Package chat.octet.model

Class LlamaService

java.lang.Object

chat.octet.model.LlamaService

public class LlamaService
extends Object

Llama.cpp API

C++ source: llamajava.h, llamajava.cpp

Since:

b1865

Author:

William

Constructor Summary

Constructors

Constructor	Description
LlamaService()

Method Summary

Modifier and Type	Method	Description
static int	batchDecode(int sequenceId, int[] tokens, int inputLength, int pastTokenSize)	Batch decoding.
static void	clearCache(int sequenceId)	Clear cache in K-V sequences.
static void	clearCache(int sequenceId, int posStart, int posEnd)	Clear cache in K-V sequences.
static void	createNewContextWithModel(LlamaContextParams params)	Create new context with model.
static int	getContextSize()	Get model context size.
static float[]	getEmbedding()	Get embedding
static LlamaContextParams	getLlamaContextDefaultParams()	Get llama context default params.
static LlamaModelParams	getLlamaModelDefaultParams()	Get llama model default params.
static LlamaModelQuantizeParams	getLlamaModelQuantizeDefaultParams()	Get llama model quantize default params.
static LlamaTokenType	getLlamaTokenType(int token)	Get token type define.
static float[]	getLogits(int index)	Get Logits based on index, and the default index must be 0.
static Metrics	getSamplingMetrics(boolean reset)	Get sampling metrics
static String	getSystemInfo()	Get system parameter information.
static int	getTokenBOS()	Get special BOS token.
static int	getTokenEOS()	Get special EOS token.
static int	getTokenType(int token)	Get token type code.
static int	getVocabSize()	Get model vocab size.
static void	initNative()	initial JNI context.
static boolean	isMlockSupported()	Check whether MLOCK is supported.
static boolean	isMmapSupported()	Check whether MMAP is supported.
static void	llamaBackendFree()	Call once at the end of the program.
static void	llamaBackendInit(boolean numa)	Initialize the llama + ggml backend If numa is true, use NUMA optimizations Call once at the start of the program.
static int	llamaModelQuantize(String sourceModelFilePath, String outputModelFilePath, LlamaModelQuantizeParams params)	Quantize the model.
static int	llamaModelQuantize(String sourceModelFilePath, String outputModelFilePath, ModelFileType modelFileType)	Quantize the model.
static boolean	loadLlamaGrammar(String grammarRules)	Load llama grammar by rules.
static void	loadLlamaModelFromFile(String modelPath, LlamaModelParams params)	Load Llama model from file.
static int	loadLoraModelFromFile(String loraPath, float loraScale, String baseModelPath, int threads)	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.
static void	release()	Close model and release all resources.
static int	sampling(float[] logits, int[] lastTokens, int lastTokensSize, float penalty, float alphaFrequency, float alphaPresence, boolean penalizeNL, int mirostatMode, float mirostatTAU, float mirostatETA, float temperature, int topK, float topP, float tsf, float typical, float minP, int sequenceId, int pastTokenSize)	Inference sampling the next token.
static int	tokenize(byte[] buf, int bufferLength, int[] tokens, int maxTokens, boolean addBos, boolean specialTokens)	Convert the provided text into tokens.
static int[]	tokenize(String text, boolean addBos, boolean specialTokens)	Convert the provided text into tokens.
static int	tokenToPiece(int token, byte[] buf, int bufferLength)	Convert the token id to text piece.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Details

LlamaService

public LlamaService()

Method Details

initNative

public static void initNative()

initial JNI context.

getLlamaModelDefaultParams

public static LlamaModelParams getLlamaModelDefaultParams()

Get llama model default params.

Returns:

LlamaModelParams

getLlamaContextDefaultParams

public static LlamaContextParams getLlamaContextDefaultParams()

Get llama context default params.

Returns:

LlamaContextParams

getLlamaModelQuantizeDefaultParams

public static LlamaModelQuantizeParams getLlamaModelQuantizeDefaultParams()

Get llama model quantize default params.

Returns:

LlamaModelQuantizeParams

llamaBackendInit

public static void llamaBackendInit(boolean numa)

Initialize the llama + ggml backend If numa is true, use NUMA optimizations Call once at the start of the program.

Parameters:

numa - Use NUMA optimizations.

llamaBackendFree

public static void llamaBackendFree()

Call once at the end of the program. NOTE: currently only used for MPI.

loadLlamaModelFromFile

public static void loadLlamaModelFromFile(String modelPath,
 LlamaModelParams params)
                                   throws ModelException

Load Llama model from file.

Parameters:

modelPath - Llama model file path.

params - Llama model params.

Throws:

ModelException

See Also:

• LlamaModelParams

createNewContextWithModel

public static void createNewContextWithModel(LlamaContextParams params)
                                      throws ModelException

Create new context with model.

Parameters:

params - Llama context params.

Throws:

ModelException

See Also:

• LlamaContextParams

release

public static void release()

Close model and release all resources.

isMmapSupported

public static boolean isMmapSupported()

Check whether MMAP is supported.

Returns:

boolean

isMlockSupported

public static boolean isMlockSupported()

Check whether MLOCK is supported.

Returns:

boolean

getVocabSize

public static int getVocabSize()

Get model vocab size.

Returns:

int

getContextSize

public static int getContextSize()

Get model context size.

Returns:

int

loadLoraModelFromFile

public static int loadLoraModelFromFile(String loraPath,
 float loraScale,
 String baseModelPath,
 int threads)
                                 throws ModelException

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.

Parameters:

loraPath - LoRA adapter file path.

loraScale - LoRA scale.

baseModelPath - Base model file path.

threads - Thread number.

Returns:

int, Returns 0 on success, else failed.

Throws:

ModelException

getLogits

public static float[] getLogits(int index)

Get Logits based on index, and the default index must be 0.

Parameters:

index - index

Returns:

float[], Returns one-dimensional float array.

getEmbedding

public static float[] getEmbedding()

Get embedding

Returns:

float[], Returns embedding float array.

getTokenType

public static int getTokenType(int token)

Get token type code.

Parameters:

token - Token id.

Returns:

int

getTokenBOS

public static int getTokenBOS()

Get special BOS token.

Returns:

int, Returns token id.

getTokenEOS

public static int getTokenEOS()

Get special EOS token.

Returns:

int, Returns token id.

tokenize

public static int tokenize(byte[] buf,
 int bufferLength,
 int[] tokens,
 int maxTokens,
 boolean addBos,
 boolean specialTokens)

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.

Parameters:

buf - Text byte buffer.

bufferLength - Text byte buffer length.

tokens - Empty token arrays, Used to receive the returned tokens.

maxTokens - Max token size, by default is context size.

addBos - Add special BOS token.

specialTokens - Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext. Does not insert a leading space.

Returns:

int, Returns a negative number on failure, else the number of tokens that would have been returned.

tokenToPiece

public static int tokenToPiece(int token,
 byte[] buf,
 int bufferLength)

Convert the token id to text piece.

Parameters:

token - Token id.

buf - Input byte buffer.

bufferLength - Input byte buffer length.

Returns:

int, Returns byte buffer length of the piece.

getSamplingMetrics

public static Metrics getSamplingMetrics(boolean reset)

Get sampling metrics

Parameters:

reset - Reset the counter when finished.

Returns:

Metrics

See Also:

• Metrics

getSystemInfo

public static String getSystemInfo()

Get system parameter information.

Returns:

String

sampling

public static int sampling(float[] logits,
 int[] lastTokens,
 int lastTokensSize,
 float penalty,
 float alphaFrequency,
 float alphaPresence,
 boolean penalizeNL,
 int mirostatMode,
 float mirostatTAU,
 float mirostatETA,
 float temperature,
 int topK,
 float topP,
 float tsf,
 float typical,
 float minP,
 int sequenceId,
 int pastTokenSize)
                    throws DecodeException

Inference sampling the next token.

Parameters:

logits - User-defined logits, Adjustments can be made via LogitsProcessor.

lastTokens - Last token array.

lastTokensSize - Last token array size.

penalty - Control the repetition of token sequences in the generated text.

alphaFrequency - Repeat alpha frequency penalty.

alphaPresence - Repeat alpha presence penalty.

penalizeNL - Disable penalization for newline tokens when applying the repeat penalty.

mirostatMode - Mirostat Sampling Use Mirostat sampling, controlling perplexity during text generation.

mirostatTAU - Mirostat Sampling Set the Mirostat target entropy.

mirostatETA - Mirostat Sampling Set the Mirostat learning rate.

temperature - Adjust the randomness of the generated text.

topK - TOP-K Sampling Limit the next token selection to the K most probable tokens.

topP - TOP-P Sampling Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P.

tsf - Tail Free Sampling (TFS) Enable tail free sampling with parameter z.

typical - Typical Sampling Enable typical sampling sampling with parameter p.

minP - MIN-P Sampling Sets a minimum base probability threshold for token selection.

sequenceId - Generation sequence id.

pastTokenSize - Past token size.

Returns:

int, Returns the sampled token id.

Throws:

DecodeException

See Also:

• GenerateParameter

loadLlamaGrammar

public static boolean loadLlamaGrammar(String grammarRules)

Load llama grammar by rules.

Parameters:

grammarRules - Grammar rules.

Returns:

boolean, Returns true on success, else failed.

batchDecode

public static int batchDecode(int sequenceId,
 int[] tokens,
 int inputLength,
 int pastTokenSize)

Batch decoding.

Parameters:

sequenceId - Specify a unique generation sequence id.

tokens - Arrays of tokens that need to be decoding.

inputLength - Input context length.

pastTokenSize - Past token size.

Returns:

int, Returns 0 on success, else failed.

clearCache

public static void clearCache(int sequenceId,
 int posStart,
 int posEnd)

Clear cache in K-V sequences.

Parameters:

sequenceId - Generation sequence id.

posStart - Start position.

posEnd - End position.

clearCache

public static void clearCache(int sequenceId)

Clear cache in K-V sequences.

Parameters:

sequenceId - Generation sequence id.

llamaModelQuantize

public static int llamaModelQuantize(String sourceModelFilePath,
 String outputModelFilePath,
 LlamaModelQuantizeParams params)

Quantize the model.

Parameters:

sourceModelFilePath - Source model file path.

outputModelFilePath - Output model file path.

params - Quantize parameters.

Returns:

int, Returns 0 on success, else failed.

llamaModelQuantize

public static int llamaModelQuantize(String sourceModelFilePath,
 String outputModelFilePath,
 ModelFileType modelFileType)

Quantize the model.

Parameters:

sourceModelFilePath - Source model file path.

outputModelFilePath - Output model file path.

modelFileType - Model file type.

Returns:

int, Returns 0 on success, else failed.

See Also:

• ModelFileType

tokenize

public static int[] tokenize(String text,
 boolean addBos,
 boolean specialTokens)

Convert the provided text into tokens.

Parameters:

text - Input text.

addBos - Add special BOS token.

specialTokens - Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext. Does not insert a leading space.

Returns:

Returns a negative number on failure, else the number of tokens that would have been returned.

getLlamaTokenType

public static LlamaTokenType getLlamaTokenType(int token)

Get token type define.

Parameters:

token - Token id.

Returns:

LlamaTokenType

See Also:

• LlamaTokenType