root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-27 03:44:35 +00:00
Code Issues Actions 4 Packages Projects Releases Wiki Activity

gguf : use UNIX line ending

Browse Source
This commit is contained in:
Georgi Gerganov 2023-08-14 13:04:35 +03:00
parent 0c19ae70d5
commit 62490f1380
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
5 changed files with 5606 additions and 5606 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

108
constants.py
View File

@ -1,54 +1,54 @@
GGUF_MAGIC = 0x47475546
GGUF_VERSION = 1
GGUF_DEFAULT_ALIGNMENT = 32
# general
KEY_GENERAL_ARCHITECTURE = "general.architecture"
KEY_GENERAL_QUANTIZATION_VERSION = "general.quantization_version"
KEY_GENERAL_ALIGNMENT = "general.alignment"
KEY_GENERAL_NAME = "general.name"
KEY_GENERAL_AUTHOR = "general.author"
KEY_GENERAL_URL = "general.url"
KEY_GENERAL_DESCRIPTION = "general.description"
KEY_GENERAL_FILE_TYPE = "general.file_type"
KEY_GENERAL_LICENSE = "general.license"
KEY_GENERAL_SOURCE_URL = "general.source.url"
KEY_GENERAL_SOURCE_HF_REPO = "general.source.hugginface.repository"
# LLM
KEY_LLM_CONTEXT_LENGTH = "{llm}.context_length"
KEY_LLM_EMBEDDING_LENGTH = "{llm}.embedding_length"
KEY_LLM_BLOCK_COUNT = "{llm}.block_count"
KEY_LLM_FEED_FORWARD_LENGTH = "{llm}.feed_forward_length"
KEY_LLM_USE_PARALLEL_RESIDUAL = "{llm}.use_parallel_residual"
KEY_LLM_TENSOR_DATA_LAYOUT = "{llm}.tensor_data_layout"
# attention
KEY_ATTENTION_HEAD_COUNT = "{llm}.attention.head_count"
KEY_ATTENTION_HEAD_COUNT_KV = "{llm}.attention.head_count_kv"
KEY_ATTENTION_MAX_ALIBI_BIAS = "{llm}.attention.max_alibi_bias"
KEY_ATTENTION_CLAMP_KQV = "{llm}.attention.clamp_kqv"
KEY_ATTENTION_LAYERNORM_EPS = "{llm}.attention.layer_norm_epsilon"
KEY_ATTENTION_LAYERNORM_RMS_EPS = "{llm}.attention.layer_norm_rms_epsilon"
# RoPE
KEY_ROPE_DIMENSION_COUNT = "{llm}.rope.dimension_count"
KEY_ROPE_SCALE = "{llm}.rope.scale"
# tokenization
KEY_TOKENIZER_MODEL = "tokenizer.ggml.model"
KEY_TOKENIZER_LIST = "tokenizer.ggml.tokens"
KEY_TOKENIZER_SCORES = "tokenizer.ggml.scores"
KEY_TOKENIZER_MERGES = "tokenizer.ggml.merges"
KEY_TOKENIZER_BOS_ID = "tokenizer.ggml.bos_token_id"
KEY_TOKENIZER_EOS_ID = "tokenizer.ggml.eos_token_id"
KEY_TOKENIZER_UNK_ID = "tokenizer.ggml.unknown_token_id"
KEY_TOKENIZER_SEP_ID = "tokenizer.ggml.seperator_token_id"
KEY_TOKENIZER_PAD_ID = "tokenizer.ggml.padding_token_id"
KEY_TOKENIZER_HF_JSON = "tokenizer.huggingface.json"
KEY_TOKENIZER_RWKV = "tokenizer.rwkv.world"
KEY_TOKENIZER_BOS_ID = "tokenizer.ggml.bos_token_id"
KEY_TOKENIZER_EOS_ID = "tokenizer.ggml.eos_token_id"
KEY_TOKENIZER_UNK_ID = "tokenizer.ggml.unknown_token_id"
KEY_TOKENIZER_SEP_ID = "tokenizer.ggml.separator_token_id"
KEY_TOKENIZER_PAD_ID = "tokenizer.ggml.padding_token_id"
GGUF_MAGIC = 0x47475546
GGUF_VERSION = 1
GGUF_DEFAULT_ALIGNMENT = 32
# general
KEY_GENERAL_ARCHITECTURE = "general.architecture"
KEY_GENERAL_QUANTIZATION_VERSION = "general.quantization_version"
KEY_GENERAL_ALIGNMENT = "general.alignment"
KEY_GENERAL_NAME = "general.name"
KEY_GENERAL_AUTHOR = "general.author"
KEY_GENERAL_URL = "general.url"
KEY_GENERAL_DESCRIPTION = "general.description"
KEY_GENERAL_FILE_TYPE = "general.file_type"
KEY_GENERAL_LICENSE = "general.license"
KEY_GENERAL_SOURCE_URL = "general.source.url"
KEY_GENERAL_SOURCE_HF_REPO = "general.source.hugginface.repository"
# LLM
KEY_LLM_CONTEXT_LENGTH = "{llm}.context_length"
KEY_LLM_EMBEDDING_LENGTH = "{llm}.embedding_length"
KEY_LLM_BLOCK_COUNT = "{llm}.block_count"
KEY_LLM_FEED_FORWARD_LENGTH = "{llm}.feed_forward_length"
KEY_LLM_USE_PARALLEL_RESIDUAL = "{llm}.use_parallel_residual"
KEY_LLM_TENSOR_DATA_LAYOUT = "{llm}.tensor_data_layout"
# attention
KEY_ATTENTION_HEAD_COUNT = "{llm}.attention.head_count"
KEY_ATTENTION_HEAD_COUNT_KV = "{llm}.attention.head_count_kv"
KEY_ATTENTION_MAX_ALIBI_BIAS = "{llm}.attention.max_alibi_bias"
KEY_ATTENTION_CLAMP_KQV = "{llm}.attention.clamp_kqv"
KEY_ATTENTION_LAYERNORM_EPS = "{llm}.attention.layer_norm_epsilon"
KEY_ATTENTION_LAYERNORM_RMS_EPS = "{llm}.attention.layer_norm_rms_epsilon"
# RoPE
KEY_ROPE_DIMENSION_COUNT = "{llm}.rope.dimension_count"
KEY_ROPE_SCALE = "{llm}.rope.scale"
# tokenization
KEY_TOKENIZER_MODEL = "tokenizer.ggml.model"
KEY_TOKENIZER_LIST = "tokenizer.ggml.tokens"
KEY_TOKENIZER_SCORES = "tokenizer.ggml.scores"
KEY_TOKENIZER_MERGES = "tokenizer.ggml.merges"
KEY_TOKENIZER_BOS_ID = "tokenizer.ggml.bos_token_id"
KEY_TOKENIZER_EOS_ID = "tokenizer.ggml.eos_token_id"
KEY_TOKENIZER_UNK_ID = "tokenizer.ggml.unknown_token_id"
KEY_TOKENIZER_SEP_ID = "tokenizer.ggml.seperator_token_id"
KEY_TOKENIZER_PAD_ID = "tokenizer.ggml.padding_token_id"
KEY_TOKENIZER_HF_JSON = "tokenizer.huggingface.json"
KEY_TOKENIZER_RWKV = "tokenizer.rwkv.world"
KEY_TOKENIZER_BOS_ID = "tokenizer.ggml.bos_token_id"
KEY_TOKENIZER_EOS_ID = "tokenizer.ggml.eos_token_id"
KEY_TOKENIZER_UNK_ID = "tokenizer.ggml.unknown_token_id"
KEY_TOKENIZER_SEP_ID = "tokenizer.ggml.separator_token_id"
KEY_TOKENIZER_PAD_ID = "tokenizer.ggml.padding_token_id"

8394
gguf-llama.cpp
View File

File diff suppressed because it is too large Load Diff

898
gguf-llama.h
View File

@ -1,449 +1,449 @@
#ifndef LLAMA_H
#define LLAMA_H
#include "ggml.h"
#ifdef GGML_USE_CUBLAS
#include "ggml-cuda.h"
#define LLAMA_MAX_DEVICES GGML_CUDA_MAX_DEVICES
#else
#define LLAMA_MAX_DEVICES 1
#endif // GGML_USE_CUBLAS
#include <stddef.h>
#include <stdint.h>
#include <stdbool.h>
#ifdef LLAMA_SHARED
# if defined(_WIN32) && !defined(__MINGW32__)
# ifdef LLAMA_BUILD
# define LLAMA_API __declspec(dllexport)
# else
# define LLAMA_API __declspec(dllimport)
# endif
# else
# define LLAMA_API __attribute__ ((visibility ("default")))
# endif
#else
# define LLAMA_API
#endif
#ifdef __GNUC__
# define DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
#elif defined(_MSC_VER)
# define DEPRECATED(func, hint) __declspec(deprecated(hint)) func
#else
# define DEPRECATED(func, hint) func
#endif
#define LLAMA_DEFAULT_SEED 0xFFFFFFFF
#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
// Defined when llama.cpp is compiled with support for offloading model layers to GPU.
#define LLAMA_SUPPORTS_GPU_OFFLOAD
#endif
#ifndef LLAMA_DEFAULT_RMS_EPS
#define LLAMA_DEFAULT_RMS_EPS 5e-6f
#endif
#ifdef __cplusplus
extern "C" {
#endif
//
// C interface
//
// TODO: show sample usage
//
struct llama_model;
struct llama_context;
typedef int llama_token;
typedef struct llama_token_data {
llama_token id; // token id
float logit; // log-odds of the token
float p; // probability of the token
} llama_token_data;
typedef struct llama_token_data_array {
llama_token_data * data;
size_t size;
bool sorted;
} llama_token_data_array;
typedef void (*llama_progress_callback)(float progress, void *ctx);
struct llama_context_params {
uint32_t seed; // RNG seed, -1 for random
int32_t n_ctx; // text context
int32_t n_batch; // prompt processing batch size
int32_t n_gqa; // grouped-query attention (TEMP - will be moved to model hparams)
float rms_norm_eps; // rms norm epsilon (TEMP - will be moved to model hparams)
int32_t n_gpu_layers; // number of layers to store in VRAM
int32_t main_gpu; // the GPU that is used for scratch and small tensors
const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
// ref: https://github.com/ggerganov/llama.cpp/pull/2054
float rope_freq_base; // RoPE base frequency
float rope_freq_scale; // RoPE frequency scaling factor
// called with a progress value between 0 and 1, pass NULL to disable
llama_progress_callback progress_callback;
// context pointer passed to the progress callback
void * progress_callback_user_data;
// Keep the booleans together to avoid misalignment during copy-by-value.
bool low_vram; // if true, reduce VRAM usage at the cost of performance
bool f16_kv; // use fp16 for KV cache
bool logits_all; // the llama_eval() call computes all logits, not just the last one
bool vocab_only; // only load the vocabulary, no weights
bool use_mmap; // use mmap if possible
bool use_mlock; // force system to keep model in RAM
bool embedding; // embedding mode only
};
// model file types
enum llama_ftype {
LLAMA_FTYPE_ALL_F32 = 0,
LLAMA_FTYPE_MOSTLY_F16 = 1, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
// LLAMA_FTYPE_MOSTLY_Q4_2 = 5, // support has been removed
// LLAMA_FTYPE_MOSTLY_Q4_3 = 6, // support has been removed
LLAMA_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q2_K = 10,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q3_K_S = 11,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q3_K_M = 12,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q3_K_L = 13,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_K_S = 14,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_K_M = 15,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_K_S = 16,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_K_M = 17,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q6_K = 18,// except 1d tensors
};
// model quantization parameters
typedef struct llama_model_quantize_params {
int nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
enum llama_ftype ftype; // quantize to this llama_ftype
bool allow_requantize; // allow quantizing non-f32/f16 tensors
bool quantize_output_tensor; // quantize output.weight
} llama_model_quantize_params;
// grammar types
struct llama_grammar;
// grammar element type
enum llama_gretype {
// end of rule definition
LLAMA_GRETYPE_END = 0,
// start of alternate definition for rule
LLAMA_GRETYPE_ALT = 1,
// non-terminal element: reference to rule
LLAMA_GRETYPE_RULE_REF = 2,
// terminal element: character (code point)
LLAMA_GRETYPE_CHAR = 3,
// inverse char(s) ([^a], [^a-b] [^abc])
LLAMA_GRETYPE_CHAR_NOT = 4,
// modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to
// be an inclusive range ([a-z])
LLAMA_GRETYPE_CHAR_RNG_UPPER = 5,
// modifies a preceding LLAMA_GRETYPE_CHAR or
// LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
LLAMA_GRETYPE_CHAR_ALT = 6,
};
typedef struct llama_grammar_element {
enum llama_gretype type;
uint32_t value; // Unicode code point or rule ID
} llama_grammar_element;
// performance timing information
struct llama_timings {
double t_start_ms;
double t_end_ms;
double t_load_ms;
double t_sample_ms;
double t_p_eval_ms;
double t_eval_ms;
int32_t n_sample;
int32_t n_p_eval;
int32_t n_eval;
};
LLAMA_API int llama_max_devices();
LLAMA_API struct llama_context_params llama_context_default_params();
LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params();
LLAMA_API bool llama_mmap_supported();
LLAMA_API bool llama_mlock_supported();
// TODO: not great API - very likely to change
// Initialize the llama + ggml backend
// If numa is true, use NUMA optimizations
// Call once at the start of the program
LLAMA_API void llama_backend_init(bool numa);
// Call once at the end of the program - currently only used for MPI
LLAMA_API void llama_backend_free();
LLAMA_API int64_t llama_time_us();
LLAMA_API struct llama_model * llama_load_model_from_file(
const char * path_model,
struct llama_context_params params);
LLAMA_API void llama_free_model(struct llama_model * model);
LLAMA_API struct llama_context * llama_new_context_with_model(
struct llama_model * model,
struct llama_context_params params);
// Frees all allocated memory
LLAMA_API void llama_free(struct llama_context * ctx);
// Returns 0 on success
LLAMA_API int llama_model_quantize(
const char * fname_inp,
const char * fname_out,
const llama_model_quantize_params * params);
// Apply a LoRA adapter to a loaded model
// path_base_model is the path to a higher quality model to use as a base for
// the layers modified by the adapter. Can be NULL to use the current loaded model.
// The model needs to be reloaded before applying a new adapter, otherwise the adapter
// will be applied on top of the previous one
// Returns 0 on success
LLAMA_API DEPRECATED(int llama_apply_lora_from_file(
struct llama_context * ctx,
const char * path_lora,
const char * path_base_model,
int n_threads),
"please use llama_model_apply_lora_from_file instead");
LLAMA_API int llama_model_apply_lora_from_file(
const struct llama_model * model,
const char * path_lora,
const char * path_base_model,
int n_threads);
// Returns the number of tokens in the KV cache
LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx);
// Sets the current rng seed.
LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
// Returns the maximum size in bytes of the state (rng, logits, embedding
// and kv_cache) - will often be smaller after compacting tokens
LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx);
// Copies the state to the specified destination address.
// Destination needs to have allocated enough memory.
// Returns the number of bytes copied
LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst);
// Set the state reading from the specified address
// Returns the number of bytes read
LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src);
// Save/load session file
LLAMA_API bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out);
LLAMA_API bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count);
// Run the llama inference to obtain the logits and probabilities for the next token.
// tokens + n_tokens is the provided batch of new tokens to process
// n_past is the number of tokens to use from previous eval calls
// Returns 0 on success
LLAMA_API int llama_eval(
struct llama_context * ctx,
const llama_token * tokens,
int n_tokens,
int n_past,
int n_threads);
// Same as llama_eval, but use float matrix input directly.
LLAMA_API int llama_eval_embd(
struct llama_context * ctx,
const float * embd,
int n_tokens,
int n_past,
int n_threads);
// Export a static computation graph for context of 511 and batch size of 1
// NOTE: since this functionality is mostly for debugging and demonstration purposes, we hardcode these
// parameters here to keep things simple
// IMPORTANT: do not use for anything else other than debugging and testing!
LLAMA_API int llama_eval_export(struct llama_context * ctx, const char * fname);
// Convert the provided text into tokens.
// The tokens pointer must be large enough to hold the resulting tokens.
// Returns the number of tokens on success, no more than n_max_tokens
// Returns a negative number on failure - the number of tokens that would have been returned
// TODO: not sure if correct
LLAMA_API int llama_tokenize(
struct llama_context * ctx,
const char * text,
llama_token * tokens,
int n_max_tokens,
bool add_bos);
LLAMA_API int llama_tokenize_with_model(
const struct llama_model * model,
const char * text,
llama_token * tokens,
int n_max_tokens,
bool add_bos);
LLAMA_API int llama_n_vocab(const struct llama_context * ctx);
LLAMA_API int llama_n_ctx (const struct llama_context * ctx);
LLAMA_API int llama_n_embd (const struct llama_context * ctx);
LLAMA_API int llama_n_vocab_from_model(const struct llama_model * model);
LLAMA_API int llama_n_ctx_from_model (const struct llama_model * model);
LLAMA_API int llama_n_embd_from_model (const struct llama_model * model);
// Get the vocabulary as output parameters.
// Returns number of results.
LLAMA_API int llama_get_vocab(
const struct llama_context * ctx,
const char * * strings,
float * scores,
int capacity);
LLAMA_API int llama_get_vocab_from_model(
const struct llama_model * model,
const char * * strings,
float * scores,
int capacity);
// Token logits obtained from the last call to llama_eval()
// The logits for the last token are stored in the last row
// Can be mutated in order to change the probabilities of the next token
// Rows: n_tokens
// Cols: n_vocab
LLAMA_API float * llama_get_logits(struct llama_context * ctx);
// Get the embeddings for the input
// shape: [n_embd] (1-dimensional)
LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
// Token Id -> String. Uses the vocabulary in the provided context
LLAMA_API const char * llama_token_to_str(
const struct llama_context * ctx,
llama_token token);
LLAMA_API const char * llama_token_to_str_with_model(
const struct llama_model * model,
llama_token token);
// Special tokens
LLAMA_API llama_token llama_token_bos(); // beginning-of-sentence
LLAMA_API llama_token llama_token_eos(); // end-of-sentence
LLAMA_API llama_token llama_token_nl(); // next-line
// Grammar
//
LLAMA_API struct llama_grammar * llama_grammar_init(
const llama_grammar_element ** rules,
size_t n_rules,
size_t start_rule_index);
LLAMA_API void llama_grammar_free(struct llama_grammar * grammar);
// Sampling functions
/// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
LLAMA_API void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float penalty);
/// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
LLAMA_API void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float alpha_frequency, float alpha_presence);
/// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806
/// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.
/// @params guidance_ctx A separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
/// @params scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
LLAMA_API void llama_sample_classifier_free_guidance(
struct llama_context * ctx,
llama_token_data_array * candidates,
struct llama_context * guidance_ctx,
float scale);
/// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
/// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep);
/// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
/// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep);
/// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
/// @details Apply constraints from grammar
LLAMA_API void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * candidates, const struct llama_grammar * grammar);
/// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
/// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
/// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
/// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
/// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
/// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
LLAMA_API llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu);
/// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
/// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
/// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
/// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
/// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
LLAMA_API llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu);
/// @details Selects the token with the highest probability.
LLAMA_API llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates);
/// @details Randomly selects a token from the candidates based on their probabilities.
LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
/// @details Accepts the sampled token into the grammar
LLAMA_API void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token);
// Performance information
LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
LLAMA_API void llama_print_timings(struct llama_context * ctx);
LLAMA_API void llama_reset_timings(struct llama_context * ctx);
// Print system information
LLAMA_API const char * llama_print_system_info(void);
#ifdef __cplusplus
}
#endif
// Internal API to be implemented by llama.cpp and used by tests/benchmarks only
#ifdef LLAMA_API_INTERNAL
#include <vector>
#include <string>
struct ggml_tensor;
const std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
#endif
#endif // LLAMA_H
#ifndef LLAMA_H
#define LLAMA_H
#include "ggml.h"
#ifdef GGML_USE_CUBLAS
#include "ggml-cuda.h"
#define LLAMA_MAX_DEVICES GGML_CUDA_MAX_DEVICES
#else
#define LLAMA_MAX_DEVICES 1
#endif // GGML_USE_CUBLAS
#include <stddef.h>
#include <stdint.h>
#include <stdbool.h>
#ifdef LLAMA_SHARED
# if defined(_WIN32) && !defined(__MINGW32__)
# ifdef LLAMA_BUILD
# define LLAMA_API __declspec(dllexport)
# else
# define LLAMA_API __declspec(dllimport)
# endif
# else
# define LLAMA_API __attribute__ ((visibility ("default")))
# endif
#else
# define LLAMA_API
#endif
#ifdef __GNUC__
# define DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
#elif defined(_MSC_VER)
# define DEPRECATED(func, hint) __declspec(deprecated(hint)) func
#else
# define DEPRECATED(func, hint) func
#endif
#define LLAMA_DEFAULT_SEED 0xFFFFFFFF
#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
// Defined when llama.cpp is compiled with support for offloading model layers to GPU.
#define LLAMA_SUPPORTS_GPU_OFFLOAD
#endif
#ifndef LLAMA_DEFAULT_RMS_EPS
#define LLAMA_DEFAULT_RMS_EPS 5e-6f
#endif
#ifdef __cplusplus
extern "C" {
#endif
//
// C interface
//
// TODO: show sample usage
//
struct llama_model;
struct llama_context;
typedef int llama_token;
typedef struct llama_token_data {
llama_token id; // token id
float logit; // log-odds of the token
float p; // probability of the token
} llama_token_data;
typedef struct llama_token_data_array {
llama_token_data * data;
size_t size;
bool sorted;
} llama_token_data_array;
typedef void (*llama_progress_callback)(float progress, void *ctx);
struct llama_context_params {
uint32_t seed; // RNG seed, -1 for random
int32_t n_ctx; // text context
int32_t n_batch; // prompt processing batch size
int32_t n_gqa; // grouped-query attention (TEMP - will be moved to model hparams)
float rms_norm_eps; // rms norm epsilon (TEMP - will be moved to model hparams)
int32_t n_gpu_layers; // number of layers to store in VRAM
int32_t main_gpu; // the GPU that is used for scratch and small tensors
const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
// ref: https://github.com/ggerganov/llama.cpp/pull/2054
float rope_freq_base; // RoPE base frequency
float rope_freq_scale; // RoPE frequency scaling factor
// called with a progress value between 0 and 1, pass NULL to disable
llama_progress_callback progress_callback;
// context pointer passed to the progress callback
void * progress_callback_user_data;
// Keep the booleans together to avoid misalignment during copy-by-value.
bool low_vram; // if true, reduce VRAM usage at the cost of performance
bool f16_kv; // use fp16 for KV cache
bool logits_all; // the llama_eval() call computes all logits, not just the last one
bool vocab_only; // only load the vocabulary, no weights
bool use_mmap; // use mmap if possible
bool use_mlock; // force system to keep model in RAM
bool embedding; // embedding mode only
};
// model file types
enum llama_ftype {
LLAMA_FTYPE_ALL_F32 = 0,
LLAMA_FTYPE_MOSTLY_F16 = 1, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
// LLAMA_FTYPE_MOSTLY_Q4_2 = 5, // support has been removed
// LLAMA_FTYPE_MOSTLY_Q4_3 = 6, // support has been removed
LLAMA_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q2_K = 10,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q3_K_S = 11,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q3_K_M = 12,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q3_K_L = 13,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_K_S = 14,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_K_M = 15,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_K_S = 16,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_K_M = 17,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q6_K = 18,// except 1d tensors
};
// model quantization parameters
typedef struct llama_model_quantize_params {
int nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
enum llama_ftype ftype; // quantize to this llama_ftype
bool allow_requantize; // allow quantizing non-f32/f16 tensors
bool quantize_output_tensor; // quantize output.weight
} llama_model_quantize_params;
// grammar types
struct llama_grammar;
// grammar element type
enum llama_gretype {
// end of rule definition
LLAMA_GRETYPE_END = 0,
// start of alternate definition for rule
LLAMA_GRETYPE_ALT = 1,
// non-terminal element: reference to rule
LLAMA_GRETYPE_RULE_REF = 2,
// terminal element: character (code point)
LLAMA_GRETYPE_CHAR = 3,
// inverse char(s) ([^a], [^a-b] [^abc])
LLAMA_GRETYPE_CHAR_NOT = 4,
// modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to
// be an inclusive range ([a-z])
LLAMA_GRETYPE_CHAR_RNG_UPPER = 5,
// modifies a preceding LLAMA_GRETYPE_CHAR or
// LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
LLAMA_GRETYPE_CHAR_ALT = 6,
};
typedef struct llama_grammar_element {
enum llama_gretype type;
uint32_t value; // Unicode code point or rule ID
} llama_grammar_element;
// performance timing information
struct llama_timings {
double t_start_ms;
double t_end_ms;
double t_load_ms;
double t_sample_ms;
double t_p_eval_ms;
double t_eval_ms;
int32_t n_sample;
int32_t n_p_eval;
int32_t n_eval;
};
LLAMA_API int llama_max_devices();
LLAMA_API struct llama_context_params llama_context_default_params();
LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params();
LLAMA_API bool llama_mmap_supported();
LLAMA_API bool llama_mlock_supported();
// TODO: not great API - very likely to change
// Initialize the llama + ggml backend
// If numa is true, use NUMA optimizations
// Call once at the start of the program
LLAMA_API void llama_backend_init(bool numa);
// Call once at the end of the program - currently only used for MPI
LLAMA_API void llama_backend_free();
LLAMA_API int64_t llama_time_us();
LLAMA_API struct llama_model * llama_load_model_from_file(
const char * path_model,
struct llama_context_params params);
LLAMA_API void llama_free_model(struct llama_model * model);
LLAMA_API struct llama_context * llama_new_context_with_model(
struct llama_model * model,
struct llama_context_params params);
// Frees all allocated memory
LLAMA_API void llama_free(struct llama_context * ctx);
// Returns 0 on success
LLAMA_API int llama_model_quantize(
const char * fname_inp,
const char * fname_out,
const llama_model_quantize_params * params);
// Apply a LoRA adapter to a loaded model
// path_base_model is the path to a higher quality model to use as a base for
// the layers modified by the adapter. Can be NULL to use the current loaded model.
// The model needs to be reloaded before applying a new adapter, otherwise the adapter
// will be applied on top of the previous one
// Returns 0 on success
LLAMA_API DEPRECATED(int llama_apply_lora_from_file(
struct llama_context * ctx,
const char * path_lora,
const char * path_base_model,
int n_threads),
"please use llama_model_apply_lora_from_file instead");
LLAMA_API int llama_model_apply_lora_from_file(
const struct llama_model * model,
const char * path_lora,
const char * path_base_model,
int n_threads);
// Returns the number of tokens in the KV cache
LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx);
// Sets the current rng seed.
LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
// Returns the maximum size in bytes of the state (rng, logits, embedding
// and kv_cache) - will often be smaller after compacting tokens
LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx);
// Copies the state to the specified destination address.
// Destination needs to have allocated enough memory.
// Returns the number of bytes copied
LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst);
// Set the state reading from the specified address
// Returns the number of bytes read
LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src);
// Save/load session file
LLAMA_API bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out);
LLAMA_API bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count);
// Run the llama inference to obtain the logits and probabilities for the next token.
// tokens + n_tokens is the provided batch of new tokens to process
// n_past is the number of tokens to use from previous eval calls
// Returns 0 on success
LLAMA_API int llama_eval(
struct llama_context * ctx,
const llama_token * tokens,
int n_tokens,
int n_past,
int n_threads);
// Same as llama_eval, but use float matrix input directly.
LLAMA_API int llama_eval_embd(
struct llama_context * ctx,
const float * embd,
int n_tokens,
int n_past,
int n_threads);
// Export a static computation graph for context of 511 and batch size of 1
// NOTE: since this functionality is mostly for debugging and demonstration purposes, we hardcode these
// parameters here to keep things simple
// IMPORTANT: do not use for anything else other than debugging and testing!
LLAMA_API int llama_eval_export(struct llama_context * ctx, const char * fname);
// Convert the provided text into tokens.
// The tokens pointer must be large enough to hold the resulting tokens.
// Returns the number of tokens on success, no more than n_max_tokens
// Returns a negative number on failure - the number of tokens that would have been returned
// TODO: not sure if correct
LLAMA_API int llama_tokenize(
struct llama_context * ctx,
const char * text,
llama_token * tokens,
int n_max_tokens,
bool add_bos);
LLAMA_API int llama_tokenize_with_model(
const struct llama_model * model,
const char * text,
llama_token * tokens,
int n_max_tokens,
bool add_bos);
LLAMA_API int llama_n_vocab(const struct llama_context * ctx);
LLAMA_API int llama_n_ctx (const struct llama_context * ctx);
LLAMA_API int llama_n_embd (const struct llama_context * ctx);
LLAMA_API int llama_n_vocab_from_model(const struct llama_model * model);
LLAMA_API int llama_n_ctx_from_model (const struct llama_model * model);
LLAMA_API int llama_n_embd_from_model (const struct llama_model * model);
// Get the vocabulary as output parameters.
// Returns number of results.
LLAMA_API int llama_get_vocab(
const struct llama_context * ctx,
const char * * strings,
float * scores,
int capacity);
LLAMA_API int llama_get_vocab_from_model(
const struct llama_model * model,
const char * * strings,
float * scores,
int capacity);
// Token logits obtained from the last call to llama_eval()
// The logits for the last token are stored in the last row
// Can be mutated in order to change the probabilities of the next token
// Rows: n_tokens
// Cols: n_vocab
LLAMA_API float * llama_get_logits(struct llama_context * ctx);
// Get the embeddings for the input
// shape: [n_embd] (1-dimensional)
LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
// Token Id -> String. Uses the vocabulary in the provided context
LLAMA_API const char * llama_token_to_str(
const struct llama_context * ctx,
llama_token token);
LLAMA_API const char * llama_token_to_str_with_model(
const struct llama_model * model,
llama_token token);
// Special tokens
LLAMA_API llama_token llama_token_bos(); // beginning-of-sentence
LLAMA_API llama_token llama_token_eos(); // end-of-sentence
LLAMA_API llama_token llama_token_nl(); // next-line
// Grammar
//
LLAMA_API struct llama_grammar * llama_grammar_init(
const llama_grammar_element ** rules,
size_t n_rules,
size_t start_rule_index);
LLAMA_API void llama_grammar_free(struct llama_grammar * grammar);
// Sampling functions
/// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
LLAMA_API void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float penalty);
/// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
LLAMA_API void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float alpha_frequency, float alpha_presence);
/// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806
/// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.
/// @params guidance_ctx A separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
/// @params scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
LLAMA_API void llama_sample_classifier_free_guidance(
struct llama_context * ctx,
llama_token_data_array * candidates,
struct llama_context * guidance_ctx,
float scale);
/// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
/// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep);
/// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
/// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep);
/// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
/// @details Apply constraints from grammar
LLAMA_API void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * candidates, const struct llama_grammar * grammar);
/// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
/// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
/// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
/// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
/// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
/// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
LLAMA_API llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu);
/// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
/// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
/// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
/// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
/// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
LLAMA_API llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu);
/// @details Selects the token with the highest probability.
LLAMA_API llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates);
/// @details Randomly selects a token from the candidates based on their probabilities.
LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
/// @details Accepts the sampled token into the grammar
LLAMA_API void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token);
// Performance information
LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
LLAMA_API void llama_print_timings(struct llama_context * ctx);
LLAMA_API void llama_reset_timings(struct llama_context * ctx);
// Print system information
LLAMA_API const char * llama_print_system_info(void);
#ifdef __cplusplus
}
#endif
// Internal API to be implemented by llama.cpp and used by tests/benchmarks only
#ifdef LLAMA_API_INTERNAL
#include <vector>
#include <string>
struct ggml_tensor;
const std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
#endif
#endif // LLAMA_H

1134
gguf-util.h
View File

File diff suppressed because it is too large Load Diff

678
gguf.py
View File

@ -1,339 +1,339 @@
"""TODOs
1. Implement writers for known architectures, LLaMA in particular.
2. Add docstrings from the format specs.
3. After development is done, Convert it to a proper pip-installable Python package, and possibly move it to its own repo under ggml-org.
"""
import struct
import constants
from enum import IntEnum
from typing import Any, IO, List
import numpy as np
import sys
class GGMLQuantizationType(IntEnum):
F32 = 0
F16 = 1
class GGUFValueType(IntEnum):
UINT8 = 0
INT8 = 1
UINT16 = 2
INT16 = 3
UINT32 = 4
INT32 = 5
FLOAT32 = 6
BOOL = 7
STRING = 8
ARRAY = 9
@staticmethod
def get_type(val):
if isinstance(val, str) or isinstance(val, bytes) or isinstance(val, bytearray):
return GGUFValueType.STRING
elif isinstance(val, list):
return GGUFValueType.ARRAY
elif isinstance(val, float):
return GGUFValueType.FLOAT32
elif isinstance(val, bool):
return GGUFValueType.BOOL
elif isinstance(val, int):
return GGUFValueType.INT32
else:
print("Unknown type: "+str(type(val)))
sys.exit()
class GGUFWriter:
def __init__(self, fout: IO):
self.fout = fout
self.offset_tensor = 0
self.data_alignment = constants.GGUF_DEFAULT_ALIGNMENT
self.kv_data = b""
self.kv_data_count = 0
self.ti_data = b""
self.ti_data_count = 0
def write_header_to_file(self):
self.fout.write(struct.pack("<I", constants.GGUF_MAGIC))
self.fout.write(struct.pack("<I", constants.GGUF_VERSION))
self.fout.write(struct.pack("<I", self.ti_data_count))
self.fout.write(struct.pack("<I", self.kv_data_count))
self.flush()
# print("tensors " + str(self.ti_data_count) + " kv " + str(self.kv_data_count))
def write_kv_data_to_file(self):
self.fout.write(self.kv_data)
self.flush()
def write_ti_data_to_file(self):
self.fout.write(self.ti_data)
self.flush()
@classmethod
def open(cls, path: str) -> "GGUFWriter":
f = open(path, "wb")
return cls(f)
def add_key(self, key: str):
self.add_val(key, GGUFValueType.STRING, add_vtype=False)
def add_uint8(self, key: str, val: int):
self.add_key(key)
self.add_val(val, GGUFValueType.UINT8)
def add_int8(self, key: str, val: int):
self.add_key(key)
self.add_val(val, GGUFValueType.INT8)
def add_uint16(self, key: str, val: int):
self.add_key(key)
self.add_val(val, GGUFValueType.UINT16)
def add_int16(self, key: str, val: int):
self.add_key(key)
self.add_val(val, GGUFValueType.INT16)
def add_uint32(self, key: str, val: int):
self.add_key(key)
self.add_val(val, GGUFValueType.UINT32)
def add_int32(self, key: str, val: int):
self.add_key(key)
self.add_val(val, GGUFValueType.INT32)
def add_float32(self, key: str, val: float):
self.add_key(key)
self.add_val(val, GGUFValueType.FLOAT32)
def add_bool(self, key: str, val: bool):
self.add_key(key)
self.add_val(val, GGUFValueType.BOOL)
def add_string(self, key: str, val: str):
if len(val) == 0: return
self.add_key(key)
self.add_val(val, GGUFValueType.STRING)
def add_array(self, key: str, val: list):
if not isinstance(val, list):
raise ValueError("Value must be a list for array type")
self.add_key(key)
self.add_val(val, GGUFValueType.ARRAY)
def add_val(self: str, val: Any, vtype: GGUFValueType = None, add_vtype: bool = True):
if vtype is None:
vtype = GGUFValueType.get_type(val)
if add_vtype:
self.kv_data += struct.pack("<I", vtype)
self.kv_data_count += 1
if vtype == GGUFValueType.UINT8:
self.kv_data += struct.pack("<B", val)
elif vtype == GGUFValueType.INT8:
self.kv_data += struct.pack("<b", val)
elif vtype == GGUFValueType.UINT16:
self.kv_data += struct.pack("<H", val)
elif vtype == GGUFValueType.INT16:
self.kv_data += struct.pack("<h", val)
elif vtype == GGUFValueType.UINT32:
self.kv_data += struct.pack("<I", val)
elif vtype == GGUFValueType.INT32:
self.kv_data += struct.pack("<i", val)
elif vtype == GGUFValueType.FLOAT32:
self.kv_data += struct.pack("<f", val)
elif vtype == GGUFValueType.BOOL:
self.kv_data += struct.pack("?", val)
elif vtype == GGUFValueType.STRING:
encoded_val = val.encode("utf8") if isinstance(val, str) else val
self.kv_data += struct.pack("<I", len(encoded_val))
self.kv_data += encoded_val
elif vtype == GGUFValueType.ARRAY:
ltype = set([GGUFValueType.get_type(item) for item in val])
assert len(ltype) == 1, "All items in a GGUF array should be of the same type"
self.kv_data += struct.pack("<I", list(ltype)[0])
self.kv_data += struct.pack("<I", len(val))
for item in val:
self.add_val(item, add_vtype=False)
else:
raise ValueError("Invalid GGUF metadata value type")
@staticmethod
def ggml_pad(x: int, n: int) -> int:
return ((x + n - 1) // n) * n
def add_tensor_info(self, name: str, tensor_shape: np.ndarray, tensor_dtype: np.dtype, tensor_nbytes: int):
encoded_name = name.encode("utf8")
self.ti_data += struct.pack("<I", len(encoded_name))
self.ti_data += encoded_name
n_dims = len(tensor_shape)
self.ti_data += struct.pack("<I", n_dims)
for i in range(n_dims):
self.ti_data += struct.pack("<I", tensor_shape[n_dims - 1 - i])
assert tensor_dtype in (np.float32, np.float16), "Only F32 and F16 tensors are supported for now"
dtype = GGMLQuantizationType.F32 if tensor_dtype == np.float32 else GGMLQuantizationType.F16
self.ti_data += struct.pack("<I", dtype)
self.ti_data += struct.pack("<Q", self.offset_tensor)
self.offset_tensor += GGUFWriter.ggml_pad(tensor_nbytes, self.data_alignment)
self.ti_data_count += 1
def write_tensor_to_file(self, tensor: np.ndarray):
pad = GGUFWriter.ggml_pad(self.fout.tell(), self.data_alignment) - self.fout.tell()
if pad != 0:
self.fout.write(bytes([0] * pad))
tensor.tofile(self.fout)
pad = GGUFWriter.ggml_pad(tensor.nbytes, self.data_alignment) - tensor.nbytes
if pad != 0:
self.fout.write(bytes([0] * pad))
def flush(self):
self.fout.flush()
def close(self):
self.fout.close()
def add_architecture(self, architecture: str):
self.add_string(constants.KEY_GENERAL_ARCHITECTURE,
architecture)
def add_author(self, author: str):
self.add_string(constants.KEY_GENERAL_AUTHOR, author)
def add_url(self, url: str):
self.add_string(constants.KEY_GENERAL_URL, url)
def add_description(self, description: str):
self.add_string(constants.KEY_GENERAL_DESCRIPTION, description)
def add_file_type(self, file_type: str):
self.add_string(constants.KEY_GENERAL_FILE_TYPE, file_type)
def add_source_url(self, url: str):
self.add_string(constants.KEY_GENERAL_SOURCE_URL, url)
def add_source_hf_repo(self, repo: str):
self.add_string(constants.KEY_GENERAL_SOURCE_HF_REPO, repo)
def add_name(self, name: str):
self.add_string(constants.KEY_GENERAL_NAME, name)
def add_quantization_version(self, quantization_version: GGMLQuantizationType):
self.add_uint32(
constants.KEY_GENERAL_QUANTIZATION_VERSION, quantization_version)
def add_custom_alignment(self, alignment: int):
self.data_alignment = alignment
self.add_uint32(constants.KEY_GENERAL_ALIGNMENT, alignment)
def add_context_length(self, llm: str, length: int):
self.add_uint32(
constants.KEY_LLM_CONTEXT_LENGTH.format(llm=llm), length)
def add_embedding_length(self, llm: str, length: int):
self.add_uint32(
constants.KEY_LLM_EMBEDDING_LENGTH.format(llm=llm), length)
def add_block_count(self, llm: str, length: int):
self.add_uint32(
constants.KEY_LLM_BLOCK_COUNT.format(llm=llm), length)
def add_feed_forward_length(self, llm: str, length: int):
self.add_uint32(
constants.KEY_LLM_FEED_FORWARD_LENGTH.format(llm=llm), length)
def add_parallel_residual(self, llm: str, use: bool):
self.add_bool(
constants.KEY_LLM_USE_PARALLEL_RESIDUAL.format(llm=llm), use)
def add_tensor_data_layout(self, llm: str, layout: str):
self.add_string(
constants.KEY_LLM_TENSOR_DATA_LAYOUT.format(llm=llm), layout)
def add_head_count(self, llm: str, count: int):
self.add_uint32(
constants.KEY_ATTENTION_HEAD_COUNT.format(llm=llm), count)
def add_head_count_kv(self, llm: str, count: int):
self.add_uint32(
constants.KEY_ATTENTION_HEAD_COUNT_KV.format(llm=llm), count)
def add_max_alibi_bias(self, llm: str, bias: float):
self.add_float32(
constants.KEY_ATTENTION_MAX_ALIBI_BIAS.format(llm=llm), bias)
def add_clamp_kqv(self, llm: str, value: float):
self.add_float32(
constants.KEY_ATTENTION_CLAMP_KQV.format(llm=llm), value)
def add_layer_norm_eps(self, llm: str, value: float):
self.add_float32(
constants.KEY_ATTENTION_LAYERNORM_EPS.format(llm=llm), value)
def add_layer_norm_rms_eps(self, llm: str, value: float):
self.add_float32(
constants.KEY_ATTENTION_LAYERNORM_RMS_EPS.format(llm=llm), value)
def add_rope_dimension_count(self, llm: str, count: int):
self.add_uint32(
constants.KEY_ROPE_DIMENSION_COUNT.format(llm=llm), count)
def add_rope_scale(self, llm: str, value: float):
self.add_float32(constants.KEY_ROPE_SCALE.format(llm=llm), value)
def add_tokenizer_model(self, model: str):
self.add_string(constants.KEY_TOKENIZER_MODEL, model)
def add_token_list(self, tokens: List):
self.add_array(constants.KEY_TOKENIZER_LIST, tokens)
def add_token_merges(self, merges: List):
self.add_array(constants.KEY_TOKENIZER_MERGES, merges)
def add_token_scores(self, scores: List[float]):
self.add_array(constants.KEY_TOKENIZER_SCORES, scores)
def add_bos_token_id(self, id: int):
self.add_uint32(constants.KEY_TOKENIZER_BOS_ID, id)
def add_eos_token_id(self, id: int):
self.add_uint32(constants.KEY_TOKENIZER_EOS_ID, id)
def add_unk_token_id(self, id: int):
self.add_uint32(constants.KEY_TOKENIZER_UNK_ID, id)
def add_sep_token_id(self, id: int):
self.add_uint32(constants.KEY_TOKENIZER_SEP_ID, id)
def add_pad_token_id(self, id: int):
self.add_uint32(constants.KEY_TOKENIZER_PAD_ID, id)
# Example usage:
if __name__ == "__main__":
# Example usage with a file
gguf_writer = GGUFWriter.open("example.gguf")
gguf_writer.add_architecture("llama")
gguf_writer.add_uint32("answer", 42) # Write a 32-bit integer
gguf_writer.add_float32("answer_in_float", 42.0) # Write a 32-bit float
gguf_writer.add_custom_alignment(64)
tensor1 = np.ones((32,), dtype=np.float32) * 100.0
tensor2 = np.ones((32,), dtype=np.float32) * 101.0
gguf_writer.add_tensor_info("tensor0", tensor1)
gguf_writer.add_tensor_info("tensor1", tensor2)
gguf_writer.write_header_to_file()
gguf_writer.write_kv_data_to_file()
gguf_writer.write_ti_data_to_file()
gguf_writer.write_tensor_to_file(tensor1)
gguf_writer.write_tensor_to_file(tensor2)
gguf_writer.close()
"""TODOs
1. Implement writers for known architectures, LLaMA in particular.
2. Add docstrings from the format specs.
3. After development is done, Convert it to a proper pip-installable Python package, and possibly move it to its own repo under ggml-org.
"""
import struct
import constants
from enum import IntEnum
from typing import Any, IO, List
import numpy as np
import sys
class GGMLQuantizationType(IntEnum):
F32 = 0
F16 = 1
class GGUFValueType(IntEnum):
UINT8 = 0
INT8 = 1
UINT16 = 2
INT16 = 3
UINT32 = 4
INT32 = 5
FLOAT32 = 6
BOOL = 7
STRING = 8
ARRAY = 9
@staticmethod
def get_type(val):
if isinstance(val, str) or isinstance(val, bytes) or isinstance(val, bytearray):
return GGUFValueType.STRING
elif isinstance(val, list):
return GGUFValueType.ARRAY
elif isinstance(val, float):
return GGUFValueType.FLOAT32
elif isinstance(val, bool):
return GGUFValueType.BOOL
elif isinstance(val, int):
return GGUFValueType.INT32
else:
print("Unknown type: "+str(type(val)))
sys.exit()
class GGUFWriter:
def __init__(self, fout: IO):
self.fout = fout
self.offset_tensor = 0
self.data_alignment = constants.GGUF_DEFAULT_ALIGNMENT
self.kv_data = b""
self.kv_data_count = 0
self.ti_data = b""
self.ti_data_count = 0
def write_header_to_file(self):
self.fout.write(struct.pack("<I", constants.GGUF_MAGIC))
self.fout.write(struct.pack("<I", constants.GGUF_VERSION))
self.fout.write(struct.pack("<I", self.ti_data_count))
self.fout.write(struct.pack("<I", self.kv_data_count))
self.flush()
# print("tensors " + str(self.ti_data_count) + " kv " + str(self.kv_data_count))
def write_kv_data_to_file(self):
self.fout.write(self.kv_data)
self.flush()
def write_ti_data_to_file(self):
self.fout.write(self.ti_data)
self.flush()
@classmethod
def open(cls, path: str) -> "GGUFWriter":
f = open(path, "wb")
return cls(f)
def add_key(self, key: str):
self.add_val(key, GGUFValueType.STRING, add_vtype=False)
def add_uint8(self, key: str, val: int):
self.add_key(key)
self.add_val(val, GGUFValueType.UINT8)
def add_int8(self, key: str, val: int):
self.add_key(key)
self.add_val(val, GGUFValueType.INT8)
def add_uint16(self, key: str, val: int):
self.add_key(key)
self.add_val(val, GGUFValueType.UINT16)
def add_int16(self, key: str, val: int):
self.add_key(key)
self.add_val(val, GGUFValueType.INT16)
def add_uint32(self, key: str, val: int):
self.add_key(key)
self.add_val(val, GGUFValueType.UINT32)
def add_int32(self, key: str, val: int):
self.add_key(key)
self.add_val(val, GGUFValueType.INT32)
def add_float32(self, key: str, val: float):
self.add_key(key)
self.add_val(val, GGUFValueType.FLOAT32)
def add_bool(self, key: str, val: bool):
self.add_key(key)
self.add_val(val, GGUFValueType.BOOL)
def add_string(self, key: str, val: str):
if len(val) == 0: return
self.add_key(key)
self.add_val(val, GGUFValueType.STRING)
def add_array(self, key: str, val: list):
if not isinstance(val, list):
raise ValueError("Value must be a list for array type")
self.add_key(key)
self.add_val(val, GGUFValueType.ARRAY)
def add_val(self: str, val: Any, vtype: GGUFValueType = None, add_vtype: bool = True):
if vtype is None:
vtype = GGUFValueType.get_type(val)
if add_vtype:
self.kv_data += struct.pack("<I", vtype)
self.kv_data_count += 1
if vtype == GGUFValueType.UINT8:
self.kv_data += struct.pack("<B", val)
elif vtype == GGUFValueType.INT8:
self.kv_data += struct.pack("<b", val)
elif vtype == GGUFValueType.UINT16:
self.kv_data += struct.pack("<H", val)
elif vtype == GGUFValueType.INT16:
self.kv_data += struct.pack("<h", val)
elif vtype == GGUFValueType.UINT32:
self.kv_data += struct.pack("<I", val)
elif vtype == GGUFValueType.INT32:
self.kv_data += struct.pack("<i", val)
elif vtype == GGUFValueType.FLOAT32:
self.kv_data += struct.pack("<f", val)
elif vtype == GGUFValueType.BOOL:
self.kv_data += struct.pack("?", val)
elif vtype == GGUFValueType.STRING:
encoded_val = val.encode("utf8") if isinstance(val, str) else val
self.kv_data += struct.pack("<I", len(encoded_val))
self.kv_data += encoded_val
elif vtype == GGUFValueType.ARRAY:
ltype = set([GGUFValueType.get_type(item) for item in val])
assert len(ltype) == 1, "All items in a GGUF array should be of the same type"
self.kv_data += struct.pack("<I", list(ltype)[0])
self.kv_data += struct.pack("<I", len(val))
for item in val:
self.add_val(item, add_vtype=False)
else:
raise ValueError("Invalid GGUF metadata value type")
@staticmethod
def ggml_pad(x: int, n: int) -> int:
return ((x + n - 1) // n) * n
def add_tensor_info(self, name: str, tensor_shape: np.ndarray, tensor_dtype: np.dtype, tensor_nbytes: int):
encoded_name = name.encode("utf8")
self.ti_data += struct.pack("<I", len(encoded_name))
self.ti_data += encoded_name
n_dims = len(tensor_shape)
self.ti_data += struct.pack("<I", n_dims)
for i in range(n_dims):
self.ti_data += struct.pack("<I", tensor_shape[n_dims - 1 - i])
assert tensor_dtype in (np.float32, np.float16), "Only F32 and F16 tensors are supported for now"
dtype = GGMLQuantizationType.F32 if tensor_dtype == np.float32 else GGMLQuantizationType.F16
self.ti_data += struct.pack("<I", dtype)
self.ti_data += struct.pack("<Q", self.offset_tensor)
self.offset_tensor += GGUFWriter.ggml_pad(tensor_nbytes, self.data_alignment)
self.ti_data_count += 1
def write_tensor_to_file(self, tensor: np.ndarray):
pad = GGUFWriter.ggml_pad(self.fout.tell(), self.data_alignment) - self.fout.tell()
if pad != 0:
self.fout.write(bytes([0] * pad))
tensor.tofile(self.fout)
pad = GGUFWriter.ggml_pad(tensor.nbytes, self.data_alignment) - tensor.nbytes
if pad != 0:
self.fout.write(bytes([0] * pad))
def flush(self):
self.fout.flush()
def close(self):
self.fout.close()
def add_architecture(self, architecture: str):
self.add_string(constants.KEY_GENERAL_ARCHITECTURE,
architecture)
def add_author(self, author: str):
self.add_string(constants.KEY_GENERAL_AUTHOR, author)
def add_url(self, url: str):
self.add_string(constants.KEY_GENERAL_URL, url)
def add_description(self, description: str):
self.add_string(constants.KEY_GENERAL_DESCRIPTION, description)
def add_file_type(self, file_type: str):
self.add_string(constants.KEY_GENERAL_FILE_TYPE, file_type)
def add_source_url(self, url: str):
self.add_string(constants.KEY_GENERAL_SOURCE_URL, url)
def add_source_hf_repo(self, repo: str):
self.add_string(constants.KEY_GENERAL_SOURCE_HF_REPO, repo)
def add_name(self, name: str):
self.add_string(constants.KEY_GENERAL_NAME, name)
def add_quantization_version(self, quantization_version: GGMLQuantizationType):
self.add_uint32(
constants.KEY_GENERAL_QUANTIZATION_VERSION, quantization_version)
def add_custom_alignment(self, alignment: int):
self.data_alignment = alignment
self.add_uint32(constants.KEY_GENERAL_ALIGNMENT, alignment)
def add_context_length(self, llm: str, length: int):
self.add_uint32(
constants.KEY_LLM_CONTEXT_LENGTH.format(llm=llm), length)
def add_embedding_length(self, llm: str, length: int):
self.add_uint32(
constants.KEY_LLM_EMBEDDING_LENGTH.format(llm=llm), length)
def add_block_count(self, llm: str, length: int):
self.add_uint32(
constants.KEY_LLM_BLOCK_COUNT.format(llm=llm), length)
def add_feed_forward_length(self, llm: str, length: int):
self.add_uint32(
constants.KEY_LLM_FEED_FORWARD_LENGTH.format(llm=llm), length)
def add_parallel_residual(self, llm: str, use: bool):
self.add_bool(
constants.KEY_LLM_USE_PARALLEL_RESIDUAL.format(llm=llm), use)
def add_tensor_data_layout(self, llm: str, layout: str):
self.add_string(
constants.KEY_LLM_TENSOR_DATA_LAYOUT.format(llm=llm), layout)
def add_head_count(self, llm: str, count: int):
self.add_uint32(
constants.KEY_ATTENTION_HEAD_COUNT.format(llm=llm), count)
def add_head_count_kv(self, llm: str, count: int):
self.add_uint32(
constants.KEY_ATTENTION_HEAD_COUNT_KV.format(llm=llm), count)
def add_max_alibi_bias(self, llm: str, bias: float):
self.add_float32(
constants.KEY_ATTENTION_MAX_ALIBI_BIAS.format(llm=llm), bias)
def add_clamp_kqv(self, llm: str, value: float):
self.add_float32(
constants.KEY_ATTENTION_CLAMP_KQV.format(llm=llm), value)
def add_layer_norm_eps(self, llm: str, value: float):
self.add_float32(
constants.KEY_ATTENTION_LAYERNORM_EPS.format(llm=llm), value)
def add_layer_norm_rms_eps(self, llm: str, value: float):
self.add_float32(
constants.KEY_ATTENTION_LAYERNORM_RMS_EPS.format(llm=llm), value)
def add_rope_dimension_count(self, llm: str, count: int):
self.add_uint32(
constants.KEY_ROPE_DIMENSION_COUNT.format(llm=llm), count)
def add_rope_scale(self, llm: str, value: float):
self.add_float32(constants.KEY_ROPE_SCALE.format(llm=llm), value)
def add_tokenizer_model(self, model: str):
self.add_string(constants.KEY_TOKENIZER_MODEL, model)
def add_token_list(self, tokens: List):
self.add_array(constants.KEY_TOKENIZER_LIST, tokens)
def add_token_merges(self, merges: List):
self.add_array(constants.KEY_TOKENIZER_MERGES, merges)
def add_token_scores(self, scores: List[float]):
self.add_array(constants.KEY_TOKENIZER_SCORES, scores)
def add_bos_token_id(self, id: int):
self.add_uint32(constants.KEY_TOKENIZER_BOS_ID, id)
def add_eos_token_id(self, id: int):
self.add_uint32(constants.KEY_TOKENIZER_EOS_ID, id)
def add_unk_token_id(self, id: int):
self.add_uint32(constants.KEY_TOKENIZER_UNK_ID, id)
def add_sep_token_id(self, id: int):
self.add_uint32(constants.KEY_TOKENIZER_SEP_ID, id)
def add_pad_token_id(self, id: int):
self.add_uint32(constants.KEY_TOKENIZER_PAD_ID, id)
# Example usage:
if __name__ == "__main__":
# Example usage with a file
gguf_writer = GGUFWriter.open("example.gguf")
gguf_writer.add_architecture("llama")
gguf_writer.add_uint32("answer", 42) # Write a 32-bit integer
gguf_writer.add_float32("answer_in_float", 42.0) # Write a 32-bit float
gguf_writer.add_custom_alignment(64)
tensor1 = np.ones((32,), dtype=np.float32) * 100.0
tensor2 = np.ones((32,), dtype=np.float32) * 101.0
gguf_writer.add_tensor_info("tensor0", tensor1)
gguf_writer.add_tensor_info("tensor1", tensor2)
gguf_writer.write_header_to_file()
gguf_writer.write_kv_data_to_file()
gguf_writer.write_ti_data_to_file()
gguf_writer.write_tensor_to_file(tensor1)
gguf_writer.write_tensor_to_file(tensor2)
gguf_writer.close()