2023-03-22 06:32:36 +01:00
# ifndef LLAMA_H
# define LLAMA_H
# include <stddef.h>
# include <stdint.h>
# 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
2023-05-11 23:23:08 +02:00
# define LLAMA_FILE_VERSION 2
2023-05-01 13:54:59 +02:00
# define LLAMA_FILE_MAGIC 'ggjt'
# define LLAMA_FILE_MAGIC_UNVERSIONED 'ggml'
# define LLAMA_SESSION_MAGIC 'ggsn'
2023-05-03 04:26:13 +02:00
# define LLAMA_SESSION_VERSION 1
2023-03-22 06:32:36 +01:00
# ifdef __cplusplus
extern " C " {
# endif
//
// C interface
//
// TODO: show sample usage
//
struct llama_context ;
typedef int llama_token ;
typedef struct llama_token_data {
llama_token id ; // token id
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
float logit ; // log-odds of the token
2023-03-22 06:32:36 +01:00
float p ; // probability of the token
} 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 ;
2023-03-28 18:48:20 +02:00
typedef void ( * llama_progress_callback ) ( float progress , void * ctx ) ;
2023-03-25 06:26:28 +01:00
2023-03-22 06:32:36 +01:00
struct llama_context_params {
int n_ctx ; // text context
int n_parts ; // -1 for default
2023-05-02 18:23:44 +02:00
int seed ; // RNG seed, -1 for random
2023-03-22 06:32:36 +01:00
bool f16_kv ; // use fp16 for KV cache
bool logits_all ; // the llama_eval() call computes all logits, not just the last one
bool vocab_only ; // only load the vocabulary, no weights
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
2023-03-24 16:05:13 +01:00
bool embedding ; // embedding mode only
2023-03-25 06:26:28 +01:00
// called with a progress value between 0 and 1, pass NULL to disable
llama_progress_callback progress_callback ;
// context pointer passed to the progress callback
void * progress_callback_user_data ;
2023-03-22 06:32:36 +01:00
} ;
2023-04-11 17:03:51 +02:00
// model file types
enum llama_ftype {
LLAMA_FTYPE_ALL_F32 = 0 ,
LLAMA_FTYPE_MOSTLY_F16 = 1 , // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_0 = 2 , // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_1 = 3 , // except 1d tensors
2023-04-12 17:06:16 +02:00
LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4 , // tok_embeddings.weight and output.weight are F16
2023-05-11 23:23:08 +02:00
// LLAMA_FTYPE_MOSTLY_Q4_2 = 5, // support has been removed
2023-04-29 01:10:43 +02:00
// LLAMA_FTYPE_MOSTLY_Q4_3 (6) support has been removed
2023-04-25 22:40:51 +02:00
LLAMA_FTYPE_MOSTLY_Q8_0 = 7 , // except 1d tensors
2023-04-26 22:14:13 +02:00
LLAMA_FTYPE_MOSTLY_Q5_0 = 8 , // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_1 = 9 , // except 1d tensors
2023-04-11 17:03:51 +02:00
} ;
2023-03-22 06:32:36 +01:00
LLAMA_API struct llama_context_params llama_context_default_params ( ) ;
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
LLAMA_API bool llama_mmap_supported ( ) ;
LLAMA_API bool llama_mlock_supported ( ) ;
2023-03-22 06:32:36 +01:00
// Various functions for loading a ggml llama model.
// Allocate (almost) all memory needed for the model.
// Return NULL on failure
LLAMA_API struct llama_context * llama_init_from_file (
const char * path_model ,
struct llama_context_params params ) ;
// Frees all allocated memory
LLAMA_API void llama_free ( struct llama_context * ctx ) ;
// TODO: not great API - very likely to change
// Returns 0 on success
2023-04-20 19:42:27 +02:00
// nthread - how many threads to use. If <=0, will use std::thread::hardware_concurrency(), else the number given
2023-03-22 06:32:36 +01:00
LLAMA_API int llama_model_quantize (
const char * fname_inp ,
const char * fname_out ,
2023-04-20 19:42:27 +02:00
enum llama_ftype ftype ,
int nthread ) ;
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
LLAMA_API int llama_apply_lora_from_file (
struct llama_context * ctx ,
const char * path_lora ,
const char * path_base_model ,
int n_threads ) ;
2023-04-02 12:23:04 +02:00
// Returns the number of tokens in the KV cache
2023-05-01 09:24:20 +02:00
LLAMA_API int llama_get_kv_cache_token_count ( const struct llama_context * ctx ) ;
2023-04-02 12:23:04 +02:00
2023-04-26 22:08:43 +02:00
// Sets the current rng seed.
LLAMA_API void llama_set_rng_seed ( struct llama_context * ctx , int seed ) ;
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
2023-05-13 08:08:52 +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
LLAMA_API size_t llama_set_state_data ( struct llama_context * ctx , const uint8_t * src ) ;
2023-04-28 17:59:37 +02:00
// Save/load session file
2023-05-01 13:54:59 +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
2023-03-22 06:32:36 +01:00
// Run the llama inference to obtain the logits and probabilities for the next token.
// tokens + n_tokens is the provided batch of new tokens to process
// n_past is the number of tokens to use from previous eval calls
// Returns 0 on success
LLAMA_API int llama_eval (
struct llama_context * ctx ,
const llama_token * tokens ,
int n_tokens ,
int n_past ,
int n_threads ) ;
// Convert the provided text into tokens.
// The tokens pointer must be large enough to hold the resulting tokens.
// Returns the number of tokens on success, no more than n_max_tokens
// Returns a negative number on failure - the number of tokens that would have been returned
// TODO: not sure if correct
LLAMA_API int llama_tokenize (
struct llama_context * ctx ,
const char * text ,
llama_token * tokens ,
int n_max_tokens ,
bool add_bos ) ;
2023-05-01 09:24:20 +02:00
LLAMA_API int llama_n_vocab ( const struct llama_context * ctx ) ;
LLAMA_API int llama_n_ctx ( const struct llama_context * ctx ) ;
LLAMA_API int llama_n_embd ( const struct llama_context * ctx ) ;
2023-03-22 06:32:36 +01:00
// Token logits obtained from the last call to llama_eval()
// The logits for the last token are stored in the last row
// Can be mutated in order to change the probabilities of the next token
// Rows: n_tokens
// Cols: n_vocab
LLAMA_API float * llama_get_logits ( struct llama_context * ctx ) ;
2023-03-24 16:05:13 +01:00
// Get the embeddings for the input
// shape: [n_embd] (1-dimensional)
LLAMA_API float * llama_get_embeddings ( struct llama_context * ctx ) ;
2023-03-22 06:32:36 +01:00
// Token Id -> String. Uses the vocabulary in the provided context
2023-05-01 09:24:20 +02:00
LLAMA_API const char * llama_token_to_str ( const struct llama_context * ctx , llama_token token ) ;
2023-03-22 06:32:36 +01:00
// Special tokens
LLAMA_API llama_token llama_token_bos ( ) ;
LLAMA_API llama_token llama_token_eos ( ) ;
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
LLAMA_API llama_token llama_token_nl ( ) ;
// Sampling functions
/// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
2023-05-02 22:09:08 +02:00
LLAMA_API void llama_sample_repetition_penalty ( struct llama_context * ctx , llama_token_data_array * candidates , const llama_token * last_tokens , size_t last_tokens_size , float penalty ) ;
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 Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
2023-05-02 22:09:08 +02:00
LLAMA_API void llama_sample_frequency_and_presence_penalties ( struct llama_context * ctx , llama_token_data_array * candidates , const llama_token * last_tokens , size_t last_tokens_size , float alpha_frequency , float alpha_presence ) ;
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.
LLAMA_API void llama_sample_softmax ( struct llama_context * ctx , llama_token_data_array * candidates ) ;
/// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
2023-05-06 23:01:47 +02:00
LLAMA_API void llama_sample_top_k ( struct llama_context * ctx , llama_token_data_array * candidates , int k , 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
2023-05-06 23:01:47 +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
/// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
2023-05-06 23:01:47 +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.
2023-05-06 23:01:47 +02:00
LLAMA_API void llama_sample_typical ( 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
LLAMA_API void llama_sample_temperature ( struct llama_context * ctx , llama_token_data_array * candidates , float temp ) ;
/// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
/// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
/// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
/// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
/// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
/// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
LLAMA_API llama_token llama_sample_token_mirostat ( struct llama_context * ctx , llama_token_data_array * candidates , float tau , float eta , int m , float * mu ) ;
/// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
/// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
/// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
/// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
/// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
LLAMA_API llama_token llama_sample_token_mirostat_v2 ( struct llama_context * ctx , llama_token_data_array * candidates , float tau , float eta , float * mu ) ;
/// @details Selects the token with the highest probability.
LLAMA_API llama_token llama_sample_token_greedy ( struct llama_context * ctx , llama_token_data_array * candidates ) ;
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.
LLAMA_API llama_token llama_sample_token ( struct llama_context * ctx , llama_token_data_array * candidates ) ;
2023-03-22 06:32:36 +01:00
// Performance information
LLAMA_API void llama_print_timings ( struct llama_context * ctx ) ;
LLAMA_API void llama_reset_timings ( struct llama_context * ctx ) ;
// Print system information
LLAMA_API const char * llama_print_system_info ( void ) ;
# ifdef __cplusplus
}
# endif
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>
struct ggml_tensor ;
std : : vector < std : : pair < std : : string , struct ggml_tensor * > > & llama_internal_get_tensor_map ( struct llama_context * ctx ) ;
# endif
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