::string input_suffix = ""; // string to suffix user inputs with
std::string logdir = ""; // directory in which to save YAML log files
std::string lookup_cache_static = ""; // path of static ngram cache file for lookup decoding
std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding
std::string logits_file = ""; // file for saving *all* logits
std::string rpc_servers = ""; // comma separated list of RPC servers
std::string model = ""; // model path // NOLINT
std::string model_alias = ""; // model alias // NOLINT
std::string model_url = ""; // model url to download // NOLINT
std::string hf_token = ""; // HF token // NOLINT
std::string hf_repo = ""; // HF repo // NOLINT
std::string hf_file = ""; // HF file // NOLINT
std::string prompt = ""; // NOLINT
std::string prompt_file = ""; // store the external prompt file name // NOLINT
std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state // NOLINT
std::string input_prefix = ""; // string to prefix user inputs with // NOLINT
std::string input_suffix = ""; // string to suffix user inputs with // NOLINT
std::string lookup_cache_static = ""; // path of static ngram cache file for lookup decoding // NOLINT
std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding // NOLINT
std::string logits_file = ""; // file for saving *all* logits // NOLINT
std::vector<std::string> in_files; // all input files
std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
std::vector<llama_model_kv_override> kv_overrides;
bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
std::vector<llama_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_adapter_lora_apply)
std::vector<common_adapter_lora_info> lora_adapters; // lora adapter path with user defined scale
std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
std::vector<common_control_vector_load_info> control_vectors; // control vector with user defined scale
int32_t verbosity = 0;
int32_t control_vector_layer_start = -1; // layer range for control vector
@ -189,13 +289,11 @@ struct gpt_params {
bool kl_divergence = false; // compute KL divergence
std::function<void(int, char **)> print_usage = nullptr; // print example-specific usage and example
bool usage = false; // print usage
bool use_color = false; // use color to distinguish generations and inputs
bool special = false; // enable special token output
bool interactive = false; // interactive mode
bool interactive_first = false; // wait for user input immediately
bool conversation = false; // conversation mode (does not print special tokens and suffix/prefix)
bool prompt_cache_all = false; // save user input and generations to prompt cache
bool prompt_cache_ro = false; // open the prompt cache read-only and do not update it
@ -204,6 +302,8 @@ struct gpt_params {
bool simple_io = false; // improves compatibility with subprocesses and limited consoles
bool cont_batching = true; // insert new sequences for decoding on-the-fly
bool flash_attn = false; // flash attention
bool no_perf = false; // disable performance metrics
bool ctx_shift = true; // context shift on inifinite text generation
bool input_prefix_bos = false; // prefix BOS to user inputs, preceding input_prefix
bool logits_all = false; // return logits for all tokens in the batch
@ -211,43 +311,49 @@ struct gpt_params {
bool use_mlock = false; // use mlock to keep model in memory
bool verbose_prompt = false; // print prompt tokens before generation
bool display_prompt = true; // print prompt before generation
bool infill = false; // use infill mode
bool dump_kv_cache = false; // dump the KV cache contents for debugging purposes
bool no_kv_offload = false; // disable KV offloading
bool warmup = true; // warmup run
bool check_tensors = false; // validate tensor data
std::string cache_type_k = "f16"; // KV cache data type for the K
std::string cache_type_v = "f16"; // KV cache data type for the V
ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;
// multimodal models (see examples/llava)
std::string mmproj = ""; // path to multimodal projector
std::string mmproj = ""; // path to multimodal projector // NOLINT
std::vector<std::string> image; // path to image file(s)
// embedding
bool embedding = false; // get only sentence embedding
int32_t embd_normalize = 2; // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
int32_t embd_normalize = 2; // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
std::string embd_out = ""; // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
std::string embd_sep = "\n"; // separator of embendings
std::string embd_sep = "\n"; // separator of embeddings
bool reranking = false; // enable reranking support on server
// server params
int32_t port = 8080; // server listens on this network port
int32_t timeout_read = 600; // http read timeout in seconds
int32_t timeout_write = timeout_read; // http write timeout in seconds
int n_threads_http = -1; // number of threads to process HTTP requests (TODO: support threadpool)
int32_t n_threads_http = -1; // number of threads to process HTTP requests (TODO: support threadpool)
int32_t n_cache_reuse = 0; // min chunk size to reuse from the cache via KV shifting
std::string hostname = "127.0.0.1";
std::string public_path = "";
std::string chat_template = "";
std::string system_prompt = "";
std::string public_path = ""; // NOLINT
std::string chat_template = ""; // NOLINT
bool use_jinja = false; // NOLINT
bool enable_chat_template = true;
std::vector<std::string> api_keys;
std::string ssl_file_key = "";
std::string ssl_file_cert = "";
std::string ssl_file_key = ""; // NOLINT
std::string ssl_file_cert = ""; // NOLINT
bool endpoint_slots = true;
// "advanced" endpoints are disabled by default for better security
bool webui = true;
bool endpoint_slots = false;
bool endpoint_props = false; // only control POST requests, not GET
bool endpoint_metrics = false;
bool log_json = false;
@ -300,112 +406,46 @@ struct gpt_params {
bool batched_bench_output_jsonl = false;
};
struct llama_arg {
std::set<enum llama_example> examples = {LLAMA_EXAMPLE_COMMON};
std::vector<const char *> args;
const char * value_hint = nullptr; // help text or example for arg value
const char * value_hint_2 = nullptr; // for second arg value
const char * env = nullptr;
std::string help;
void (*handler_void) (gpt_params & params) = nullptr;
void (*handler_string) (gpt_params & params, const std::string &) = nullptr;
void (*handler_str_str)(gpt_params & params, const std::string &, const std::string &) = nullptr;
void (*handler_int) (gpt_params & params, int) = nullptr;
// call once at the start of a program if it uses libcommon
// initializes the logging system and prints info about the build
void common_init();
llama_arg(
const std::initializer_list<const char *> & args,
const char * value_hint,
const std::string & help,
void (*handler)(gpt_params & params, const std::string &)
) : args(args), value_hint(value_hint), help(help), handler_string(handler) {}
std::string common_params_get_system_info(const common_params & params);
llama_arg(
const std::initializer_list<const char *> & args,
const char * value_hint,
const std::string & help,
void (*handler)(gpt_params & params, int)
) : args(args), value_hint(value_hint), help(help), handler_int(handler) {}
llama_arg(
const std::initializer_list<const char *> & args,
const std::string & help,
void (*handler)(gpt_params & params)
) : args(args), help(help), handler_void(handler) {}
// support 2 values for arg
llama_arg(
const std::initializer_list<const char *> & args,
const char * value_hint,
const char * value_hint_2,
const std::string & help,
void (*handler)(gpt_params & params, const std::string &, const std::string &)
) : args(args), value_hint(value_hint), value_hint_2(value_hint_2), help(help), handler_str_str(handler) {}
llama_arg & set_examples(std::initializer_list<enum llama_example> examples) {
this->examples = std::move(examples);
return *this;
}
llama_arg & set_env(const char * env) {
help = help + "\n(env: " + env + ")";
this->env = env;
return *this;
}
bool in_example(enum llama_example ex) {
return examples.find(ex) != examples.end();
}
bool get_value_from_env(std::string & output) const {
if (env == nullptr) return false;
char * value = std::getenv(env);
if (value) {
output = value;
return true;
}
return false;
}
bool has_value_from_env() const {
return env != nullptr && std::getenv(env);
}
std::string to_string();
};
// initialize list of options (arguments) that can be used by the current example
std::vector<llama_arg> gpt_params_parser_init(gpt_params & params, llama_example ex);
// optionally, we can provide "print_usage" to print example usage
std::vector<llama_arg> gpt_params_parser_init(gpt_params & params, llama_example ex, std::function<void(int, char **)> print_usage);
// parse input arguments from CLI
// if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message)
bool gpt_params_parse (int argc, char ** argv, gpt_params & params, std::vector<llama_arg> & options);
bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params, std::vector<llama_arg> & options);
// print full usage message; it will be called internally by gpt_params_parse() if "-h" is set
void gpt_params_print_usage(gpt_params & params, std::vector<llama_arg> & options);
std::string gpt_params_get_system_info(const gpt_params & params);
bool parse_cpu_range(const std::string& range, bool(&boolmask)[GGML_MAX_N_THREADS]);
bool parse_cpu_mask(const std::string& mask, bool(&boolmask)[GGML_MAX_N_THREADS]);
void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model = nullptr);
bool parse_cpu_range(const std::string & range, bool(&boolmask)[GGML_MAX_N_THREADS]);
bool parse_cpu_mask(const std::string & mask, bool(&boolmask)[GGML_MAX_N_THREADS]);
void postprocess_cpu_params(cpu_params & cpuparams, const cpu_params * role_model = nullptr);
bool set_process_priority(enum ggml_sched_priority prio);
//
// String utils
//
std::vector<std::string> string_split(std::string input, char separator);
#ifdef __GNUC__
#ifdef __MINGW32__
#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
#else
#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
#endif
#else
#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...)
#endif
LLAMA_COMMON_ATTRIBUTE_FORMAT(1, 2)
std::string string_format(const char * fmt, ...);
std::string string_strip(const std::string & str);
std::string string_get_sortable_timestamp();
std::string string_join(const std::vector<std::string> & values, const std::string & separator);
std::vector<std::string> string_split(const std::string & str, const std::string & delimiter);
std::string string_repeat(const std::string & str, size_t n);
void string_replace_all(std::string & s, const std::string & search, const std::string & replace);
template<class T>
static std::vector<T> string_split(const std::string & str, char delim) {
static_assert(!std::is_same<T, std::string>::value, "Please use the specialized version for std::string");
std::vector<T> values;
std::istringstream str_stream(str);
std::string token;
@ -418,9 +458,40 @@ static std::vector<T> string_split(const std::string & str, char delim) {
return values;
}
template<>
std::vector<std::string> string_split<std::string>(const std::string & input, char separator)
{
std::vector<std::string> parts;
size_t begin_pos = 0;
size_t separator_pos = input.find(separator);
while (separator_pos != std::string::npos) {
std::string part = input.substr(begin_pos, separator_pos - begin_pos);
parts.emplace_back(part);
begin_pos = separator_pos + 1;
separator_pos = input.find(separator, begin_pos);
}
parts.emplace_back(input.substr(begin_pos, separator_pos - begin_pos));
return parts;
}
static bool string_starts_with(const std::string & str,
const std::string & prefix) { // While we wait for C++20's std::string::starts_with...
return str.rfind(prefix, 0) == 0;
}
static bool string_ends_with(const std::string & str,
const std::string & suffix) { // While we wait for C++20's std::string::ends_with...
return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
}
bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
void string_process_escapes(std::string & input);
std::string string_from(bool value);
std::string string_from(const std::vector<int> & values);
std::string string_from(const struct llama_context * ctx, const std::vector<llama_token> & tokens);
std::string string_from(const struct llama_context * ctx, const struct llama_batch & batch);
//
// Filesystem utils
//
@ -435,65 +506,103 @@ std::string fs_get_cache_file(const std::string & filename);
// Model utils
//
struct llama_init_result {
struct llama_model * model = nullptr;
struct llama_context * context = nullptr;
std::vector<llama_lora_adapter_container> lora_adapters;
// note: defines object's lifetime
struct common_init_result {
llama_model_ptr model;
llama_context_ptr context;
std::vector<llama_adapter_lora_ptr> lora;
};
struct llama_init_result llama_init_from_gpt_params(gpt_params & params);
struct common_init_result common_init_from_params(common_params & params);
struct llama_model_params llama_model_params_from_gpt_params (const gpt_params & params);
struct llama_context_params llama_context_params_from_gpt_params (const gpt_params & params);
struct llama_model_params common_model_params_to_llama ( common_params & params);
struct llama_context_params common_context_params_to_llama(const common_params & params);
struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params);
struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);
struct llama_model * common_load_model_from_url(
const std::string & model_url,
const std::string & local_path,
const std::string & hf_token,
const struct llama_model_params & params);
struct llama_model * common_load_model_from_hf(
const std::string & repo,
const std::string & remote_path,
const std::string & local_path,
const std::string & hf_token,
const struct llama_model_params & params);
std::pair<std::string, std::string> common_get_hf_file(
const std::string & hf_repo_with_tag,
const std::string & hf_token);
// clear LoRA adapters from context, then apply new list of adapters
void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters);
void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
//
// Batch utils
//
void llama_batch_clear(struct llama_batch & batch);
void common_batch_clear(struct llama_batch & batch);
void llama_batch_add(
void common_batch_add(
struct llama_batch & batch,
llama_token id,
llama_pos pos,
const std::vector<llama_seq_id> & seq_ids,
bool logits);
//
// Token utils
//
// longest common prefix
size_t common_lcp(const llama_tokens & a, const llama_tokens & b);
// longet common subsequence
size_t common_lcs(const llama_tokens & a, const llama_tokens & b);
//
// Vocab utils
//
// tokenizes a string into a vector of tokens
// should work similar to Python's `tokenizer.encode`
std::vector<llama_token> llama_tokenize(
std::vector<llama_token> common_tokenize(
const struct llama_context * ctx,
const std::string & text,
bool add_special,
bool parse_special = false);
std::vector<llama_token> llama_tokenize(
const struct llama_model * model,
std::vector<llama_token> common_tokenize(
const struct llama_vocab * vocab,
const std::string & text,
bool add_special,
bool parse_special = false);
// tokenizes a token into a piece, optionally renders special/control tokens
// should work similar to Python's `tokenizer.id_to_piece`
std::string llama_token_to_piece(
std::string common_token_to_piece(
const struct llama_context * ctx,
llama_token token,
bool special = true);
std::string common_token_to_piece(
const struct llama_vocab * vocab,
llama_token token,
bool special = true);
// detokenizes a vector of tokens into a string
// should work similar to Python's `tokenizer.decode`
// optionally renders special/control tokens
std::string llama_detokenize(
llama_context * ctx,
std::string common_detokenize(
const struct llama_context * ctx,
const std::vector<llama_token> & tokens,
bool special = true);
std::string common_detokenize(
const struct llama_vocab * vocab,
const std::vector<llama_token> & tokens,
bool special = true);
@ -501,64 +610,88 @@ std::string llama_detokenize(
// Chat template utils
//
struct common_tool_call {
std::string name;
std::string arguments;
std::string id;
};
// same with llama_chat_message, but uses std::string
struct llama_chat_msg {
struct common_chat_msg {
std::string role;
std::string content;
std::vector<common_tool_call> tool_calls;
};
// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
bool llama_chat_verify_template(const std::string & tmpl);
bool common_chat_verify_template(const std::string & tmpl, bool use_jinja);
namespace minja {
class chat_template;
}
typedef minja::chat_template common_chat_template;
struct common_chat_templates {
bool has_explicit_template; // Model had builtin template or template overridde was specified.
std::unique_ptr<common_chat_template> template_default; // always set (defaults to chatml)
std::unique_ptr<common_chat_template> template_tool_use;
};
// CPP wrapper for llama_chat_apply_template
// If the built-in template is not supported, we default to chatml
// If the custom "tmpl" is not supported, we throw an error
std::string llama_chat_apply_template(const struct llama_model * model,
const std::string & tmpl,
const std::vector<llama_chat_msg> & chat,
bool add_ass);
std::string common_chat_apply_template(
const common_chat_template & tmpl,
const std::vector<common_chat_msg> & chat,
bool add_ass,
bool use_jinja);
// Format single message, while taking into account the position of that message in chat history
std::string llama_chat_format_single(const struct llama_model * model,
const std::string & tmpl,
const std::vector<llama_chat_msg> & past_msg,
const llama_chat_msg & new_msg,
bool add_ass);
std::string common_chat_format_single(
const common_chat_template & tmpl,
const std::vector<common_chat_msg> & past_msg,
const common_chat_msg & new_msg,
bool add_ass,
bool use_jinja);
// Returns an example of formatted chat
std::string llama_chat_format_example(const struct llama_model * model,
const std::string & tmpl);
std::string common_chat_format_example(
const common_chat_template & tmpl, bool use_jinja);
common_chat_templates common_chat_templates_from_model(const struct llama_model * model, const std::string & chat_template_override);
//
// KV cache utils
//
// Dump the KV cache view with the number of sequences per cell.
void llama_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
void common_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
// Dump the KV cache view showing individual sequences in each cell (long output).
void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
//
// Embedding utils
//
void llama_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2);
// TODO: repace embd_norm with an enum
void common_embd_normalize(const float * inp, float * out, int n, int embd_norm);
float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n);
float common_embd_similarity_cos(const float * embd1, const float * embd2, int n);
//
// Control vector utils
//
struct llama_control_vector_data {
struct common_control_vector_data {
int n_embd;
// stores data for layers [1, n_layer] where n_layer = data.size() / n_embd
std::vector<float> data;
};
struct llama_control_vector_load_info {
struct common_control_vector_load_info {
float strength;
std::string fname;
@ -566,24 +699,16 @@ struct llama_control_vector_load_info {
// Load control vectors, scale each by strength, and add them together.
// On error, returns {-1, empty}
llama_control_vector_data llama_control_vector_load(const std::vector<llama_control_vector_load_info> & load_infos);
common_control_vector_data common_control_vector_load(const std::vector<common_control_vector_load_info> & load_infos);
//
// Split utils
//
static const char * const LLM_KV_SPLIT_NO = "split.no";
static const char * const LLM_KV_SPLIT_COUNT = "split.count";
static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
namespace {
//
// YAML utils
//
const char * const LLM_KV_SPLIT_NO = "split.no";
const char * const LLM_KV_SPLIT_COUNT = "split.count";
const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
void yaml_dump_vector_float (FILE * stream, const char * prop_name, const std::vector<float> & data);
void yaml_dump_vector_int (FILE * stream, const char * prop_name, const std::vector<int> & data);
void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data);
void yaml_dump_non_result_info(
FILE * stream, const gpt_params & params, const llama_context * lctx,
const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
}