2023-03-22 06:32:36 +01:00
# ifndef LLAMA_H
# define LLAMA_H
2023-06-06 21:33:23 +02:00
# include "ggml.h"
2024-01-17 17:39:41 +01:00
# include "ggml-backend.h"
2024-01-31 16:30:17 +01:00
2023-03-22 06:32:36 +01:00
# include <stddef.h>
# include <stdint.h>
2023-08-28 17:59:39 +02:00
# include <stdio.h>
2023-03-22 06:32:36 +01:00
# include <stdbool.h>
# ifdef LLAMA_SHARED
2023-03-29 15:19:29 +02:00
# if defined(_WIN32) && !defined(__MINGW32__)
2023-03-22 06:32:36 +01:00
# 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
2023-06-24 10:47:58 +02:00
# 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
2023-08-21 22:07:43 +02:00
# define LLAMA_DEFAULT_SEED 0xFFFFFFFF
2023-05-20 14:58:15 +02:00
2023-09-28 18:04:36 +02:00
# define LLAMA_MAX_RNG_STATE (64*1024)
2023-12-16 18:58:46 +01:00
# define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla'
2023-08-21 22:07:43 +02:00
# define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
2023-03-22 06:32:36 +01:00
2023-08-21 22:07:43 +02:00
# define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN
2024-01-13 17:29:43 +01:00
# define LLAMA_SESSION_VERSION 4
2023-06-29 15:15:15 +02:00
2023-03-22 06:32:36 +01:00
# ifdef __cplusplus
extern " C " {
# endif
//
// C interface
//
// TODO: show sample usage
//
2023-06-24 10:47:58 +02:00
struct llama_model ;
2023-03-22 06:32:36 +01:00
struct llama_context ;
2023-09-28 18:04:36 +02:00
typedef int32_t llama_pos ;
typedef int32_t llama_token ;
typedef int32_t llama_seq_id ;
2023-03-22 06:32:36 +01:00
2023-08-21 22:07:43 +02:00
enum llama_vocab_type {
LLAMA_VOCAB_TYPE_SPM = 0 , // SentencePiece
LLAMA_VOCAB_TYPE_BPE = 1 , // Byte Pair Encoding
2024-02-11 17:21:38 +01:00
LLAMA_VOCAB_TYPE_WPM = 2 , // WordPiece
2023-08-21 22:07:43 +02:00
} ;
enum llama_token_type {
LLAMA_TOKEN_TYPE_UNDEFINED = 0 ,
LLAMA_TOKEN_TYPE_NORMAL = 1 ,
LLAMA_TOKEN_TYPE_UNKNOWN = 2 ,
LLAMA_TOKEN_TYPE_CONTROL = 3 ,
LLAMA_TOKEN_TYPE_USER_DEFINED = 4 ,
LLAMA_TOKEN_TYPE_UNUSED = 5 ,
LLAMA_TOKEN_TYPE_BYTE = 6 ,
} ;
// model file types
enum llama_ftype {
LLAMA_FTYPE_ALL_F32 = 0 ,
2023-09-28 18:04:36 +02:00
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
2024-01-08 16:02:32 +01:00
LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19 , // except 1d tensors
2024-01-11 20:39:39 +01:00
LLAMA_FTYPE_MOSTLY_IQ2_XS = 20 , // except 1d tensors
2024-01-11 20:43:15 +01:00
LLAMA_FTYPE_MOSTLY_Q2_K_S = 21 , // except 1d tensors
2024-01-22 11:43:33 +01:00
LLAMA_FTYPE_MOSTLY_Q3_K_XS = 22 , // except 1d tensors
2024-01-30 14:14:12 +01:00
LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23 , // except 1d tensors
2023-08-22 19:05:59 +02:00
LLAMA_FTYPE_GUESSED = 1024 , // not specified in the model file
2023-08-21 22:07:43 +02:00
} ;
2023-11-01 23:04:33 +01:00
enum llama_rope_scaling_type {
LLAMA_ROPE_SCALING_UNSPECIFIED = - 1 ,
LLAMA_ROPE_SCALING_NONE = 0 ,
LLAMA_ROPE_SCALING_LINEAR = 1 ,
LLAMA_ROPE_SCALING_YARN = 2 ,
LLAMA_ROPE_SCALING_MAX_VALUE = LLAMA_ROPE_SCALING_YARN ,
} ;
2024-01-12 20:07:38 +01:00
enum llama_split_mode {
LLAMA_SPLIT_NONE = 0 , // single GPU
LLAMA_SPLIT_LAYER = 1 , // split layers and KV across GPUs
LLAMA_SPLIT_ROW = 2 , // split rows across GPUs
} ;
2023-03-22 06:32:36 +01:00
typedef struct llama_token_data {
2023-05-20 10:06:11 +02:00
llama_token id ; // token id
float logit ; // log-odds of the token
float p ; // probability of the token
2023-03-22 06:32:36 +01:00
} llama_token_data ;
llama : new sampling algorithms (#1126)
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
2023-04-29 07:34:41 +02:00
typedef struct llama_token_data_array {
llama_token_data * data ;
size_t size ;
bool sorted ;
} llama_token_data_array ;
2023-12-22 07:19:36 +01:00
typedef bool ( * llama_progress_callback ) ( float progress , void * ctx ) ;
2023-03-25 06:26:28 +01:00
2023-09-28 18:04:36 +02:00
// Input data for llama_decode
// A llama_batch object can contain input about one or many sequences
// The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens
//
// - token : the token ids of the input (used when embd is NULL)
// - embd : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
// - pos : the positions of the respective token in the sequence
// - seq_id : the sequence to which the respective token belongs
// - logits : if zero, the logits for the respective token will not be output
//
typedef struct llama_batch {
int32_t n_tokens ;
2023-10-18 15:21:57 +02:00
llama_token * token ;
float * embd ;
llama_pos * pos ;
int32_t * n_seq_id ;
llama_seq_id * * seq_id ;
int8_t * logits ;
2023-09-28 18:04:36 +02:00
// NOTE: helpers for smooth API transition - can be deprecated in the future
// for future-proof code, use the above fields instead and ignore everything below
//
// pos[i] = all_pos_0 + i*all_pos_1
//
llama_pos all_pos_0 ; // used if pos == NULL
llama_pos all_pos_1 ; // used if pos == NULL
llama_seq_id all_seq_id ; // used if seq_id == NULL
} llama_batch ;
2023-12-05 18:19:18 +01:00
enum llama_model_kv_override_type {
LLAMA_KV_OVERRIDE_INT ,
LLAMA_KV_OVERRIDE_FLOAT ,
LLAMA_KV_OVERRIDE_BOOL ,
} ;
struct llama_model_kv_override {
char key [ 128 ] ;
enum llama_model_kv_override_type tag ;
union {
int64_t int_value ;
double float_value ;
bool bool_value ;
} ;
} ;
2023-09-28 21:42:38 +02:00
struct llama_model_params {
int32_t n_gpu_layers ; // number of layers to store in VRAM
2024-01-12 20:07:38 +01:00
enum llama_split_mode split_mode ; // how to split the model across multiple GPUs
// main_gpu interpretation depends on split_mode:
// LLAMA_SPLIT_NONE: the GPU that is used for the entire model
// LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results
// LLAMA_SPLIT_LAYER: ignored
int32_t main_gpu ;
2024-01-31 16:30:17 +01:00
// proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
2024-01-12 20:07:38 +01:00
const float * tensor_split ;
llama : add custom RoPE (#2054)
* Implement customizable RoPE
The original RoPE has pre-defined parameters
theta_i = 10000^(−2(i−1)/d), for i in [1, 2, ..., d/2]
Our customizable RoPE, ggml_rope_custom_inplace, uses
theta_i = scale * base^(−2(i−1)/d), for i in [1, 2, ..., d/2]
with the default matches the original
scale = 1.0
base = 10000
The new command line arguments
--rope-freq-base
--rope-freq-scale
set the two new RoPE parameter.
Recent researches show changing these two parameters extends the context limit with minimal loss.
1. Extending Context to 8K
kaiokendev
https://kaiokendev.github.io/til#extending-context-to-8k
2. Extending Context Window of Large Language Models via Positional Interpolation
Shouyuan Chen, Sherman Wong, Liangjian Chen, Yuandong Tian
https://arxiv.org/abs/2306.15595
3. NTK-Aware Scaled RoPE allows LLaMA models to have extended (8k+) context size without any fine-tuning and minimal perplexity degradation.
https://www.reddit.com/user/bloc97
https://www.reddit.com/r/LocalLLaMA/comments/14lz7j5/ntkaware_scaled_rope_allows_llama_models_to_have/
For the bold, try adding the following command line parameters to your favorite model:
-c 16384 --rope-freq-base 80000 --rope-freq-scale 0.5
* ggml-metal: fix custom rope
* common: fix argument names in help
* llama: increase MEM_REQ_EVAL for MODEL_3B
It avoids crashing for quantized weights on CPU.
Better ways to calculate the required buffer size would be better.
* llama: make MEM_REQ_EVAL depend on n_ctx
* server: use proper Content-Type in curl examples
Without the header Content-Type: application/json, curl will POST with
Content-Type: application/x-www-form-urlencoded
Though our simple server doesn't care, the httplib.h used has a limit
with CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 8192
With Content-Type: application/json, we can send large json data.
* style : minor fixes, mostly indentations
* ggml : fix asserts
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-07-15 12:34:16 +02:00
2023-12-22 07:19:36 +01:00
// Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
// If the provided progress_callback returns true, model loading continues.
// If it returns false, model loading is immediately aborted.
2023-06-20 03:24:39 +02:00
llama_progress_callback progress_callback ;
2023-12-05 18:19:18 +01:00
2023-06-20 03:24:39 +02:00
// context pointer passed to the progress callback
void * progress_callback_user_data ;
2023-03-22 06:32:36 +01:00
2023-12-05 18:19:18 +01:00
// override key-value pairs of the model meta data
const struct llama_model_kv_override * kv_overrides ;
2023-06-20 03:24:39 +02:00
// Keep the booleans together to avoid misalignment during copy-by-value.
2023-03-22 06:32:36 +01:00
bool vocab_only ; // only load the vocabulary, no weights
Rewrite loading code to try to satisfy everyone:
- Support all three formats (ggml, ggmf, ggjt). (However, I didn't
include the hack needed to support GPT4All files without conversion.
Those can still be used after converting them with convert.py from my
other PR.)
- Support both mmap and read (mmap is used by default, but can be
disabled with `--no-mmap`, and is automatically disabled for pre-ggjt
files or on platforms where mmap is not supported).
- Support multi-file models like before, but automatically determine the
number of parts rather than requiring `--n_parts`.
- Improve validation and error checking.
- Stop using the per-file type field (f16) entirely in favor of just
relying on the per-tensor type/size fields. This has no immediate
benefit, but makes it easier to experiment with different formats, and
should make it easier to support the new GPTQ-for-LLaMa models in the
future (I have some work in progress on that front).
- Support VirtualLock on Windows (using the same `--mlock` option as on
Unix).
- Indicate loading progress when using mmap + mlock. (Which led me
to the interesting observation that on my Linux machine, with a
warm file cache, mlock actually takes some time, whereas mmap
without mlock starts almost instantly...)
- To help implement this, move mlock support from ggml to the
loading code.
- madvise/PrefetchVirtualMemory support (based on #740)
- Switch from ifstream to the `fopen` family of functions to avoid
unnecessary copying and, when mmap is enabled, allow reusing the same
file descriptor for both metadata reads and mmap (whereas the existing
implementation opens the file a second time to mmap).
- Quantization now produces a single-file output even with multi-file
inputs (not really a feature as much as 'it was easier this way').
Implementation notes:
I tried to factor the code into more discrete pieces than before.
Regarding code style: I tried to follow the code style, but I'm naughty
and used a few advanced C++ features repeatedly:
- Destructors to make it easier to ensure everything gets cleaned up.
- Exceptions. I don't even usually use exceptions when writing C++, and
I can remove them if desired... but here they make the loading code
much more succinct while still properly handling a variety of errors,
ranging from API calls failing to integer overflow and allocation
failure. The exceptions are converted to error codes at the
API boundary.)
Co-authored-by: Pavol Rusnak <pavol@rusnak.io> (for the bit I copied from #740)
2023-04-08 21:24:37 +02:00
bool use_mmap ; // use mmap if possible
2023-03-24 16:19:05 +01:00
bool use_mlock ; // force system to keep model in RAM
2023-09-28 21:42:38 +02:00
} ;
struct llama_context_params {
2023-11-03 08:24:00 +01:00
uint32_t seed ; // RNG seed, -1 for random
uint32_t n_ctx ; // text context, 0 = from model
uint32_t n_batch ; // prompt processing maximum batch size
uint32_t n_threads ; // number of threads to use for generation
uint32_t n_threads_batch ; // number of threads to use for batch processing
2024-02-03 12:22:06 +01:00
int32_t rope_scaling_type ; // RoPE scaling type, from `enum llama_rope_scaling_type`
2023-09-28 21:42:38 +02:00
// ref: https://github.com/ggerganov/llama.cpp/pull/2054
2023-11-01 23:04:33 +01:00
float rope_freq_base ; // RoPE base frequency, 0 = from model
float rope_freq_scale ; // RoPE frequency scaling factor, 0 = from model
2023-11-25 16:47:07 +01:00
float yarn_ext_factor ; // YaRN extrapolation mix factor, negative = from model
2023-11-01 23:04:33 +01:00
float yarn_attn_factor ; // YaRN magnitude scaling factor
float yarn_beta_fast ; // YaRN low correction dim
float yarn_beta_slow ; // YaRN high correction dim
uint32_t yarn_orig_ctx ; // YaRN original context size
2023-09-28 21:42:38 +02:00
2024-01-17 17:39:41 +01:00
ggml_backend_sched_eval_callback cb_eval ;
void * cb_eval_user_data ;
2023-12-07 12:03:17 +01:00
enum ggml_type type_k ; // data type for K cache
enum ggml_type type_v ; // data type for V cache
2023-09-28 21:42:38 +02:00
// Keep the booleans together to avoid misalignment during copy-by-value.
2023-12-07 12:03:17 +01:00
bool mul_mat_q ; // if true, use experimental mul_mat_q kernels (DEPRECATED - always true)
2023-12-12 10:25:57 +01:00
bool logits_all ; // the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
2023-12-07 12:03:17 +01:00
bool embedding ; // embedding mode only
bool offload_kqv ; // whether to offload the KQV ops (including the KV cache) to GPU
2024-02-13 13:06:58 +01:00
bool do_pooling ; // whether to pool (sum) embedding results by sequence id (ignored if no pooling layer)
2023-03-22 06:32:36 +01:00
} ;
2023-08-21 22:07:43 +02:00
2023-06-10 09:59:17 +02:00
// model quantization parameters
typedef struct llama_model_quantize_params {
2024-01-02 15:15:16 +01:00
int32_t nthread ; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
2023-08-21 22:07:43 +02:00
enum llama_ftype ftype ; // quantize to this llama_ftype
2023-06-10 09:59:17 +02:00
bool allow_requantize ; // allow quantizing non-f32/f16 tensors
bool quantize_output_tensor ; // quantize output.weight
2023-09-01 16:02:48 +02:00
bool only_copy ; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
2023-10-29 17:32:28 +01:00
bool pure ; // disable k-quant mixtures and quantize all tensors to the same type
2024-01-14 08:45:56 +01:00
void * imatrix ; // pointer to importance matrix data
2023-06-10 09:59:17 +02:00
} llama_model_quantize_params ;
2023-07-24 05:58:10 +02:00
// 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 ;
2023-07-05 22:51:13 +02:00
// 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 ;
} ;
2023-09-28 18:04:36 +02:00
// Helpers for getting default parameters
2023-09-28 21:42:38 +02:00
LLAMA_API struct llama_model_params llama_model_default_params ( void ) ;
2023-08-21 22:07:43 +02:00
LLAMA_API struct llama_context_params llama_context_default_params ( void ) ;
LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params ( void ) ;
2023-07-19 09:06:40 +02:00
2023-05-20 10:06:11 +02:00
// Initialize the llama + ggml backend
2023-06-26 19:57:59 +02:00
// If numa is true, use NUMA optimizations
2023-05-20 10:06:11 +02:00
// Call once at the start of the program
2023-07-10 17:49:56 +02:00
LLAMA_API void llama_backend_init ( bool numa ) ;
2023-05-20 10:06:11 +02:00
2023-08-21 22:07:43 +02:00
// Call once at the end of the program - currently only used for MPI
LLAMA_API void llama_backend_free ( void ) ;
2023-05-20 10:06:11 +02:00
2023-06-24 10:47:58 +02:00
LLAMA_API struct llama_model * llama_load_model_from_file (
const char * path_model ,
2023-09-28 21:42:38 +02:00
struct llama_model_params params ) ;
2023-06-24 10:47:58 +02:00
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 ) ;
2023-03-22 06:32:36 +01:00
// Frees all allocated memory
LLAMA_API void llama_free ( struct llama_context * ctx ) ;
2023-08-21 22:07:43 +02:00
LLAMA_API int64_t llama_time_us ( void ) ;
2024-01-31 16:30:17 +01:00
LLAMA_API size_t llama_max_devices ( void ) ;
LLAMA_API bool llama_supports_mmap ( void ) ;
LLAMA_API bool llama_supports_mlock ( void ) ;
LLAMA_API bool llama_supports_gpu_offload ( void ) ;
LLAMA_API DEPRECATED ( bool llama_mmap_supported ( void ) , " use llama_supports_mmap() instead " ) ;
LLAMA_API DEPRECATED ( bool llama_mlock_supported ( void ) , " use llama_supports_mlock() instead " ) ;
2023-08-21 22:07:43 +02:00
2023-09-28 21:42:38 +02:00
LLAMA_API const struct llama_model * llama_get_model ( const struct llama_context * ctx ) ;
2023-12-21 20:57:48 +01:00
LLAMA_API uint32_t llama_n_ctx ( const struct llama_context * ctx ) ;
LLAMA_API uint32_t llama_n_batch ( const struct llama_context * ctx ) ;
2023-08-21 22:07:43 +02:00
2023-09-28 21:42:38 +02:00
LLAMA_API enum llama_vocab_type llama_vocab_type ( const struct llama_model * model ) ;
2023-08-23 22:08:04 +02:00
2024-01-02 15:15:16 +01:00
LLAMA_API int32_t llama_n_vocab ( const struct llama_model * model ) ;
LLAMA_API int32_t llama_n_ctx_train ( const struct llama_model * model ) ;
LLAMA_API int32_t llama_n_embd ( const struct llama_model * model ) ;
2023-08-21 22:07:43 +02:00
2023-10-03 19:09:28 +02:00
// Get the model's RoPE frequency scaling factor
LLAMA_API float llama_rope_freq_scale_train ( const struct llama_model * model ) ;
2023-11-17 16:17:37 +01:00
// Functions to access the model's GGUF metadata scalar values
// - The functions return the length of the string on success, or -1 on failure
// - The output string is always null-terminated and cleared on failure
// - GGUF array values are not supported by these functions
// Get metadata value as a string by key name
2024-01-02 15:15:16 +01:00
LLAMA_API int32_t llama_model_meta_val_str ( const struct llama_model * model , const char * key , char * buf , size_t buf_size ) ;
2023-11-17 16:17:37 +01:00
// Get the number of metadata key/value pairs
2024-01-02 15:15:16 +01:00
LLAMA_API int32_t llama_model_meta_count ( const struct llama_model * model ) ;
2023-11-17 16:17:37 +01:00
// Get metadata key name by index
2024-01-02 15:15:16 +01:00
LLAMA_API int32_t llama_model_meta_key_by_index ( const struct llama_model * model , int32_t i , char * buf , size_t buf_size ) ;
2023-11-17 16:17:37 +01:00
// Get metadata value as a string by index
2024-01-02 15:15:16 +01:00
LLAMA_API int32_t llama_model_meta_val_str_by_index ( const struct llama_model * model , int32_t i , char * buf , size_t buf_size ) ;
2023-11-17 16:17:37 +01:00
2023-08-21 22:07:43 +02:00
// Get a string describing the model type
2024-01-02 15:15:16 +01:00
LLAMA_API int32_t llama_model_desc ( const struct llama_model * model , char * buf , size_t buf_size ) ;
2023-09-28 18:04:36 +02:00
2023-08-25 15:16:19 +02:00
// Returns the total size of all the tensors in the model in bytes
LLAMA_API uint64_t llama_model_size ( const struct llama_model * model ) ;
2023-09-28 18:04:36 +02:00
2023-08-25 15:16:19 +02:00
// Returns the total number of parameters in the model
LLAMA_API uint64_t llama_model_n_params ( const struct llama_model * model ) ;
2023-08-21 22:07:43 +02:00
train : finetune LORA (#2632)
* fix track_max_mem in forward_batch_wo_cache_flash_attn_train
* remove unnecessary Adam(W) optimizer tensors.
reduces optimizer memory overhead from 7*modelsize to 2*modelsize.
additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t.
bumps training checkpoint file version, but old checkpoints can still be read.
new version with less tensors is saved.
* add gradient clipping to AdamW
* Fix reset of unused g->nodes and g->grads to NULL
* implement gradient checkpointing for training
reduces memory overhead from O(n_layer) to O(sqrt(n_layer))
as explained in readme of https://github.com/cybertronai/gradient-checkpointing
* remove unused compute buffer 3
* add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes
GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep);
* change AdamW decay parameter to work like the torch AdamW decay parameter
It is now relative to Adam learning rate `alpha*sched`.
Before that it was relative to `sched` only.
`alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1]
* change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT
* change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW
btw: the default weight decay parameter for torch.optim.AdamW is 0.01
* bug fixes for cross entropy loss
ggml_cross_entropy_loss: sums where not correctly added in workload of each thread
ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues
guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16
cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup.
so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance.
* fix test-grad0 for cross_entropy_loss
the second argument to cross_entropy_loss must sum up to 1 for each row
* fix test-grad0 for soft_max
dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work
instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0)
* improve finite differences of test-grad0 by using double instead of float
* change cross_entropy_loss to output average over all rows
this helps keeping the loss and gradients in a sane range
* improve gradient checkpointing
sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal.
since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different:
```
given: n, u, v
objective: minimize(a*u+b*v) where a*b=n, a>0, b>0
b=n/a
minimize(a*u+v*n/a)
diff(a*u+v*n/a, a) = u - (v*n/a)/a
diff(a*u+v*n/a, a) == 0
u - (v*n/a)/a == 0
u == v*n/(a*a)
u*a*a = v*n
a*a = v*n/u
a = sqrt(n*v/u)
```
this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage.
* disable gradient checkpointing debug output
* llama : fix rope usage in train-text-from-scratch after ChatGLM change
* add more training parameters:
--enable-restart N Only for Adam optimizer. Enable restarts of cos-decay
--disable-restart N Only for Adam optimizer. Disable restarts of cos-decay
--opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero.
--opt-delta N Maximum delta for delta convergence test. Disabled when <= zero.
--opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero.
--adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero.
--adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha
* replace memcpy with reshape operation so that the graph is not cut at the input
this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it
* remove unused function argument from get_example_targets_batch
* measure and print total training time
* add optimization callback to ggml_opt_resume_g
this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)).
can be used for dynamic learning schedule and setting input data for batches before each iteration
* use optimization callback in training
allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters
reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration
* add minimum number of tensor dimensions to apply weight decay (default 2)
this allows to not apply weight decay to bias parameters
* rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup
* fix increase of model.train_samples and model.train_tokens
now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations
* change sampling parameters for prediction after training to defaults of common.h
and clarify what is context for prediction and what are generated tokens
* tighten abs error bounds for cross_entropy_loss in test-grad0
* add conditional compilation of using F16 exp in flash attention
uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention
* tighten abs error bounds for flash_attn in test-grad0
* tighten abs error bounds for sqrt in test-grad0
* remove out-commented vectorized code of opt_adam
the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead
* ggml : update ggml_rms_norm_back with configurable eps
* llama training : fix ggml_rms_norm_back calls to pass configurable eps
* remove trailing whitespace
* add train function using automatic gradient checkpointing backward pass and allocator
* in train function replace add_inplace by regular add
because using add_inplace seems to result in different gradients
* don't use allocate hash_map on context
because the context has no_alloc=True when using memory allocator resulting in NULL data pointers
* correctly clone reshape and permute operations by also cloning tensor->nb values
* fix variable name and add missing type cast
* terminate recursive tensor cloning when reaching tensor without src tensors
* correctly clone view tensors by setting data pointers
without this the checkpointing would only work when being used together with memory allocator
* fix variable names
* swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn`
* add input tensors as checkpoints
so that recursive tensor cloning of gradient checkpointing terminates on input tensors
* fix variable name and add missing boolean negation
* make sure some tensors are not reallocated by inserting new temporary nodes depending on them:
output and parameter gradient tensors need to be available at the end of the graph execution
parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration
checkpoint tensors are allocated all together to reduce memory allocator fragmentation
afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs
* fix ASSERT to work with zero layers
* add training options whether to use allocator and/or unified training function
* integrate unified training function which may use memory allocator
the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing
* format name of cloned tensors with " (clone)" suffix
* set names for tensors in unified train function for easier debugging
* allocate graph on context using ggml_new_graph
* remove handwritten training functions
* remove unused training parameters "use_scratch" and "use_unified"
* remove trailing whitespace
* remove unused train params: mem_compute1_gb & mem_compute2_gb
mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions
mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented)
* remove unused forward_batch function
* add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly
* only use ggml_allocr_alloc when tensor has NULL data and is no view
* fix test when to create temporary backward graph
temporary backward graph is only necessary when using checkpointing
* fix memory "leak" in optimizers
each iteration a new cplan with new memory for work data was allocated.
now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data.
* reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator
with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory.
the computation results are the same
* add API functions to access llama model tensors
* add stub example for finetuning, based on train-text-from-scratch
* move and remove code
* add API functions to access remaining model parameters:
mult, head and rot
* first draft for LORA finetune training
* remove const model and layer arguments in API functions for accessing model tensors
* bug fixes to make finetune compile
automatic allocator does not work yet
* add debug prints for training memory improvements
* fix names of lora tensors
* avoid stack overflow resulting from big ggml_cgraph
replace stack allocation and ggml_build_forward by ggml_new_graph in combination with ggml_build_forward_expand
* replace llama API functions to get model tensors by one function to get model tensor by name
LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name);
* remove unused call to not existing llama_get_layer_from_model
* implement ggml_compute_forward_out_prod_q_f32
* remove trailing whitespace
* add lora finetune support on quantized base model tensors
* add ggml_add_cast API function
this function works like ggml_add, but accepts a data type for the resulting tensor.
only supported for quantized src0 input.
* use ggml_add_cast in finetuning
lora-applied weights will now have data type F32, which improves gradients when finetuning quantized base models
* bug fix: actually use result type passed to ggml_add_cast
* make sure base model tensors data cannot be used in viewable operations
memory allocator would try to make lora application inplace on base model tensors.
since those are memory mapped this will result in memory access violations
* fix bug in ggml_out_prod which resulted in wrong n_dims of result tensors
* avoid keeping in memory ALL of the gradients
The problem here stems from ggml_graph_reset. This function is called in the optimization function, before each graph computation, to reset the gradients to zero. This required a unique memory slot for each gradient: allocating memory from a previosly freed memory location might lead to non-zero input gradients.
During ggml_compute_backward the gradients are build stepwise by adding or substracting new values, starting from a OP_NONE tensor which needs to contain zero-values. This requires the graph reset.
To avoid this I now remember in ggml_build_backward_expand the original OP_NONE gradient tensors in a hash table, which is passed to ggml_compute_backward. There instead of using add (or sub or similar) I test whether the existing gradient to be changed is a zero-valued-tensor by looking up its existence in the hash table. When it is such a zero-tensor it will not be modified, but replaced by the value to be added, otherwise the regular add (not inplace, allocator will take care of this) will be used. This way none of those zero-tensor values will be necessary in the final backward graph and more importantly they won't need a unique memory slot, just to make them zero.
* remove trailing whitespace
* remove debug prints and function to compute tensor data hash
* improve optimization iteration prints
* adjust maximal values to support finetuning 3B models
* change default finetune params lora_r and lora_alpha to match the n_rank parameters of 4
* bug fix: make sure finetune input gradient is allocated at begin and kept until end
* remove unnecessary src tensor from ggml_get_rows_back
we don't need data of src[2] for computation, only to setup the correct output shape.
remove dependency on src[2], so that allocator can work more freely.
the computational graph is still completely determined, because the output shape is naturally included.
this is similar to how ggml_reshape does it.
* remove unnecessary src tensor from ggml_repeat & ggml_repeat_back
we don't need data of src[1] for computation, only to setup the correct output shape.
remove dependency on src[1], so that allocator can work more freely.
the computational graph is still completely determined, because the output shape is naturally included
* resolve todo
allocator will only make it inplace when they are of the same type
* mixing multiple LORA adapters is now possible
pass more than one '--lora FNAME' argument to apply more than one LORA.
use '--lora-scaled FNAME S' when you want to specify a user-defined scale for an adapter.
* add option to save finetune output every N iterations
* also save latest finetune output with ITERATION="LATEST" and print where files are saved
saving with LATEST makes it easier to resume training from the latest checkpoint
the string "LATEST" can be configured with command line option "--fn-latest STR"
* update checkpoint train stats before saving via "--save-every"
* add command line option `--rank-wo N` for rank of wo tensor
* update finetune README
* fix dump_non_result_info_yaml to output multiple lora adapters
* bug fix: replace GGML_TYPE_SIZE[t] by ggml_type_size(t)
* replace llama_n_mult by llama_n_ff
* finetune bug fixes to compile with merged in code from master
* remove prediction related code to reduce duplicated code with main
use main instead
* reduce large memory overhead in train-text-from-scratch
all gradients had to be pinned so that graph_reset works correctly.
this is no longer necessary with the changes to ggml_compute_backward introduced in this PR.
* add comment explaining why finetune checkpoints are allocated in one block
* make default value of float member a float literal
* handle rms_norm and rope parameters the same as in train-text-from-scratch
* remove unused code
* remove vocab related code as it is unnecessary
* add LLM_KV_TRAINING_TYPE to train-text-from-scratch checkpoints
so that they can be differentiated from lora finetune checkpoints
* add gguf constants and load/save functions from train-text-from-scratch
* add load & save lora finetune checkpoints via gguf
* add python script to convert old finetune checkpoint files to gguf
* remove old checkpoint save & load code
* remove code to print data checksums which was used to verify correctness of new gguf code
* omit tokenization when training is disabled, only save llama lora adapter
training can be disabled by passing '-n 0' to finetune
* remove trailing whitespace
* update README.md
* implement ggml_compute_forward_repeat_f16
* avoid stack overflow of large cgraphs in test-grad0
* add ggml API functions ggml_unravel_index, ggml_get_i32_nd and its analogs for set and for f32
ggml_get_i32_1d, ggml_set_i32_1d, ggml_get_f32_1d, ggml_set_f32_1d now support non-contiguous tensors.
in case of non-contiguous tensor, the 1d index is unraveled into a multi index using ggml_unravel_index to be passed to '_nd' function equivalent.
this fixes a bug in test-grad0 which happens due to ggml_build_backward not building purely contiguous tensors anymore
* increase test-grad0 context mem size to accommodate for bigger cgraph
* add sanity check to ggml_compute_backward, asserting the correct shape of gradients
* fix ggml_acc_or_set to return tensor of correct shape
* remove unused 'inplace' argument from ggml_compute_backward function
inplace operations to add gradients are no longer created by ggml_compute_backward
use allocator to automatically make inplace operations
* add missing argument 'int i0' to ggml_get_i32_nd & ggml_set_i32_nd header declarations
* fix error message in ggml_allocr_alloc to display actual max_avail
* fix check_gradient
ggml_build_backward_expand was previously replaced by ggml_build_backward, but the assignment of forward graph to backward graph missing
* use tensor->view_src instead of ggml_is_view and get_view_source
* move gradient checkpointing code into ggml, new API function:
// build gradient checkpointing backward graph gb for gf using provided checkpoints
// gb_tmp will contain original backward graph with rewritten backward process nodes,
// but without the second forward pass nodes.
GGML_API void ggml_build_backward_gradient_checkpointing(
struct ggml_context * ctx,
struct ggml_cgraph * gf,
struct ggml_cgraph * gb,
struct ggml_cgraph * gb_tmp,
struct ggml_tensor * * checkpoints,
int n_checkpoints);
* replace custom data getters and setters by ggml functions
* train-text-from-scratch can train (full finetune) gguf models
just pass the gguf model via `--checkpoint-in FN`.
after this, to continue training, pass the generated checkpoint instead of the original gguf model.
tested with smaller models, bigger models may exceed available memory.
use (LORA) finetune for those.
* remove trailing whitespace
* add option to save train-text-from-scratch output every N iterations
* update README.md
* fix warnings
* fix warnings
* remove finetune option to disable allocator
the allocator should always be used.
by making sure that it is always used it gets easier to implement automatic memory requirements computation
* add tensor checkpoints only when gradient checkpointing is enabled
* initialize opt ggml context if none was provided
* add ggml-alloc API function 'ggml_allocr_max_size' to get max size of alloc
GGML_API size_t ggml_allocr_max_size(struct ggml_allocr * alloc);
* finetune: automatically allocate all memory and changes to command line options
remove '--n_examples N' parameter, as it no longer makes sense to call optimization process multiple times in a loop.
add '--only_write_lora' command line option: will skip tokenization and training, to only write a llama.cpp comptabile LORA adapter.
remove memory buffer related command line options.
improve iteration console output.
* add finetune to Makefile
* update README.md
* print time per iteration and estimate remaining time
* increase measured alloc size by tensor_alignment
ggml_allocr_reset will reduce the given size by up to tensor_alignment-1
* fix README.md
* add some more allocator debug prints
* bug fix, probably solves the 'ggml_allocr_alloc: not enough space in the buffer' issue
* revert last commit
"bug fix, probably solves the 'ggml_allocr_alloc: not enough space in the buffer' issue"
"alloc was freeing an externally allocated tensor, because it calculated the end of allocator memory as alloc->data + alloc->max_size instead of alloc->data + alloc->size."
This is intentional to reduce the risk of freeing external tensors when measuring. Unless max_size is not properly calculated, I don't see why this is an issue.
* remove unnecessary "0x" before "%p" output
* move measurement memory segment to upper region of the address space
* update README.md
* fix printf format warnings
* add missing gguf_free in load_checkpoint_lora_file
* load default rms_norm and rope parameters from base model
* add gradient accumulation
specify number accumulation steps with '--grad-acc N'.
this will simulate a bigger batch size of grad_acc*batch.
* fix tracking of train_samples and train_tokens
* build : fix compile warnings
* ggml : fix L-BFGS linesearch loop
* improve finetune time measurement
fix printf warnings on system where int64_t is (long int).
change time datatypes to double because values get big with long training times.
exclude file saving from time measurement.
converge faster to actual time per iteration by removing very small first duration before first iteration was performed.
fix bug in output of total training time, the reported value was 1000 times to small.
* specify default lora rank with '--lora-r N'
'--lora-r N' will specify default rank for all tensors
'--rank-wq N', etc. will override this default rank for specific tensor types.
* fix gradient accumulation bug where the same batch was used for each microstep
* fix gradient accumulation bug where the same batch was used for each microstep
* support grouped-query-attention in ggml_flash_attn and ggml_flash_attn_back
k and v can now be repeated in q along ne[2]
in forward pass just use modulo to compute k and v indices, like ik2 = iq2 % nek2.
in backard pass this won't work as easy, because multiple threads will compete to accumulate to the same k->grad[:,ik1,ik2,ik3] and v->grad[:,iv1,iv2,iv3].
so we change the parallelization over q rows to be over k rows. this ensures non-overlapping (ik2,ik3) across threads.
in each thread we then iterate over the number of repetitions of k/v in q to compute iq2 as iq2 = ik2 + irep*nek2.
since ne2 is not the same for q,k and v we also change how the gradients are concatenated into the result tensor.
additionally the offsets of gradq, gradk and gradv in the result tensor are now memory aligned.
we also simplify the compute_backward part of flash_attn to use ggml_reshape instead of switching over the number of dimensions.
this needs a small change to ggml_reshape, removing the assertion of second argument to be contiguous.
since only the shape (ne) of the second reshape argument is of relevance, its memory layout (nb) is irrelevant -> it can very well be non-contiguous.
change test-grad0 to also test for repeated k/v in q.
this changes the rng and now results in small gradient differences in softmax. these solely come from using f16 exp table lookup in forward softmax: when temporarily changing softmax to use actual exp function, the reported gradient differences go away. gradient differences coming solely from f16 table lookup are acceptable.
added a note to explain this.
* add llama API functions to get grouped-query-attention n_head parameter 'n_head_kv'.
* fix finetune to support grouped-query-attention (using flash-attention)
note: ggml changes to ggml_out_prod are necessary to support grouped-query-attention without flash-attention.
* support broadcastable a in out_prod(a, b) and backward pass of broadcasting mul_mat(a, b)
* test broadcasting mul_mat backward pass
* decouple random number generator of each operation test
when changing one test the rng of others tests is not influenced anymore
* add comment briefly describing what ggml_repeat_back does
* simplify broadcasting mul_mat backward using ggml_repeat_back
* add cgraph evaluation order member and corresponding enum type
this controls in which order ggml_build_forward visits source nodes.
by default the nodes are visited left to right, i.e. src[0] first.
in some cases it is beneficial for ggml-alloc to visit in a different order.
two possible orders are supported: left-to-right (src[0] first) and right-to-left (src[0] last).
* measure max compute size for each cgraph eval order and use best order
this can bring huge memory savings:
e.g. codellama-34b with n_ctx=64, n_batch=1 goes from 92927.8mb down to 4627.6 MB
* remove unused command line options
* add sample start patterns and options to force new or by default resume last shuffling
* update shuffle rng state on reshuffle
* exclude known zero values from computations in flash_attn_f32 & flash_attn_back_f32
* remove probably unnecessary exception type flags from stringstream
* pass correct max number of tokens to llama_tokenize
* account for possible leading whitespace that will be added by tokenizer
e.g. '\t' will be tokenized by llama spm tokenizer to [29871, 12]
* use unrolled vec_mad in out_prod
y is vec_mad result vec.
x is vec_mad input vec.
v is vec_mad input scalar.
ggml_vec_mad_f32_unroll will internally loop over x and v with same y.
GGML_VEC_MAD_UNROLL is by default defined to 32.
This value is empirical optimized using performance test runs of out-prod in openllama-3b finetune with 256 context length and batch size 1. It gives 23% performance boost for out_prod.
Full measurements of out-prod runtime in ms:
unroll_xv unroll_yv
1 67014.643 87826.469
2 77117.552 89077.656
4 72091.311 109121.657
8 61077.543 88678.334
16 56914.67 79514.947
24 59024.595 84350.254
28 55952.446 83368.73
32 51476.658 85177.745
36 55973.792 84659.92
40 55139.616 93844.738
48 60736.392 93330.267
64 99856.878 116994.99
Second column is when unrollying yv instead of xv
* set lora_alpha to value of lora_r if it is not set via command line
otherwise only changing lora_r will change scaling of lora adapter used in prediction
* reshuffle original sample order instead of the previous shuffled order
otherwise resumed reshuffle will not result in same sample order
* block tiling for out-prod inspired by mul-mat
block sizes are empirically optimized
roughly doubles the flops of out-prod
* exclude some more known zero values from computations in flash_attn_f32 & flash_attn_back_f32
* add static keywords
* remove outcommented old code
* update train-text-from-scratch with tokenization, sample selection and shuffling from finetune
* remove lbfgs related train parameters
* move common train functions into common/train.[h|cpp]
* move train state into struct train_state
* move train data saving code into callback to unify code of opt_callback
train_params are still different in finetune and train-text-from-scratch, so it can't yet be moved to train.h|cpp
* move common train params into common/train
* move common opt_callback into common/train
* fix consume_common_train_arg
* save and load head_count_kv in lora checkpoints
* increase train_samples by used_samples instead of number of batches
on batch can contain more than one sample when option "fill_with_next_samples" is used
* fix usage of llama_tokenize
* remove static from process_escape since we need it exposed in header
* fix code formating of long function declarations
* fix condition in load_train_state_gguf
* use die("msg") instead of replace GGML_ASSERT(!"msg") or throw std::runtime_error("msg")
* fix saving and loading of training type
* remove terminating '\0' from tokenization
(llama_tokenize is now passed the string length instead of relying on terminating '\0')
* fix compile warnings
* fix compile warnings
* use new/delete for train_state instead of malloc/free
using malloc may result in seg faults when trying to assign string fields
* assert that sample_count > 0, avoiding division by zero
* fix frand to return value in interval [0,1)
* add train option "--sample-random-offsets"
Use samples beginning at random offsets.
The offset is only applied to the first sample in each batch context window.
Together with "--fill-with-next-samples" this may help for training endless text generation.
For example given a dataset containing samples "abcd", "ABCD", "0123".
With context size of 8 and options "--fill-with-next-samples", "--no-separate-with-eos", "--no-separate-with-bos",
the context windows of batches could only be filled with "abcdABCD", "ABCDabcd", "0123abcd", etc.
With "--sample-random-offsets" it can also be filled with "23abcdAB", "bcd0123A", etc.
* deduplicate code into function
* remove n_rot hparam, as it must always be hparam.n_embd_head()
* align code
* assert correct base model tensor shapes
* move some params from lora hparams into model hparams and load model params from gguf
this equalizes the model definition in finetune and text-from-scratch and removes the need for additional llama api functions to get model parameters
* remove now unnecessary llama API functions to get model params that where added by this PR
* train-text-from-scratch: automatically allocate model tensors, remove option '--mem-model N'
* train-text-from-scratch: automatically allocate opt context
* train-text-from-scratch: automatically allocate input tensors
* train-text-from-scratch: automatically allocate compute memory
* remove unused options and equalize train-text-from-scratch with finetune
* initialize opt->loss_after with zero
* add export-lora program
* remove trailing whitespace
* add export-lora build in Makefile
* remove unused struct tensor_info from export-lora
* add export-lora build dependency to llama
because it depends on common, which depends on llama
* update finetune README.md
* cancel optimization when specified number of epochs is completed
* improve handling of export-lora arguments
print errors and warnings when files could not be read or created
* Fix export-lora.cpp "not enough space in the context's memory pool" (#1)
* Fix export-lora.cpp "not enough space in the context's memory pool"
Without this patch, export-lora would sometimes error with "not enough space in the context's memory pool (needed 656784, available 656800)".
* increase required context size by 5*GGML_MEM_ALIGN instead of plain 16
---------
Co-authored-by: xaedes <xaedes@gmail.com>
* improve handling of not yet supported tensor types
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: meatbag-18a <145869052+meatbag-18a@users.noreply.github.com>
2023-09-28 20:40:11 +02:00
// Get a llama model tensor
LLAMA_API struct ggml_tensor * llama_get_model_tensor ( struct llama_model * model , const char * name ) ;
2023-03-22 06:32:36 +01:00
// Returns 0 on success
2024-01-02 15:15:16 +01:00
LLAMA_API uint32_t llama_model_quantize (
2023-03-22 06:32:36 +01:00
const char * fname_inp ,
const char * fname_out ,
2023-06-10 09:59:17 +02:00
const llama_model_quantize_params * params ) ;
2023-03-22 06:32:36 +01:00
2023-04-17 17:28:55 +02:00
// 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
2024-01-02 15:15:16 +01:00
LLAMA_API DEPRECATED ( int32_t llama_apply_lora_from_file (
2023-04-17 17:28:55 +02:00
struct llama_context * ctx ,
2023-06-24 10:47:58 +02:00
const char * path_lora ,
train : finetune LORA (#2632)
* fix track_max_mem in forward_batch_wo_cache_flash_attn_train
* remove unnecessary Adam(W) optimizer tensors.
reduces optimizer memory overhead from 7*modelsize to 2*modelsize.
additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t.
bumps training checkpoint file version, but old checkpoints can still be read.
new version with less tensors is saved.
* add gradient clipping to AdamW
* Fix reset of unused g->nodes and g->grads to NULL
* implement gradient checkpointing for training
reduces memory overhead from O(n_layer) to O(sqrt(n_layer))
as explained in readme of https://github.com/cybertronai/gradient-checkpointing
* remove unused compute buffer 3
* add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes
GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep);
* change AdamW decay parameter to work like the torch AdamW decay parameter
It is now relative to Adam learning rate `alpha*sched`.
Before that it was relative to `sched` only.
`alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1]
* change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT
* change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW
btw: the default weight decay parameter for torch.optim.AdamW is 0.01
* bug fixes for cross entropy loss
ggml_cross_entropy_loss: sums where not correctly added in workload of each thread
ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues
guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16
cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup.
so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance.
* fix test-grad0 for cross_entropy_loss
the second argument to cross_entropy_loss must sum up to 1 for each row
* fix test-grad0 for soft_max
dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work
instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0)
* improve finite differences of test-grad0 by using double instead of float
* change cross_entropy_loss to output average over all rows
this helps keeping the loss and gradients in a sane range
* improve gradient checkpointing
sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal.
since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different:
```
given: n, u, v
objective: minimize(a*u+b*v) where a*b=n, a>0, b>0
b=n/a
minimize(a*u+v*n/a)
diff(a*u+v*n/a, a) = u - (v*n/a)/a
diff(a*u+v*n/a, a) == 0
u - (v*n/a)/a == 0
u == v*n/(a*a)
u*a*a = v*n
a*a = v*n/u
a = sqrt(n*v/u)
```
this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage.
* disable gradient checkpointing debug output
* llama : fix rope usage in train-text-from-scratch after ChatGLM change
* add more training parameters:
--enable-restart N Only for Adam optimizer. Enable restarts of cos-decay
--disable-restart N Only for Adam optimizer. Disable restarts of cos-decay
--opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero.
--opt-delta N Maximum delta for delta convergence test. Disabled when <= zero.
--opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero.
--adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero.
--adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha
* replace memcpy with reshape operation so that the graph is not cut at the input
this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it
* remove unused function argument from get_example_targets_batch
* measure and print total training time
* add optimization callback to ggml_opt_resume_g
this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)).
can be used for dynamic learning schedule and setting input data for batches before each iteration
* use optimization callback in training
allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters
reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration
* add minimum number of tensor dimensions to apply weight decay (default 2)
this allows to not apply weight decay to bias parameters
* rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup
* fix increase of model.train_samples and model.train_tokens
now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations
* change sampling parameters for prediction after training to defaults of common.h
and clarify what is context for prediction and what are generated tokens
* tighten abs error bounds for cross_entropy_loss in test-grad0
* add conditional compilation of using F16 exp in flash attention
uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention
* tighten abs error bounds for flash_attn in test-grad0
* tighten abs error bounds for sqrt in test-grad0
* remove out-commented vectorized code of opt_adam
the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead
* ggml : update ggml_rms_norm_back with configurable eps
* llama training : fix ggml_rms_norm_back calls to pass configurable eps
* remove trailing whitespace
* add train function using automatic gradient checkpointing backward pass and allocator
* in train function replace add_inplace by regular add
because using add_inplace seems to result in different gradients
* don't use allocate hash_map on context
because the context has no_alloc=True when using memory allocator resulting in NULL data pointers
* correctly clone reshape and permute operations by also cloning tensor->nb values
* fix variable name and add missing type cast
* terminate recursive tensor cloning when reaching tensor without src tensors
* correctly clone view tensors by setting data pointers
without this the checkpointing would only work when being used together with memory allocator
* fix variable names
* swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn`
* add input tensors as checkpoints
so that recursive tensor cloning of gradient checkpointing terminates on input tensors
* fix variable name and add missing boolean negation
* make sure some tensors are not reallocated by inserting new temporary nodes depending on them:
output and parameter gradient tensors need to be available at the end of the graph execution
parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration
checkpoint tensors are allocated all together to reduce memory allocator fragmentation
afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs
* fix ASSERT to work with zero layers
* add training options whether to use allocator and/or unified training function
* integrate unified training function which may use memory allocator
the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing
* format name of cloned tensors with " (clone)" suffix
* set names for tensors in unified train function for easier debugging
* allocate graph on context using ggml_new_graph
* remove handwritten training functions
* remove unused training parameters "use_scratch" and "use_unified"
* remove trailing whitespace
* remove unused train params: mem_compute1_gb & mem_compute2_gb
mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions
mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented)
* remove unused forward_batch function
* add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly
* only use ggml_allocr_alloc when tensor has NULL data and is no view
* fix test when to create temporary backward graph
temporary backward graph is only necessary when using checkpointing
* fix memory "leak" in optimizers
each iteration a new cplan with new memory for work data was allocated.
now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data.
* reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator
with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory.
the computation results are the same
* add API functions to access llama model tensors
* add stub example for finetuning, based on train-text-from-scratch
* move and remove code
* add API functions to access remaining model parameters:
mult, head and rot
* first draft for LORA finetune training
* remove const model and layer arguments in API functions for accessing model tensors
* bug fixes to make finetune compile
automatic allocator does not work yet
* add debug prints for training memory improvements
* fix names of lora tensors
* avoid stack overflow resulting from big ggml_cgraph
replace stack allocation and ggml_build_forward by ggml_new_graph in combination with ggml_build_forward_expand
* replace llama API functions to get model tensors by one function to get model tensor by name
LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name);
* remove unused call to not existing llama_get_layer_from_model
* implement ggml_compute_forward_out_prod_q_f32
* remove trailing whitespace
* add lora finetune support on quantized base model tensors
* add ggml_add_cast API function
this function works like ggml_add, but accepts a data type for the resulting tensor.
only supported for quantized src0 input.
* use ggml_add_cast in finetuning
lora-applied weights will now have data type F32, which improves gradients when finetuning quantized base models
* bug fix: actually use result type passed to ggml_add_cast
* make sure base model tensors data cannot be used in viewable operations
memory allocator would try to make lora application inplace on base model tensors.
since those are memory mapped this will result in memory access violations
* fix bug in ggml_out_prod which resulted in wrong n_dims of result tensors
* avoid keeping in memory ALL of the gradients
The problem here stems from ggml_graph_reset. This function is called in the optimization function, before each graph computation, to reset the gradients to zero. This required a unique memory slot for each gradient: allocating memory from a previosly freed memory location might lead to non-zero input gradients.
During ggml_compute_backward the gradients are build stepwise by adding or substracting new values, starting from a OP_NONE tensor which needs to contain zero-values. This requires the graph reset.
To avoid this I now remember in ggml_build_backward_expand the original OP_NONE gradient tensors in a hash table, which is passed to ggml_compute_backward. There instead of using add (or sub or similar) I test whether the existing gradient to be changed is a zero-valued-tensor by looking up its existence in the hash table. When it is such a zero-tensor it will not be modified, but replaced by the value to be added, otherwise the regular add (not inplace, allocator will take care of this) will be used. This way none of those zero-tensor values will be necessary in the final backward graph and more importantly they won't need a unique memory slot, just to make them zero.
* remove trailing whitespace
* remove debug prints and function to compute tensor data hash
* improve optimization iteration prints
* adjust maximal values to support finetuning 3B models
* change default finetune params lora_r and lora_alpha to match the n_rank parameters of 4
* bug fix: make sure finetune input gradient is allocated at begin and kept until end
* remove unnecessary src tensor from ggml_get_rows_back
we don't need data of src[2] for computation, only to setup the correct output shape.
remove dependency on src[2], so that allocator can work more freely.
the computational graph is still completely determined, because the output shape is naturally included.
this is similar to how ggml_reshape does it.
* remove unnecessary src tensor from ggml_repeat & ggml_repeat_back
we don't need data of src[1] for computation, only to setup the correct output shape.
remove dependency on src[1], so that allocator can work more freely.
the computational graph is still completely determined, because the output shape is naturally included
* resolve todo
allocator will only make it inplace when they are of the same type
* mixing multiple LORA adapters is now possible
pass more than one '--lora FNAME' argument to apply more than one LORA.
use '--lora-scaled FNAME S' when you want to specify a user-defined scale for an adapter.
* add option to save finetune output every N iterations
* also save latest finetune output with ITERATION="LATEST" and print where files are saved
saving with LATEST makes it easier to resume training from the latest checkpoint
the string "LATEST" can be configured with command line option "--fn-latest STR"
* update checkpoint train stats before saving via "--save-every"
* add command line option `--rank-wo N` for rank of wo tensor
* update finetune README
* fix dump_non_result_info_yaml to output multiple lora adapters
* bug fix: replace GGML_TYPE_SIZE[t] by ggml_type_size(t)
* replace llama_n_mult by llama_n_ff
* finetune bug fixes to compile with merged in code from master
* remove prediction related code to reduce duplicated code with main
use main instead
* reduce large memory overhead in train-text-from-scratch
all gradients had to be pinned so that graph_reset works correctly.
this is no longer necessary with the changes to ggml_compute_backward introduced in this PR.
* add comment explaining why finetune checkpoints are allocated in one block
* make default value of float member a float literal
* handle rms_norm and rope parameters the same as in train-text-from-scratch
* remove unused code
* remove vocab related code as it is unnecessary
* add LLM_KV_TRAINING_TYPE to train-text-from-scratch checkpoints
so that they can be differentiated from lora finetune checkpoints
* add gguf constants and load/save functions from train-text-from-scratch
* add load & save lora finetune checkpoints via gguf
* add python script to convert old finetune checkpoint files to gguf
* remove old checkpoint save & load code
* remove code to print data checksums which was used to verify correctness of new gguf code
* omit tokenization when training is disabled, only save llama lora adapter
training can be disabled by passing '-n 0' to finetune
* remove trailing whitespace
* update README.md
* implement ggml_compute_forward_repeat_f16
* avoid stack overflow of large cgraphs in test-grad0
* add ggml API functions ggml_unravel_index, ggml_get_i32_nd and its analogs for set and for f32
ggml_get_i32_1d, ggml_set_i32_1d, ggml_get_f32_1d, ggml_set_f32_1d now support non-contiguous tensors.
in case of non-contiguous tensor, the 1d index is unraveled into a multi index using ggml_unravel_index to be passed to '_nd' function equivalent.
this fixes a bug in test-grad0 which happens due to ggml_build_backward not building purely contiguous tensors anymore
* increase test-grad0 context mem size to accommodate for bigger cgraph
* add sanity check to ggml_compute_backward, asserting the correct shape of gradients
* fix ggml_acc_or_set to return tensor of correct shape
* remove unused 'inplace' argument from ggml_compute_backward function
inplace operations to add gradients are no longer created by ggml_compute_backward
use allocator to automatically make inplace operations
* add missing argument 'int i0' to ggml_get_i32_nd & ggml_set_i32_nd header declarations
* fix error message in ggml_allocr_alloc to display actual max_avail
* fix check_gradient
ggml_build_backward_expand was previously replaced by ggml_build_backward, but the assignment of forward graph to backward graph missing
* use tensor->view_src instead of ggml_is_view and get_view_source
* move gradient checkpointing code into ggml, new API function:
// build gradient checkpointing backward graph gb for gf using provided checkpoints
// gb_tmp will contain original backward graph with rewritten backward process nodes,
// but without the second forward pass nodes.
GGML_API void ggml_build_backward_gradient_checkpointing(
struct ggml_context * ctx,
struct ggml_cgraph * gf,
struct ggml_cgraph * gb,
struct ggml_cgraph * gb_tmp,
struct ggml_tensor * * checkpoints,
int n_checkpoints);
* replace custom data getters and setters by ggml functions
* train-text-from-scratch can train (full finetune) gguf models
just pass the gguf model via `--checkpoint-in FN`.
after this, to continue training, pass the generated checkpoint instead of the original gguf model.
tested with smaller models, bigger models may exceed available memory.
use (LORA) finetune for those.
* remove trailing whitespace
* add option to save train-text-from-scratch output every N iterations
* update README.md
* fix warnings
* fix warnings
* remove finetune option to disable allocator
the allocator should always be used.
by making sure that it is always used it gets easier to implement automatic memory requirements computation
* add tensor checkpoints only when gradient checkpointing is enabled
* initialize opt ggml context if none was provided
* add ggml-alloc API function 'ggml_allocr_max_size' to get max size of alloc
GGML_API size_t ggml_allocr_max_size(struct ggml_allocr * alloc);
* finetune: automatically allocate all memory and changes to command line options
remove '--n_examples N' parameter, as it no longer makes sense to call optimization process multiple times in a loop.
add '--only_write_lora' command line option: will skip tokenization and training, to only write a llama.cpp comptabile LORA adapter.
remove memory buffer related command line options.
improve iteration console output.
* add finetune to Makefile
* update README.md
* print time per iteration and estimate remaining time
* increase measured alloc size by tensor_alignment
ggml_allocr_reset will reduce the given size by up to tensor_alignment-1
* fix README.md
* add some more allocator debug prints
* bug fix, probably solves the 'ggml_allocr_alloc: not enough space in the buffer' issue
* revert last commit
"bug fix, probably solves the 'ggml_allocr_alloc: not enough space in the buffer' issue"
"alloc was freeing an externally allocated tensor, because it calculated the end of allocator memory as alloc->data + alloc->max_size instead of alloc->data + alloc->size."
This is intentional to reduce the risk of freeing external tensors when measuring. Unless max_size is not properly calculated, I don't see why this is an issue.
* remove unnecessary "0x" before "%p" output
* move measurement memory segment to upper region of the address space
* update README.md
* fix printf format warnings
* add missing gguf_free in load_checkpoint_lora_file
* load default rms_norm and rope parameters from base model
* add gradient accumulation
specify number accumulation steps with '--grad-acc N'.
this will simulate a bigger batch size of grad_acc*batch.
* fix tracking of train_samples and train_tokens
* build : fix compile warnings
* ggml : fix L-BFGS linesearch loop
* improve finetune time measurement
fix printf warnings on system where int64_t is (long int).
change time datatypes to double because values get big with long training times.
exclude file saving from time measurement.
converge faster to actual time per iteration by removing very small first duration before first iteration was performed.
fix bug in output of total training time, the reported value was 1000 times to small.
* specify default lora rank with '--lora-r N'
'--lora-r N' will specify default rank for all tensors
'--rank-wq N', etc. will override this default rank for specific tensor types.
* fix gradient accumulation bug where the same batch was used for each microstep
* fix gradient accumulation bug where the same batch was used for each microstep
* support grouped-query-attention in ggml_flash_attn and ggml_flash_attn_back
k and v can now be repeated in q along ne[2]
in forward pass just use modulo to compute k and v indices, like ik2 = iq2 % nek2.
in backard pass this won't work as easy, because multiple threads will compete to accumulate to the same k->grad[:,ik1,ik2,ik3] and v->grad[:,iv1,iv2,iv3].
so we change the parallelization over q rows to be over k rows. this ensures non-overlapping (ik2,ik3) across threads.
in each thread we then iterate over the number of repetitions of k/v in q to compute iq2 as iq2 = ik2 + irep*nek2.
since ne2 is not the same for q,k and v we also change how the gradients are concatenated into the result tensor.
additionally the offsets of gradq, gradk and gradv in the result tensor are now memory aligned.
we also simplify the compute_backward part of flash_attn to use ggml_reshape instead of switching over the number of dimensions.
this needs a small change to ggml_reshape, removing the assertion of second argument to be contiguous.
since only the shape (ne) of the second reshape argument is of relevance, its memory layout (nb) is irrelevant -> it can very well be non-contiguous.
change test-grad0 to also test for repeated k/v in q.
this changes the rng and now results in small gradient differences in softmax. these solely come from using f16 exp table lookup in forward softmax: when temporarily changing softmax to use actual exp function, the reported gradient differences go away. gradient differences coming solely from f16 table lookup are acceptable.
added a note to explain this.
* add llama API functions to get grouped-query-attention n_head parameter 'n_head_kv'.
* fix finetune to support grouped-query-attention (using flash-attention)
note: ggml changes to ggml_out_prod are necessary to support grouped-query-attention without flash-attention.
* support broadcastable a in out_prod(a, b) and backward pass of broadcasting mul_mat(a, b)
* test broadcasting mul_mat backward pass
* decouple random number generator of each operation test
when changing one test the rng of others tests is not influenced anymore
* add comment briefly describing what ggml_repeat_back does
* simplify broadcasting mul_mat backward using ggml_repeat_back
* add cgraph evaluation order member and corresponding enum type
this controls in which order ggml_build_forward visits source nodes.
by default the nodes are visited left to right, i.e. src[0] first.
in some cases it is beneficial for ggml-alloc to visit in a different order.
two possible orders are supported: left-to-right (src[0] first) and right-to-left (src[0] last).
* measure max compute size for each cgraph eval order and use best order
this can bring huge memory savings:
e.g. codellama-34b with n_ctx=64, n_batch=1 goes from 92927.8mb down to 4627.6 MB
* remove unused command line options
* add sample start patterns and options to force new or by default resume last shuffling
* update shuffle rng state on reshuffle
* exclude known zero values from computations in flash_attn_f32 & flash_attn_back_f32
* remove probably unnecessary exception type flags from stringstream
* pass correct max number of tokens to llama_tokenize
* account for possible leading whitespace that will be added by tokenizer
e.g. '\t' will be tokenized by llama spm tokenizer to [29871, 12]
* use unrolled vec_mad in out_prod
y is vec_mad result vec.
x is vec_mad input vec.
v is vec_mad input scalar.
ggml_vec_mad_f32_unroll will internally loop over x and v with same y.
GGML_VEC_MAD_UNROLL is by default defined to 32.
This value is empirical optimized using performance test runs of out-prod in openllama-3b finetune with 256 context length and batch size 1. It gives 23% performance boost for out_prod.
Full measurements of out-prod runtime in ms:
unroll_xv unroll_yv
1 67014.643 87826.469
2 77117.552 89077.656
4 72091.311 109121.657
8 61077.543 88678.334
16 56914.67 79514.947
24 59024.595 84350.254
28 55952.446 83368.73
32 51476.658 85177.745
36 55973.792 84659.92
40 55139.616 93844.738
48 60736.392 93330.267
64 99856.878 116994.99
Second column is when unrollying yv instead of xv
* set lora_alpha to value of lora_r if it is not set via command line
otherwise only changing lora_r will change scaling of lora adapter used in prediction
* reshuffle original sample order instead of the previous shuffled order
otherwise resumed reshuffle will not result in same sample order
* block tiling for out-prod inspired by mul-mat
block sizes are empirically optimized
roughly doubles the flops of out-prod
* exclude some more known zero values from computations in flash_attn_f32 & flash_attn_back_f32
* add static keywords
* remove outcommented old code
* update train-text-from-scratch with tokenization, sample selection and shuffling from finetune
* remove lbfgs related train parameters
* move common train functions into common/train.[h|cpp]
* move train state into struct train_state
* move train data saving code into callback to unify code of opt_callback
train_params are still different in finetune and train-text-from-scratch, so it can't yet be moved to train.h|cpp
* move common train params into common/train
* move common opt_callback into common/train
* fix consume_common_train_arg
* save and load head_count_kv in lora checkpoints
* increase train_samples by used_samples instead of number of batches
on batch can contain more than one sample when option "fill_with_next_samples" is used
* fix usage of llama_tokenize
* remove static from process_escape since we need it exposed in header
* fix code formating of long function declarations
* fix condition in load_train_state_gguf
* use die("msg") instead of replace GGML_ASSERT(!"msg") or throw std::runtime_error("msg")
* fix saving and loading of training type
* remove terminating '\0' from tokenization
(llama_tokenize is now passed the string length instead of relying on terminating '\0')
* fix compile warnings
* fix compile warnings
* use new/delete for train_state instead of malloc/free
using malloc may result in seg faults when trying to assign string fields
* assert that sample_count > 0, avoiding division by zero
* fix frand to return value in interval [0,1)
* add train option "--sample-random-offsets"
Use samples beginning at random offsets.
The offset is only applied to the first sample in each batch context window.
Together with "--fill-with-next-samples" this may help for training endless text generation.
For example given a dataset containing samples "abcd", "ABCD", "0123".
With context size of 8 and options "--fill-with-next-samples", "--no-separate-with-eos", "--no-separate-with-bos",
the context windows of batches could only be filled with "abcdABCD", "ABCDabcd", "0123abcd", etc.
With "--sample-random-offsets" it can also be filled with "23abcdAB", "bcd0123A", etc.
* deduplicate code into function
* remove n_rot hparam, as it must always be hparam.n_embd_head()
* align code
* assert correct base model tensor shapes
* move some params from lora hparams into model hparams and load model params from gguf
this equalizes the model definition in finetune and text-from-scratch and removes the need for additional llama api functions to get model parameters
* remove now unnecessary llama API functions to get model params that where added by this PR
* train-text-from-scratch: automatically allocate model tensors, remove option '--mem-model N'
* train-text-from-scratch: automatically allocate opt context
* train-text-from-scratch: automatically allocate input tensors
* train-text-from-scratch: automatically allocate compute memory
* remove unused options and equalize train-text-from-scratch with finetune
* initialize opt->loss_after with zero
* add export-lora program
* remove trailing whitespace
* add export-lora build in Makefile
* remove unused struct tensor_info from export-lora
* add export-lora build dependency to llama
because it depends on common, which depends on llama
* update finetune README.md
* cancel optimization when specified number of epochs is completed
* improve handling of export-lora arguments
print errors and warnings when files could not be read or created
* Fix export-lora.cpp "not enough space in the context's memory pool" (#1)
* Fix export-lora.cpp "not enough space in the context's memory pool"
Without this patch, export-lora would sometimes error with "not enough space in the context's memory pool (needed 656784, available 656800)".
* increase required context size by 5*GGML_MEM_ALIGN instead of plain 16
---------
Co-authored-by: xaedes <xaedes@gmail.com>
* improve handling of not yet supported tensor types
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: meatbag-18a <145869052+meatbag-18a@users.noreply.github.com>
2023-09-28 20:40:11 +02:00
float scale ,
2023-06-24 10:47:58 +02:00
const char * path_base_model ,
2024-01-02 15:15:16 +01:00
int32_t n_threads ) ,
2023-09-28 18:04:36 +02:00
" use llama_model_apply_lora_from_file instead " ) ;
2023-06-24 10:47:58 +02:00
2024-01-02 15:15:16 +01:00
LLAMA_API int32_t llama_model_apply_lora_from_file (
2023-06-24 10:47:58 +02:00
const struct llama_model * model ,
train : finetune LORA (#2632)
* fix track_max_mem in forward_batch_wo_cache_flash_attn_train
* remove unnecessary Adam(W) optimizer tensors.
reduces optimizer memory overhead from 7*modelsize to 2*modelsize.
additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t.
bumps training checkpoint file version, but old checkpoints can still be read.
new version with less tensors is saved.
* add gradient clipping to AdamW
* Fix reset of unused g->nodes and g->grads to NULL
* implement gradient checkpointing for training
reduces memory overhead from O(n_layer) to O(sqrt(n_layer))
as explained in readme of https://github.com/cybertronai/gradient-checkpointing
* remove unused compute buffer 3
* add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes
GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep);
* change AdamW decay parameter to work like the torch AdamW decay parameter
It is now relative to Adam learning rate `alpha*sched`.
Before that it was relative to `sched` only.
`alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1]
* change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT
* change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW
btw: the default weight decay parameter for torch.optim.AdamW is 0.01
* bug fixes for cross entropy loss
ggml_cross_entropy_loss: sums where not correctly added in workload of each thread
ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues
guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16
cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup.
so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance.
* fix test-grad0 for cross_entropy_loss
the second argument to cross_entropy_loss must sum up to 1 for each row
* fix test-grad0 for soft_max
dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work
instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0)
* improve finite differences of test-grad0 by using double instead of float
* change cross_entropy_loss to output average over all rows
this helps keeping the loss and gradients in a sane range
* improve gradient checkpointing
sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal.
since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different:
```
given: n, u, v
objective: minimize(a*u+b*v) where a*b=n, a>0, b>0
b=n/a
minimize(a*u+v*n/a)
diff(a*u+v*n/a, a) = u - (v*n/a)/a
diff(a*u+v*n/a, a) == 0
u - (v*n/a)/a == 0
u == v*n/(a*a)
u*a*a = v*n
a*a = v*n/u
a = sqrt(n*v/u)
```
this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage.
* disable gradient checkpointing debug output
* llama : fix rope usage in train-text-from-scratch after ChatGLM change
* add more training parameters:
--enable-restart N Only for Adam optimizer. Enable restarts of cos-decay
--disable-restart N Only for Adam optimizer. Disable restarts of cos-decay
--opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero.
--opt-delta N Maximum delta for delta convergence test. Disabled when <= zero.
--opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero.
--adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero.
--adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha
* replace memcpy with reshape operation so that the graph is not cut at the input
this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it
* remove unused function argument from get_example_targets_batch
* measure and print total training time
* add optimization callback to ggml_opt_resume_g
this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)).
can be used for dynamic learning schedule and setting input data for batches before each iteration
* use optimization callback in training
allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters
reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration
* add minimum number of tensor dimensions to apply weight decay (default 2)
this allows to not apply weight decay to bias parameters
* rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup
* fix increase of model.train_samples and model.train_tokens
now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations
* change sampling parameters for prediction after training to defaults of common.h
and clarify what is context for prediction and what are generated tokens
* tighten abs error bounds for cross_entropy_loss in test-grad0
* add conditional compilation of using F16 exp in flash attention
uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention
* tighten abs error bounds for flash_attn in test-grad0
* tighten abs error bounds for sqrt in test-grad0
* remove out-commented vectorized code of opt_adam
the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead
* ggml : update ggml_rms_norm_back with configurable eps
* llama training : fix ggml_rms_norm_back calls to pass configurable eps
* remove trailing whitespace
* add train function using automatic gradient checkpointing backward pass and allocator
* in train function replace add_inplace by regular add
because using add_inplace seems to result in different gradients
* don't use allocate hash_map on context
because the context has no_alloc=True when using memory allocator resulting in NULL data pointers
* correctly clone reshape and permute operations by also cloning tensor->nb values
* fix variable name and add missing type cast
* terminate recursive tensor cloning when reaching tensor without src tensors
* correctly clone view tensors by setting data pointers
without this the checkpointing would only work when being used together with memory allocator
* fix variable names
* swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn`
* add input tensors as checkpoints
so that recursive tensor cloning of gradient checkpointing terminates on input tensors
* fix variable name and add missing boolean negation
* make sure some tensors are not reallocated by inserting new temporary nodes depending on them:
output and parameter gradient tensors need to be available at the end of the graph execution
parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration
checkpoint tensors are allocated all together to reduce memory allocator fragmentation
afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs
* fix ASSERT to work with zero layers
* add training options whether to use allocator and/or unified training function
* integrate unified training function which may use memory allocator
the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing
* format name of cloned tensors with " (clone)" suffix
* set names for tensors in unified train function for easier debugging
* allocate graph on context using ggml_new_graph
* remove handwritten training functions
* remove unused training parameters "use_scratch" and "use_unified"
* remove trailing whitespace
* remove unused train params: mem_compute1_gb & mem_compute2_gb
mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions
mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented)
* remove unused forward_batch function
* add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly
* only use ggml_allocr_alloc when tensor has NULL data and is no view
* fix test when to create temporary backward graph
temporary backward graph is only necessary when using checkpointing
* fix memory "leak" in optimizers
each iteration a new cplan with new memory for work data was allocated.
now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data.
* reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator
with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory.
the computation results are the same
* add API functions to access llama model tensors
* add stub example for finetuning, based on train-text-from-scratch
* move and remove code
* add API functions to access remaining model parameters:
mult, head and rot
* first draft for LORA finetune training
* remove const model and layer arguments in API functions for accessing model tensors
* bug fixes to make finetune compile
automatic allocator does not work yet
* add debug prints for training memory improvements
* fix names of lora tensors
* avoid stack overflow resulting from big ggml_cgraph
replace stack allocation and ggml_build_forward by ggml_new_graph in combination with ggml_build_forward_expand
* replace llama API functions to get model tensors by one function to get model tensor by name
LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name);
* remove unused call to not existing llama_get_layer_from_model
* implement ggml_compute_forward_out_prod_q_f32
* remove trailing whitespace
* add lora finetune support on quantized base model tensors
* add ggml_add_cast API function
this function works like ggml_add, but accepts a data type for the resulting tensor.
only supported for quantized src0 input.
* use ggml_add_cast in finetuning
lora-applied weights will now have data type F32, which improves gradients when finetuning quantized base models
* bug fix: actually use result type passed to ggml_add_cast
* make sure base model tensors data cannot be used in viewable operations
memory allocator would try to make lora application inplace on base model tensors.
since those are memory mapped this will result in memory access violations
* fix bug in ggml_out_prod which resulted in wrong n_dims of result tensors
* avoid keeping in memory ALL of the gradients
The problem here stems from ggml_graph_reset. This function is called in the optimization function, before each graph computation, to reset the gradients to zero. This required a unique memory slot for each gradient: allocating memory from a previosly freed memory location might lead to non-zero input gradients.
During ggml_compute_backward the gradients are build stepwise by adding or substracting new values, starting from a OP_NONE tensor which needs to contain zero-values. This requires the graph reset.
To avoid this I now remember in ggml_build_backward_expand the original OP_NONE gradient tensors in a hash table, which is passed to ggml_compute_backward. There instead of using add (or sub or similar) I test whether the existing gradient to be changed is a zero-valued-tensor by looking up its existence in the hash table. When it is such a zero-tensor it will not be modified, but replaced by the value to be added, otherwise the regular add (not inplace, allocator will take care of this) will be used. This way none of those zero-tensor values will be necessary in the final backward graph and more importantly they won't need a unique memory slot, just to make them zero.
* remove trailing whitespace
* remove debug prints and function to compute tensor data hash
* improve optimization iteration prints
* adjust maximal values to support finetuning 3B models
* change default finetune params lora_r and lora_alpha to match the n_rank parameters of 4
* bug fix: make sure finetune input gradient is allocated at begin and kept until end
* remove unnecessary src tensor from ggml_get_rows_back
we don't need data of src[2] for computation, only to setup the correct output shape.
remove dependency on src[2], so that allocator can work more freely.
the computational graph is still completely determined, because the output shape is naturally included.
this is similar to how ggml_reshape does it.
* remove unnecessary src tensor from ggml_repeat & ggml_repeat_back
we don't need data of src[1] for computation, only to setup the correct output shape.
remove dependency on src[1], so that allocator can work more freely.
the computational graph is still completely determined, because the output shape is naturally included
* resolve todo
allocator will only make it inplace when they are of the same type
* mixing multiple LORA adapters is now possible
pass more than one '--lora FNAME' argument to apply more than one LORA.
use '--lora-scaled FNAME S' when you want to specify a user-defined scale for an adapter.
* add option to save finetune output every N iterations
* also save latest finetune output with ITERATION="LATEST" and print where files are saved
saving with LATEST makes it easier to resume training from the latest checkpoint
the string "LATEST" can be configured with command line option "--fn-latest STR"
* update checkpoint train stats before saving via "--save-every"
* add command line option `--rank-wo N` for rank of wo tensor
* update finetune README
* fix dump_non_result_info_yaml to output multiple lora adapters
* bug fix: replace GGML_TYPE_SIZE[t] by ggml_type_size(t)
* replace llama_n_mult by llama_n_ff
* finetune bug fixes to compile with merged in code from master
* remove prediction related code to reduce duplicated code with main
use main instead
* reduce large memory overhead in train-text-from-scratch
all gradients had to be pinned so that graph_reset works correctly.
this is no longer necessary with the changes to ggml_compute_backward introduced in this PR.
* add comment explaining why finetune checkpoints are allocated in one block
* make default value of float member a float literal
* handle rms_norm and rope parameters the same as in train-text-from-scratch
* remove unused code
* remove vocab related code as it is unnecessary
* add LLM_KV_TRAINING_TYPE to train-text-from-scratch checkpoints
so that they can be differentiated from lora finetune checkpoints
* add gguf constants and load/save functions from train-text-from-scratch
* add load & save lora finetune checkpoints via gguf
* add python script to convert old finetune checkpoint files to gguf
* remove old checkpoint save & load code
* remove code to print data checksums which was used to verify correctness of new gguf code
* omit tokenization when training is disabled, only save llama lora adapter
training can be disabled by passing '-n 0' to finetune
* remove trailing whitespace
* update README.md
* implement ggml_compute_forward_repeat_f16
* avoid stack overflow of large cgraphs in test-grad0
* add ggml API functions ggml_unravel_index, ggml_get_i32_nd and its analogs for set and for f32
ggml_get_i32_1d, ggml_set_i32_1d, ggml_get_f32_1d, ggml_set_f32_1d now support non-contiguous tensors.
in case of non-contiguous tensor, the 1d index is unraveled into a multi index using ggml_unravel_index to be passed to '_nd' function equivalent.
this fixes a bug in test-grad0 which happens due to ggml_build_backward not building purely contiguous tensors anymore
* increase test-grad0 context mem size to accommodate for bigger cgraph
* add sanity check to ggml_compute_backward, asserting the correct shape of gradients
* fix ggml_acc_or_set to return tensor of correct shape
* remove unused 'inplace' argument from ggml_compute_backward function
inplace operations to add gradients are no longer created by ggml_compute_backward
use allocator to automatically make inplace operations
* add missing argument 'int i0' to ggml_get_i32_nd & ggml_set_i32_nd header declarations
* fix error message in ggml_allocr_alloc to display actual max_avail
* fix check_gradient
ggml_build_backward_expand was previously replaced by ggml_build_backward, but the assignment of forward graph to backward graph missing
* use tensor->view_src instead of ggml_is_view and get_view_source
* move gradient checkpointing code into ggml, new API function:
// build gradient checkpointing backward graph gb for gf using provided checkpoints
// gb_tmp will contain original backward graph with rewritten backward process nodes,
// but without the second forward pass nodes.
GGML_API void ggml_build_backward_gradient_checkpointing(
struct ggml_context * ctx,
struct ggml_cgraph * gf,
struct ggml_cgraph * gb,
struct ggml_cgraph * gb_tmp,
struct ggml_tensor * * checkpoints,
int n_checkpoints);
* replace custom data getters and setters by ggml functions
* train-text-from-scratch can train (full finetune) gguf models
just pass the gguf model via `--checkpoint-in FN`.
after this, to continue training, pass the generated checkpoint instead of the original gguf model.
tested with smaller models, bigger models may exceed available memory.
use (LORA) finetune for those.
* remove trailing whitespace
* add option to save train-text-from-scratch output every N iterations
* update README.md
* fix warnings
* fix warnings
* remove finetune option to disable allocator
the allocator should always be used.
by making sure that it is always used it gets easier to implement automatic memory requirements computation
* add tensor checkpoints only when gradient checkpointing is enabled
* initialize opt ggml context if none was provided
* add ggml-alloc API function 'ggml_allocr_max_size' to get max size of alloc
GGML_API size_t ggml_allocr_max_size(struct ggml_allocr * alloc);
* finetune: automatically allocate all memory and changes to command line options
remove '--n_examples N' parameter, as it no longer makes sense to call optimization process multiple times in a loop.
add '--only_write_lora' command line option: will skip tokenization and training, to only write a llama.cpp comptabile LORA adapter.
remove memory buffer related command line options.
improve iteration console output.
* add finetune to Makefile
* update README.md
* print time per iteration and estimate remaining time
* increase measured alloc size by tensor_alignment
ggml_allocr_reset will reduce the given size by up to tensor_alignment-1
* fix README.md
* add some more allocator debug prints
* bug fix, probably solves the 'ggml_allocr_alloc: not enough space in the buffer' issue
* revert last commit
"bug fix, probably solves the 'ggml_allocr_alloc: not enough space in the buffer' issue"
"alloc was freeing an externally allocated tensor, because it calculated the end of allocator memory as alloc->data + alloc->max_size instead of alloc->data + alloc->size."
This is intentional to reduce the risk of freeing external tensors when measuring. Unless max_size is not properly calculated, I don't see why this is an issue.
* remove unnecessary "0x" before "%p" output
* move measurement memory segment to upper region of the address space
* update README.md
* fix printf format warnings
* add missing gguf_free in load_checkpoint_lora_file
* load default rms_norm and rope parameters from base model
* add gradient accumulation
specify number accumulation steps with '--grad-acc N'.
this will simulate a bigger batch size of grad_acc*batch.
* fix tracking of train_samples and train_tokens
* build : fix compile warnings
* ggml : fix L-BFGS linesearch loop
* improve finetune time measurement
fix printf warnings on system where int64_t is (long int).
change time datatypes to double because values get big with long training times.
exclude file saving from time measurement.
converge faster to actual time per iteration by removing very small first duration before first iteration was performed.
fix bug in output of total training time, the reported value was 1000 times to small.
* specify default lora rank with '--lora-r N'
'--lora-r N' will specify default rank for all tensors
'--rank-wq N', etc. will override this default rank for specific tensor types.
* fix gradient accumulation bug where the same batch was used for each microstep
* fix gradient accumulation bug where the same batch was used for each microstep
* support grouped-query-attention in ggml_flash_attn and ggml_flash_attn_back
k and v can now be repeated in q along ne[2]
in forward pass just use modulo to compute k and v indices, like ik2 = iq2 % nek2.
in backard pass this won't work as easy, because multiple threads will compete to accumulate to the same k->grad[:,ik1,ik2,ik3] and v->grad[:,iv1,iv2,iv3].
so we change the parallelization over q rows to be over k rows. this ensures non-overlapping (ik2,ik3) across threads.
in each thread we then iterate over the number of repetitions of k/v in q to compute iq2 as iq2 = ik2 + irep*nek2.
since ne2 is not the same for q,k and v we also change how the gradients are concatenated into the result tensor.
additionally the offsets of gradq, gradk and gradv in the result tensor are now memory aligned.
we also simplify the compute_backward part of flash_attn to use ggml_reshape instead of switching over the number of dimensions.
this needs a small change to ggml_reshape, removing the assertion of second argument to be contiguous.
since only the shape (ne) of the second reshape argument is of relevance, its memory layout (nb) is irrelevant -> it can very well be non-contiguous.
change test-grad0 to also test for repeated k/v in q.
this changes the rng and now results in small gradient differences in softmax. these solely come from using f16 exp table lookup in forward softmax: when temporarily changing softmax to use actual exp function, the reported gradient differences go away. gradient differences coming solely from f16 table lookup are acceptable.
added a note to explain this.
* add llama API functions to get grouped-query-attention n_head parameter 'n_head_kv'.
* fix finetune to support grouped-query-attention (using flash-attention)
note: ggml changes to ggml_out_prod are necessary to support grouped-query-attention without flash-attention.
* support broadcastable a in out_prod(a, b) and backward pass of broadcasting mul_mat(a, b)
* test broadcasting mul_mat backward pass
* decouple random number generator of each operation test
when changing one test the rng of others tests is not influenced anymore
* add comment briefly describing what ggml_repeat_back does
* simplify broadcasting mul_mat backward using ggml_repeat_back
* add cgraph evaluation order member and corresponding enum type
this controls in which order ggml_build_forward visits source nodes.
by default the nodes are visited left to right, i.e. src[0] first.
in some cases it is beneficial for ggml-alloc to visit in a different order.
two possible orders are supported: left-to-right (src[0] first) and right-to-left (src[0] last).
* measure max compute size for each cgraph eval order and use best order
this can bring huge memory savings:
e.g. codellama-34b with n_ctx=64, n_batch=1 goes from 92927.8mb down to 4627.6 MB
* remove unused command line options
* add sample start patterns and options to force new or by default resume last shuffling
* update shuffle rng state on reshuffle
* exclude known zero values from computations in flash_attn_f32 & flash_attn_back_f32
* remove probably unnecessary exception type flags from stringstream
* pass correct max number of tokens to llama_tokenize
* account for possible leading whitespace that will be added by tokenizer
e.g. '\t' will be tokenized by llama spm tokenizer to [29871, 12]
* use unrolled vec_mad in out_prod
y is vec_mad result vec.
x is vec_mad input vec.
v is vec_mad input scalar.
ggml_vec_mad_f32_unroll will internally loop over x and v with same y.
GGML_VEC_MAD_UNROLL is by default defined to 32.
This value is empirical optimized using performance test runs of out-prod in openllama-3b finetune with 256 context length and batch size 1. It gives 23% performance boost for out_prod.
Full measurements of out-prod runtime in ms:
unroll_xv unroll_yv
1 67014.643 87826.469
2 77117.552 89077.656
4 72091.311 109121.657
8 61077.543 88678.334
16 56914.67 79514.947
24 59024.595 84350.254
28 55952.446 83368.73
32 51476.658 85177.745
36 55973.792 84659.92
40 55139.616 93844.738
48 60736.392 93330.267
64 99856.878 116994.99
Second column is when unrollying yv instead of xv
* set lora_alpha to value of lora_r if it is not set via command line
otherwise only changing lora_r will change scaling of lora adapter used in prediction
* reshuffle original sample order instead of the previous shuffled order
otherwise resumed reshuffle will not result in same sample order
* block tiling for out-prod inspired by mul-mat
block sizes are empirically optimized
roughly doubles the flops of out-prod
* exclude some more known zero values from computations in flash_attn_f32 & flash_attn_back_f32
* add static keywords
* remove outcommented old code
* update train-text-from-scratch with tokenization, sample selection and shuffling from finetune
* remove lbfgs related train parameters
* move common train functions into common/train.[h|cpp]
* move train state into struct train_state
* move train data saving code into callback to unify code of opt_callback
train_params are still different in finetune and train-text-from-scratch, so it can't yet be moved to train.h|cpp
* move common train params into common/train
* move common opt_callback into common/train
* fix consume_common_train_arg
* save and load head_count_kv in lora checkpoints
* increase train_samples by used_samples instead of number of batches
on batch can contain more than one sample when option "fill_with_next_samples" is used
* fix usage of llama_tokenize
* remove static from process_escape since we need it exposed in header
* fix code formating of long function declarations
* fix condition in load_train_state_gguf
* use die("msg") instead of replace GGML_ASSERT(!"msg") or throw std::runtime_error("msg")
* fix saving and loading of training type
* remove terminating '\0' from tokenization
(llama_tokenize is now passed the string length instead of relying on terminating '\0')
* fix compile warnings
* fix compile warnings
* use new/delete for train_state instead of malloc/free
using malloc may result in seg faults when trying to assign string fields
* assert that sample_count > 0, avoiding division by zero
* fix frand to return value in interval [0,1)
* add train option "--sample-random-offsets"
Use samples beginning at random offsets.
The offset is only applied to the first sample in each batch context window.
Together with "--fill-with-next-samples" this may help for training endless text generation.
For example given a dataset containing samples "abcd", "ABCD", "0123".
With context size of 8 and options "--fill-with-next-samples", "--no-separate-with-eos", "--no-separate-with-bos",
the context windows of batches could only be filled with "abcdABCD", "ABCDabcd", "0123abcd", etc.
With "--sample-random-offsets" it can also be filled with "23abcdAB", "bcd0123A", etc.
* deduplicate code into function
* remove n_rot hparam, as it must always be hparam.n_embd_head()
* align code
* assert correct base model tensor shapes
* move some params from lora hparams into model hparams and load model params from gguf
this equalizes the model definition in finetune and text-from-scratch and removes the need for additional llama api functions to get model parameters
* remove now unnecessary llama API functions to get model params that where added by this PR
* train-text-from-scratch: automatically allocate model tensors, remove option '--mem-model N'
* train-text-from-scratch: automatically allocate opt context
* train-text-from-scratch: automatically allocate input tensors
* train-text-from-scratch: automatically allocate compute memory
* remove unused options and equalize train-text-from-scratch with finetune
* initialize opt->loss_after with zero
* add export-lora program
* remove trailing whitespace
* add export-lora build in Makefile
* remove unused struct tensor_info from export-lora
* add export-lora build dependency to llama
because it depends on common, which depends on llama
* update finetune README.md
* cancel optimization when specified number of epochs is completed
* improve handling of export-lora arguments
print errors and warnings when files could not be read or created
* Fix export-lora.cpp "not enough space in the context's memory pool" (#1)
* Fix export-lora.cpp "not enough space in the context's memory pool"
Without this patch, export-lora would sometimes error with "not enough space in the context's memory pool (needed 656784, available 656800)".
* increase required context size by 5*GGML_MEM_ALIGN instead of plain 16
---------
Co-authored-by: xaedes <xaedes@gmail.com>
* improve handling of not yet supported tensor types
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: meatbag-18a <145869052+meatbag-18a@users.noreply.github.com>
2023-09-28 20:40:11 +02:00
const char * path_lora ,
float scale ,
const char * path_base_model ,
2024-01-02 15:15:16 +01:00
int32_t n_threads ) ;
2023-04-17 17:28:55 +02:00
2023-09-28 18:04:36 +02:00
//
// KV cache
//
2023-11-23 18:07:56 +01:00
// Information associated with an individual cell in the KV cache view.
struct llama_kv_cache_view_cell {
// The position for this cell. Takes KV cache shifts into account.
// May be negative if the cell is not populated.
llama_pos pos ;
} ;
// An updateable view of the KV cache.
struct llama_kv_cache_view {
// Number of KV cache cells. This will be the same as the context size.
int32_t n_cells ;
// Maximum number of sequences that can exist in a cell. It's not an error
// if there are more sequences in a cell than this value, however they will
// not be visible in the view cells_sequences.
int32_t n_max_seq ;
// Number of tokens in the cache. For example, if there are two populated
// cells, the first with 1 sequence id in it and the second with 2 sequence
// ids then you'll have 3 tokens.
int32_t token_count ;
// Number of populated cache cells.
int32_t used_cells ;
// Maximum contiguous empty slots in the cache.
int32_t max_contiguous ;
// Index to the start of the max_contiguous slot range. Can be negative
// when cache is full.
int32_t max_contiguous_idx ;
// Information for an individual cell.
struct llama_kv_cache_view_cell * cells ;
// The sequences for each cell. There will be n_max_seq items per cell.
llama_seq_id * cells_sequences ;
} ;
// Create an empty KV cache view. (use only for debugging purposes)
LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init ( const struct llama_context * ctx , int32_t n_max_seq ) ;
// Free a KV cache view. (use only for debugging purposes)
LLAMA_API void llama_kv_cache_view_free ( struct llama_kv_cache_view * view ) ;
// Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes)
LLAMA_API void llama_kv_cache_view_update ( const struct llama_context * ctx , struct llama_kv_cache_view * view ) ;
// Returns the number of tokens in the KV cache (slow, use only for debug)
// If a KV cell has multiple sequences assigned to it, it will be counted multiple times
2024-01-02 15:15:16 +01:00
LLAMA_API int32_t llama_get_kv_cache_token_count ( const struct llama_context * ctx ) ;
2023-11-23 18:07:56 +01:00
// Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
2024-01-02 15:15:16 +01:00
LLAMA_API int32_t llama_get_kv_cache_used_cells ( const struct llama_context * ctx ) ;
2023-04-02 12:23:04 +02:00
2023-10-29 18:31:40 +01:00
// Clear the KV cache
LLAMA_API void llama_kv_cache_clear (
struct llama_context * ctx ) ;
2023-09-28 18:04:36 +02:00
// Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
2023-10-29 18:31:40 +01:00
// seq_id < 0 : match any sequence
// p0 < 0 : [0, p1]
// p1 < 0 : [p0, inf)
2023-09-28 18:04:36 +02:00
LLAMA_API void llama_kv_cache_seq_rm (
struct llama_context * ctx ,
llama_seq_id seq_id ,
llama_pos p0 ,
llama_pos p1 ) ;
// Copy all tokens that belong to the specified sequence to another sequence
// Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence
2023-10-03 20:04:01 +02:00
// p0 < 0 : [0, p1]
// p1 < 0 : [p0, inf)
2023-09-28 18:04:36 +02:00
LLAMA_API void llama_kv_cache_seq_cp (
struct llama_context * ctx ,
llama_seq_id seq_id_src ,
llama_seq_id seq_id_dst ,
llama_pos p0 ,
llama_pos p1 ) ;
// Removes all tokens that do not belong to the specified sequence
LLAMA_API void llama_kv_cache_seq_keep (
struct llama_context * ctx ,
llama_seq_id seq_id ) ;
// Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
// If the KV cache is RoPEd, the KV data is updated accordingly
2023-10-03 20:04:01 +02:00
// p0 < 0 : [0, p1]
// p1 < 0 : [p0, inf)
2023-09-28 18:04:36 +02:00
LLAMA_API void llama_kv_cache_seq_shift (
struct llama_context * ctx ,
llama_seq_id seq_id ,
llama_pos p0 ,
llama_pos p1 ,
llama_pos delta ) ;
2024-01-08 10:18:32 +01:00
// Integer division of the positions by factor of `d > 1`
// If the KV cache is RoPEd, the KV data is updated accordingly
// p0 < 0 : [0, p1]
// p1 < 0 : [p0, inf)
2024-01-08 10:14:04 +01:00
LLAMA_API void llama_kv_cache_seq_div (
struct llama_context * ctx ,
llama_seq_id seq_id ,
llama_pos p0 ,
llama_pos p1 ,
int d ) ;
2023-09-28 18:04:36 +02:00
//
// State / sessions
//
2023-04-26 22:08:43 +02:00
2023-05-03 04:26:13 +02:00
// Returns the maximum size in bytes of the state (rng, logits, embedding
// and kv_cache) - will often be smaller after compacting tokens
2023-05-01 09:24:20 +02:00
LLAMA_API size_t llama_get_state_size ( const struct llama_context * ctx ) ;
2023-04-22 08:21:32 +02:00
// Copies the state to the specified destination address.
// Destination needs to have allocated enough memory.
// Returns the number of bytes copied
2023-09-28 18:04:36 +02:00
LLAMA_API size_t llama_copy_state_data (
struct llama_context * ctx ,
uint8_t * dst ) ;
2023-04-22 08:21:32 +02:00
// Set the state reading from the specified address
// Returns the number of bytes read
2023-09-28 18:04:36 +02:00
LLAMA_API size_t llama_set_state_data (
struct llama_context * ctx ,
uint8_t * src ) ;
2023-04-22 08:21:32 +02:00
2023-04-28 17:59:37 +02:00
// Save/load session file
2023-09-28 18:04:36 +02:00
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 ) ;
2023-04-28 17:59:37 +02:00
2023-09-28 18:04:36 +02:00
//
// Decoding
//
// Run the llama inference to obtain the logits and probabilities for the next token(s).
2023-03-22 06:32:36 +01:00
// 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
2023-09-28 18:04:36 +02:00
// DEPRECATED: use llama_decode() instead
LLAMA_API DEPRECATED ( int llama_eval (
2023-03-22 06:32:36 +01:00
struct llama_context * ctx ,
2023-09-28 18:04:36 +02:00
llama_token * tokens ,
int32_t n_tokens ,
2024-01-02 15:15:16 +01:00
int32_t n_past ) ,
2023-09-28 18:04:36 +02:00
" use llama_decode() instead " ) ;
2023-06-28 17:53:37 +02:00
// Same as llama_eval, but use float matrix input directly.
2023-09-28 18:04:36 +02:00
// DEPRECATED: use llama_decode() instead
LLAMA_API DEPRECATED ( int llama_eval_embd (
2023-06-28 17:53:37 +02:00
struct llama_context * ctx ,
2023-09-28 18:04:36 +02:00
float * embd ,
int32_t n_tokens ,
2024-01-02 15:15:16 +01:00
int32_t n_past ) ,
2023-09-28 18:04:36 +02:00
" use llama_decode() instead " ) ;
2023-03-22 06:32:36 +01:00
2023-09-28 18:04:36 +02:00
// Return batch for single sequence of tokens starting at pos_0
//
// NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it
//
LLAMA_API struct llama_batch llama_batch_get_one (
llama_token * tokens ,
int32_t n_tokens ,
llama_pos pos_0 ,
llama_seq_id seq_id ) ;
2023-10-18 15:21:57 +02:00
// Allocates a batch of tokens on the heap that can hold a maximum of n_tokens
// Each token can be assigned up to n_seq_max sequence ids
2023-09-28 18:04:36 +02:00
// The batch has to be freed with llama_batch_free()
// If embd != 0, llama_batch.embd will be allocated with size of n_tokens * embd * sizeof(float)
// Otherwise, llama_batch.token will be allocated to store n_tokens llama_token
// The rest of the llama_batch members are allocated with size n_tokens
// All members are left uninitialized
LLAMA_API struct llama_batch llama_batch_init (
int32_t n_tokens ,
2023-10-18 15:21:57 +02:00
int32_t embd ,
int32_t n_seq_max ) ;
2023-09-28 18:04:36 +02:00
// Frees a batch of tokens allocated with llama_batch_init()
LLAMA_API void llama_batch_free ( struct llama_batch batch ) ;
// Positive return values does not mean a fatal error, but rather a warning.
// 0 - success
// 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
// < 0 - error
2024-01-02 15:15:16 +01:00
LLAMA_API int32_t llama_decode (
2023-09-28 18:04:36 +02:00
struct llama_context * ctx ,
2023-09-28 21:42:38 +02:00
struct llama_batch batch ) ;
// Set the number of threads used for decoding
// n_threads is the number of threads used for generation (single token)
// n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens)
LLAMA_API void llama_set_n_threads ( struct llama_context * ctx , uint32_t n_threads , uint32_t n_threads_batch ) ;
2023-06-04 22:34:30 +02:00
2023-08-21 22:07:43 +02:00
// Token logits obtained from the last call to llama_eval()
// The logits for the last token are stored in the last row
2023-09-28 18:04:36 +02:00
// Logits for which llama_batch.logits[i] == 0 are undefined
// Rows: n_tokens provided with llama_batch
2023-08-21 22:07:43 +02:00
// Cols: n_vocab
LLAMA_API float * llama_get_logits ( struct llama_context * ctx ) ;
2023-09-28 18:04:36 +02:00
// Logits for the ith token. Equivalent to:
// llama_get_logits(ctx) + i*n_vocab
LLAMA_API float * llama_get_logits_ith ( struct llama_context * ctx , int32_t i ) ;
2023-08-21 22:07:43 +02:00
// Get the embeddings for the input
// shape: [n_embd] (1-dimensional)
LLAMA_API float * llama_get_embeddings ( struct llama_context * ctx ) ;
2024-02-13 13:06:58 +01:00
// Get the embeddings for the ith sequence
// llama_get_embeddings(ctx) + i*n_embd
LLAMA_API float * llama_get_embeddings_ith ( struct llama_context * ctx , int32_t i ) ;
2023-08-21 22:07:43 +02:00
//
// Vocab
//
2023-10-23 21:40:03 +02:00
LLAMA_API const char * llama_token_get_text ( const struct llama_model * model , llama_token token ) ;
2023-08-21 22:07:43 +02:00
2023-10-23 21:40:03 +02:00
LLAMA_API float llama_token_get_score ( const struct llama_model * model , llama_token token ) ;
2023-08-21 22:07:43 +02:00
2023-10-23 21:40:03 +02:00
LLAMA_API enum llama_token_type llama_token_get_type ( const struct llama_model * model , llama_token token ) ;
2023-08-21 22:07:43 +02:00
// Special tokens
2023-10-23 21:40:03 +02:00
LLAMA_API llama_token llama_token_bos ( const struct llama_model * model ) ; // beginning-of-sentence
LLAMA_API llama_token llama_token_eos ( const struct llama_model * model ) ; // end-of-sentence
LLAMA_API llama_token llama_token_nl ( const struct llama_model * model ) ; // next-line
2023-11-17 03:14:37 +01:00
// Returns -1 if unknown, 1 for true or 0 for false.
2024-01-02 15:15:16 +01:00
LLAMA_API int32_t llama_add_bos_token ( const struct llama_model * model ) ;
2023-11-17 03:14:37 +01:00
// Returns -1 if unknown, 1 for true or 0 for false.
2024-01-02 15:15:16 +01:00
LLAMA_API int32_t llama_add_eos_token ( const struct llama_model * model ) ;
2023-11-17 03:14:37 +01:00
2023-10-02 09:42:02 +02:00
// codellama infill tokens
2023-10-23 21:40:03 +02:00
LLAMA_API llama_token llama_token_prefix ( const struct llama_model * model ) ; // Beginning of infill prefix
LLAMA_API llama_token llama_token_middle ( const struct llama_model * model ) ; // Beginning of infill middle
LLAMA_API llama_token llama_token_suffix ( const struct llama_model * model ) ; // Beginning of infill suffix
LLAMA_API llama_token llama_token_eot ( const struct llama_model * model ) ; // End of infill middle
2023-08-21 22:07:43 +02:00
//
// Tokenization
//
2023-10-17 17:11:01 +02:00
/// @details Convert the provided text into tokens.
/// @param tokens The tokens pointer must be large enough to hold the resulting tokens.
/// @return Returns the number of tokens on success, no more than n_max_tokens
/// @return Returns a negative number on failure - the number of tokens that would have been returned
/// @param special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext.
/// Does not insert a leading space.
2024-01-02 15:15:16 +01:00
LLAMA_API int32_t llama_tokenize (
2023-07-14 20:55:24 +02:00
const struct llama_model * model ,
const char * text ,
2024-01-02 15:15:16 +01:00
int32_t text_len ,
2023-07-14 20:55:24 +02:00
llama_token * tokens ,
2024-01-02 15:15:16 +01:00
int32_t n_max_tokens ,
2023-10-17 17:11:01 +02:00
bool add_bos ,
bool special ) ;
2023-07-14 20:55:24 +02:00
2023-08-27 13:19:19 +02:00
// Token Id -> Piece.
// Uses the vocabulary in the provided context.
// Does not write null terminator to the buffer.
// User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens.
2024-01-02 15:15:16 +01:00
LLAMA_API int32_t llama_token_to_piece (
2023-07-14 20:55:24 +02:00
const struct llama_model * model ,
2023-08-21 22:07:43 +02:00
llama_token token ,
char * buf ,
2024-01-02 15:15:16 +01:00
int32_t length ) ;
llama : new sampling algorithms (#1126)
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
2023-04-29 07:34:41 +02:00
2023-08-21 22:07:43 +02:00
//
2023-07-24 05:58:10 +02:00
// Grammar
//
2023-08-21 22:07:43 +02:00
2023-07-24 05:58:10 +02:00
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 ) ;
2023-09-05 07:46:17 +02:00
LLAMA_API struct llama_grammar * llama_grammar_copy ( const struct llama_grammar * grammar ) ;
2023-08-21 22:07:43 +02:00
//
llama : new sampling algorithms (#1126)
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
2023-04-29 07:34:41 +02:00
// Sampling functions
2023-08-21 22:07:43 +02:00
//
llama : new sampling algorithms (#1126)
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
2023-04-29 07:34:41 +02:00
2023-09-28 18:04:36 +02:00
// Sets the current rng seed.
LLAMA_API void llama_set_rng_seed ( struct llama_context * ctx , uint32_t seed ) ;
llama : new sampling algorithms (#1126)
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
2023-04-29 07:34:41 +02:00
/// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
/// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
2023-10-20 20:07:23 +02:00
LLAMA_API void llama_sample_repetition_penalties (
2023-09-28 18:04:36 +02:00
struct llama_context * ctx ,
llama_token_data_array * candidates ,
const llama_token * last_tokens ,
2023-10-20 20:07:23 +02:00
size_t penalty_last_n ,
float penalty_repeat ,
float penalty_freq ,
float penalty_present ) ;
llama : new sampling algorithms (#1126)
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
2023-04-29 07:34:41 +02:00
2023-07-11 18:18:43 +02:00
/// @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
2024-01-15 14:06:52 +01:00
/// @param logits Logits extracted from the original generation context.
/// @param logits_guidance Logits extracted from 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.
/// @param scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
LLAMA_API void llama_sample_apply_guidance (
struct llama_context * ctx ,
float * logits ,
float * logits_guidance ,
float scale ) ;
LLAMA_API DEPRECATED ( void llama_sample_classifier_free_guidance (
2023-07-11 18:18:43 +02:00
struct llama_context * ctx ,
llama_token_data_array * candidates ,
struct llama_context * guidance_ctx ,
2024-01-15 14:06:52 +01:00
float scale ) ,
" use llama_sample_apply_guidance() instead " ) ;
2023-07-11 18:18:43 +02:00
llama : new sampling algorithms (#1126)
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
2023-04-29 07:34:41 +02:00
/// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
2023-09-28 18:04:36 +02:00
LLAMA_API void llama_sample_softmax (
struct llama_context * ctx ,
llama_token_data_array * candidates ) ;
llama : new sampling algorithms (#1126)
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
2023-04-29 07:34:41 +02:00
/// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
2023-09-28 18:04:36 +02:00
LLAMA_API void llama_sample_top_k (
struct llama_context * ctx ,
llama_token_data_array * candidates ,
2024-01-02 15:15:16 +01:00
int32_t k ,
2023-09-28 18:04:36 +02:00
size_t min_keep ) ;
llama : new sampling algorithms (#1126)
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
2023-04-29 07:34:41 +02:00
/// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
2023-09-28 18:04:36 +02:00
LLAMA_API void llama_sample_top_p (
struct llama_context * ctx ,
llama_token_data_array * candidates ,
float p ,
size_t min_keep ) ;
llama : new sampling algorithms (#1126)
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
2023-04-29 07:34:41 +02:00
2023-10-31 20:44:49 +01:00
/// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
LLAMA_API void llama_sample_min_p (
struct llama_context * ctx ,
llama_token_data_array * candidates ,
float p ,
size_t min_keep ) ;
llama : new sampling algorithms (#1126)
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
2023-04-29 07:34:41 +02:00
/// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
2023-09-28 18:04:36 +02:00
LLAMA_API void llama_sample_tail_free (
struct llama_context * ctx ,
llama_token_data_array * candidates ,
float z ,
size_t min_keep ) ;
llama : new sampling algorithms (#1126)
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
2023-04-29 07:34:41 +02:00
/// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
2023-09-28 18:04:36 +02:00
LLAMA_API void llama_sample_typical (
struct llama_context * ctx ,
llama_token_data_array * candidates ,
float p ,
size_t min_keep ) ;
2024-01-25 21:06:22 +01:00
/// @details Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772.
LLAMA_API void llama_sample_entropy (
struct llama_context * ctx ,
llama_token_data_array * candidates_p ,
float min_temp ,
float max_temp ,
float exponent_val ) ;
2023-09-28 18:04:36 +02:00
LLAMA_API void llama_sample_temp (
struct llama_context * ctx ,
llama_token_data_array * candidates ,
float temp ) ;
LLAMA_API DEPRECATED ( void llama_sample_temperature (
struct llama_context * ctx ,
llama_token_data_array * candidates ,
float temp ) ,
" use llama_sample_temp instead " ) ;
llama : new sampling algorithms (#1126)
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
2023-04-29 07:34:41 +02:00
2023-07-24 05:58:10 +02:00
/// @details Apply constraints from grammar
2023-09-28 18:04:36 +02:00
LLAMA_API void llama_sample_grammar (
struct llama_context * ctx ,
llama_token_data_array * candidates ,
const struct llama_grammar * grammar ) ;
2023-07-24 05:58:10 +02:00
llama : new sampling algorithms (#1126)
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
2023-04-29 07:34:41 +02:00
/// @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.
2023-09-28 18:04:36 +02:00
LLAMA_API llama_token llama_sample_token_mirostat (
struct llama_context * ctx ,
llama_token_data_array * candidates ,
float tau ,
float eta ,
2024-01-02 15:15:16 +01:00
int32_t m ,
2023-09-28 18:04:36 +02:00
float * mu ) ;
llama : new sampling algorithms (#1126)
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
2023-04-29 07:34:41 +02:00
/// @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.
2023-09-28 18:04:36 +02:00
LLAMA_API llama_token llama_sample_token_mirostat_v2 (
struct llama_context * ctx ,
llama_token_data_array * candidates ,
float tau ,
float eta ,
float * mu ) ;
llama : new sampling algorithms (#1126)
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
2023-04-29 07:34:41 +02:00
/// @details Selects the token with the highest probability.
2023-10-28 13:23:11 +02:00
/// Does not compute the token probabilities. Use llama_sample_softmax() instead.
2023-09-28 18:04:36 +02:00
LLAMA_API llama_token llama_sample_token_greedy (
struct llama_context * ctx ,
llama_token_data_array * candidates ) ;
2023-03-22 06:32:36 +01:00
llama : new sampling algorithms (#1126)
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
2023-04-29 07:34:41 +02:00
/// @details Randomly selects a token from the candidates based on their probabilities.
2023-09-28 18:04:36 +02:00
LLAMA_API llama_token llama_sample_token (
struct llama_context * ctx ,
llama_token_data_array * candidates ) ;
2023-03-22 06:32:36 +01:00
2023-07-24 05:58:10 +02:00
/// @details Accepts the sampled token into the grammar
2023-09-28 18:04:36 +02:00
LLAMA_API void llama_grammar_accept_token (
struct llama_context * ctx ,
struct llama_grammar * grammar ,
llama_token token ) ;
2023-07-24 05:58:10 +02:00
2023-08-25 17:18:48 +02:00
//
// Beam search
//
struct llama_beam_view {
const llama_token * tokens ;
2023-09-28 18:04:36 +02:00
2023-08-25 17:18:48 +02:00
size_t n_tokens ;
2023-09-28 18:04:36 +02:00
float p ; // Cumulative beam probability (renormalized relative to all beams)
bool eob ; // Callback should set this to true when a beam is at end-of-beam.
2023-08-25 17:18:48 +02:00
} ;
// Passed to beam_search_callback function.
// Whenever 0 < common_prefix_length, this number of tokens should be copied from any of the beams
// (e.g. beams[0]) as they will be removed (shifted) from all beams in all subsequent callbacks.
// These pointers are valid only during the synchronous callback, so should not be saved.
struct llama_beams_state {
2023-08-25 18:17:15 +02:00
struct llama_beam_view * beam_views ;
2023-09-28 18:04:36 +02:00
2023-08-25 17:18:48 +02:00
size_t n_beams ; // Number of elements in beam_views[].
size_t common_prefix_length ; // Current max length of prefix tokens shared by all beams.
2023-09-28 18:04:36 +02:00
bool last_call ; // True iff this is the last callback invocation.
2023-08-25 17:18:48 +02:00
} ;
// Type of pointer to the beam_search_callback function.
// void* callback_data is any custom data passed to llama_beam_search, that is subsequently
// passed back to beam_search_callback. This avoids having to use global variables in the callback.
2023-08-28 10:19:59 +02:00
typedef void ( * llama_beam_search_callback_fn_t ) ( void * callback_data , struct llama_beams_state ) ;
2023-08-25 17:18:48 +02:00
/// @details Deterministically returns entire sentence constructed by a beam search.
/// @param ctx Pointer to the llama_context.
/// @param callback Invoked for each iteration of the beam_search loop, passing in beams_state.
/// @param callback_data A pointer that is simply passed back to callback.
/// @param n_beams Number of beams to use.
/// @param n_past Number of tokens already evaluated.
/// @param n_predict Maximum number of tokens to predict. EOS may occur earlier.
2023-09-28 18:04:36 +02:00
LLAMA_API void llama_beam_search (
struct llama_context * ctx ,
llama_beam_search_callback_fn_t callback ,
void * callback_data ,
size_t n_beams ,
2024-01-02 15:15:16 +01:00
int32_t n_past ,
int32_t n_predict ) ;
2023-08-25 17:18:48 +02:00
2023-03-22 06:32:36 +01:00
// Performance information
2023-07-05 22:51:13 +02:00
LLAMA_API struct llama_timings llama_get_timings ( struct llama_context * ctx ) ;
2023-09-28 18:04:36 +02:00
2023-03-22 06:32:36 +01:00
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 ) ;
2023-08-21 22:07:43 +02:00
// Set callback for all future logging events.
// If this is not called, or NULL is supplied, everything is output on stderr.
2023-09-27 17:48:33 +02:00
LLAMA_API void llama_log_set ( ggml_log_callback log_callback , void * user_data ) ;
2023-08-21 22:07:43 +02:00
2023-08-29 08:33:27 +02:00
LLAMA_API void llama_dump_timing_info_yaml ( FILE * stream , const struct llama_context * ctx ) ;
2023-08-28 17:59:39 +02:00
2023-03-22 06:32:36 +01:00
# ifdef __cplusplus
}
# endif
2023-04-13 17:04:45 +02:00
// Internal API to be implemented by llama.cpp and used by tests/benchmarks only
# ifdef LLAMA_API_INTERNAL
# include <vector>
# include <string>
2023-08-21 22:07:43 +02:00
2023-04-13 17:04:45 +02:00
struct ggml_tensor ;
2023-09-15 21:38:27 +02:00
const std : : vector < std : : pair < std : : string , struct ggml_tensor * > > & llama_internal_get_tensor_map (
struct llama_context * ctx
) ;
2023-04-13 17:04:45 +02:00
2023-08-21 22:07:43 +02:00
# endif // LLAMA_API_INTERNAL
2023-04-13 17:04:45 +02:00
Rewrite loading code to try to satisfy everyone:
- Support all three formats (ggml, ggmf, ggjt). (However, I didn't
include the hack needed to support GPT4All files without conversion.
Those can still be used after converting them with convert.py from my
other PR.)
- Support both mmap and read (mmap is used by default, but can be
disabled with `--no-mmap`, and is automatically disabled for pre-ggjt
files or on platforms where mmap is not supported).
- Support multi-file models like before, but automatically determine the
number of parts rather than requiring `--n_parts`.
- Improve validation and error checking.
- Stop using the per-file type field (f16) entirely in favor of just
relying on the per-tensor type/size fields. This has no immediate
benefit, but makes it easier to experiment with different formats, and
should make it easier to support the new GPTQ-for-LLaMa models in the
future (I have some work in progress on that front).
- Support VirtualLock on Windows (using the same `--mlock` option as on
Unix).
- Indicate loading progress when using mmap + mlock. (Which led me
to the interesting observation that on my Linux machine, with a
warm file cache, mlock actually takes some time, whereas mmap
without mlock starts almost instantly...)
- To help implement this, move mlock support from ggml to the
loading code.
- madvise/PrefetchVirtualMemory support (based on #740)
- Switch from ifstream to the `fopen` family of functions to avoid
unnecessary copying and, when mmap is enabled, allow reusing the same
file descriptor for both metadata reads and mmap (whereas the existing
implementation opens the file a second time to mmap).
- Quantization now produces a single-file output even with multi-file
inputs (not really a feature as much as 'it was easier this way').
Implementation notes:
I tried to factor the code into more discrete pieces than before.
Regarding code style: I tried to follow the code style, but I'm naughty
and used a few advanced C++ features repeatedly:
- Destructors to make it easier to ensure everything gets cleaned up.
- Exceptions. I don't even usually use exceptions when writing C++, and
I can remove them if desired... but here they make the loading code
much more succinct while still properly handling a variety of errors,
ranging from API calls failing to integer overflow and allocation
failure. The exceptions are converted to error codes at the
API boundary.)
Co-authored-by: Pavol Rusnak <pavol@rusnak.io> (for the bit I copied from #740)
2023-04-08 21:24:37 +02:00
# endif // LLAMA_H