ml

package
v0.0.0-...-0e1d3c2 Latest Latest
Warning

This package is not in the latest version of its module.

 Go to latest Published: May 16, 2023 License: MIT Imports: 9 Imported by: 0

Documentation

Index

Constants

View Source 
const (
	DEBUG = false

	MAX_DIMS   = 4
	MAX_NODES  = 4096
	MAX_PARAMS = 16
	MAX_OPT    = 4

	QK = 32 // quantization

	TOKEN_BOS = 1
	TOKEN_EOS = 2
)
View Source 
const (
	OP_NONE optype = iota
	OP_DUP
	OP_ADD
	OP_SUB
	OP_MUL
	OP_DIV
	OP_SQR
	OP_SQRT
	OP_SUM
	OP_MEAN
	OP_REPEAT
	OP_ABS
	OP_SGN
	OP_NEG
	OP_STEP
	OP_RELU
	OP_GELU
	OP_SILU
	OP_NORM
	OP_RMS_NORM

	OP_MUL_MAT

	OP_SCALE
	OP_CPY
	OP_RESHAPE
	OP_VIEW
	OP_PERMUTE
	OP_TRANSPOSE
	OP_GET_ROWS
	OP_DIAG_MASK_INF
	OP_SOFT_MAX
	OP_ROPE
	OP_CONV_1D_1S
	OP_CONV_1D_2S

	OP_FLASH_ATTN
	OP_FLASH_FF

	OP_COUNT
)
View Source 
const MaxMem = 0 // 28_000_000_000
View Source 
const MaxPool = 0 // 2_000_000_000

TODO: Precompute max needed RAM size

View Source 
const NewLineToken = 13 // ml.Tokenize(Ctx.Vocab, "\n", false)[0]

Variables

View Source 
var BLCK_SIZE [TYPE_COUNT]uint32 = [TYPE_COUNT]uint32{1, 1, QK, QK, 1, 1, 1, 0}
View Source 
var TYPE_SIZE [TYPE_COUNT]uint32 = [TYPE_COUNT]uint32{4, 2, 4 + QK/2, 4*2 + QK/2, 1, 2, 4, 0}
View Source 
var TableExpFP16 [1 << 16]float16.Float16

precomputed exp table for f16 (128 KB) static ggml_fp16_t table_exp_f16[1 << 16];

Functions

func AreSameShape 

func AreSameShape(a, b *Tensor) bool

func BuildForwardExpand 

func BuildForwardExpand(graph *Graph, tensor *Tensor)

ggml_build_forward_expand

func BuildForwardImpl 

func BuildForwardImpl(graph *Graph, tensor *Tensor, expand bool)

ggml_build_forward_impl

func CanMulMat 

func CanMulMat(t0, t1 *Tensor) bool

ggml_can_mul_mat

func CheckGraph 

func CheckGraph()

TODO: Implement all the tensor asserts BEFORE the real computing

func ComputeBackward 

func ComputeBackward(ctx *Context, tensor *Tensor, inplace bool)

func ComputeForward 

func ComputeForward(ctx *Context, graph *Graph, params *ComputeParams, tensor *Tensor)

func ComputeForwardAddFP32 

func ComputeForwardAddFP32(params *ComputeParams, src0, src1, dst *Tensor)

ggml_compute_forward_add

func ComputeForwardCopy 

func ComputeForwardCopy(params *ComputeParams, src0, dst *Tensor)

func ComputeForwardDiagMaskInfFP32 

func ComputeForwardDiagMaskInfFP32(params *ComputeParams, src0, src1, dst *Tensor)

ggml_compute_forward_diag_mask_inf

func ComputeForwardDupFP32 

func ComputeForwardDupFP32(params *ComputeParams, src0, dst *Tensor)

ggml_compute_forward_dup_f32

func ComputeForwardGetRows 

func ComputeForwardGetRows(params *ComputeParams, src0, src1, dst *Tensor)

ggml_compute_forward_get_rows_f32

func ComputeForwardMulFP32 

func ComputeForwardMulFP32(params *ComputeParams, src0, src1, dst *Tensor)

ggml_compute_forward_mul

func ComputeForwardMulMatFP32 

func ComputeForwardMulMatFP32(params *ComputeParams, src0, src1, dst *Tensor)

ggml_compute_forward_mul_mat_f32

func ComputeForwardPermute 

func ComputeForwardPermute(params *ComputeParams, src0 *Tensor)

ggml_compute_forward_permute

func ComputeForwardRMSNormFP32 

func ComputeForwardRMSNormFP32(params *ComputeParams, src0, dst *Tensor)

ggml_compute_forward_rms_norm_f32

func ComputeForwardRepeatFP32 

func ComputeForwardRepeatFP32(params *ComputeParams, src0, dst *Tensor)

ggml_compute_forward_repeat

func ComputeForwardReshape 

func ComputeForwardReshape(params *ComputeParams, src0, dst *Tensor)

ggml_compute_forward_reshape

func ComputeForwardRopeFP32 

func ComputeForwardRopeFP32(params *ComputeParams, src0, src1, dst *Tensor)

ggml_compute_forward_rope

func ComputeForwardScaleFP32 

func ComputeForwardScaleFP32(params *ComputeParams, src0, src1, dst *Tensor)

ggml_compute_forward_scale_f32

func ComputeForwardSiluFP32 

func ComputeForwardSiluFP32(params *ComputeParams, src0, dst *Tensor)

ggml_compute_forward_silu

func ComputeForwardSoftMaxFP32 

func ComputeForwardSoftMaxFP32(params *ComputeParams, src0, dst *Tensor)

ggml_compute_forward_soft_max

func ComputeForwardView 

func ComputeForwardView(params *ComputeParams, src0 *Tensor)

ggml_compute_forward_view

func Do 

func Do(params *ComputeParams, id int)

Do is an experimental alternative for always waiting Job threads

func GraphCompute 

func GraphCompute(ctx *Context, graph *Graph)

func Init 

func Init(params InitParams)

TODO Do we need this?

func IsMatrix 

func IsMatrix(tensor *Tensor) bool

func IsScalar 

func IsScalar(tensor *Tensor) bool

func IsVector 

func IsVector(tensor *Tensor) bool

func Job 

func Job(listen <-chan *ComputeParams, id int)

Job is goroutine existing while the computation loop is active The main purpose of the Job is to perform some part of time consuming matrix multiplications TODO: Investigate https://pkg.go.dev/runtime#LockOSThread

func SiluFP32 

func SiluFP32(x float32) float32

Sigmoid Linear Unit (SiLU) function

func Token2Str 

func Token2Str(vocab *Vocab, token uint32) string

func Tokenize 

func Tokenize(vocab *Vocab, text string, bos bool) []uint32

void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {

func TryAddBigram 

func TryAddBigram(vocab *Vocab, symbols []Symbol, workQueue *[]Bigram, left, right int)

func TypeSizeFloat 

func TypeSizeFloat(dt DType) float32

func VecAccFP32 

func VecAccFP32(n uint32, y, x []float32)

ggml_vec_acc_f32

func VecAddFP32 

func VecAddFP32(n uint32, z, x, y []float32)

inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] + y[i]; }

func VecCopyFP32 

func VecCopyFP32(n uint32, y, x []float32)

func VecDotFP32 

func VecDotFP32(n uint32, x, y []float32) float32

ggml_vec_dot_f32

func VecMadFP32 

func VecMadFP32(n uint32, y, x []float32, v float32)

ggml_vec_mad_f32

func VecMaxFP32 

func VecMaxFP32(n uint32, x []float32) float32

func VecMulFP32 

func VecMulFP32(n uint32, z, x, y []float32)

func VecScaleFP32 

func VecScaleFP32(n uint32, y []float32, v float32)

ggml_vec_scale_f32

func VecSiluFP32 

func VecSiluFP32(n uint32, y, x []float32)

inline static void ggml_vec_silu_f32(const int n, float * y, const float * x) {

func VisitParents 

func VisitParents(graph *Graph, node *Tensor)

ggml_visit_parents

Types

type Allocator 

type Allocator struct {
	sync.Mutex

	PoolSize int
	MemSize  int

	Pool []byte
	Mem  []byte
}

Allocator is an experimental memory pool for FP32 slices TODO: Investigate https://github.com/valyala/bytebufferpool

func NewAllocator 

func NewAllocator() *Allocator

func (*Allocator) Get 

func (a *Allocator) Get(size uint32) *[]float32

Get new or reuse memory buffer of size bytes

func (*Allocator) GetFixed 

func (a *Allocator) GetFixed(size uint32) *[]float32

Get fixed memory buffer of size bytes

func (*Allocator) Reset 

func (a *Allocator) Reset()

type Bigram 

type Bigram struct {

	// NB! Allow -1
	Left  int
	Right int

	Score float32
	Size  uint32
}

struct llama_sp_bigram {

func PopMax 

func PopMax(queue *[]Bigram) Bigram

type ComputeParams 

type ComputeParams struct {
	Type TaskType

	UseAVX  bool
	UseNEON bool
	// contains filtered or unexported fields
}

type Context 

type Context struct {
	MaxThreads int
	UseAVX     bool
	UseNEON    bool
	//Graph      *Graph
	Compute   chan *ComputeParams
	Allocator *Allocator
}

func NewContext 

func NewContext(maxThreads int, useAVX, useNEON bool) *Context

func (*Context) ReleaseContext 

func (ctx *Context) ReleaseContext()

ReleaseContext frees all context resources - channel will be closed and goroutines stopped

type DType 

type DType uint8
const (
	TYPE_F32   DType = 0
	TYPE_F16   DType = 1
	TYPE_Q4_0  DType = 2
	TYPE_Q4_1  DType = 3
	TYPE_I8    DType = 4
	TYPE_I16   DType = 5
	TYPE_I32   DType = 6
	TYPE_COUNT DType = 8
)

Data types are the same as in llama.cpp so full compatibility there

type Graph 

type Graph struct {
	NodesCount uint32
	LeafsCount uint32

	Jobs chan *ComputeParams

	Nodes [MAX_NODES]*Tensor
	Grads [MAX_NODES]*Tensor
	Leafs [MAX_NODES]*Tensor
}

computation graph

func BuildBackward 

func BuildBackward(ctx *Context, gf *Graph, keep bool) Graph

func BuildForward 

func BuildForward(tensor *Tensor) *Graph

type InitParams 

type InitParams struct {
}

type Symbol 

type Symbol struct {

	// NB! Allow -1
	Prev int
	Next int

	Text string
	N    uint32
}

struct llama_sp_symbol {

type TaskType 

type TaskType uint8
const (
	TASK_INIT     TaskType = 0
	TASK_COMPUTE  TaskType = 1
	TASK_FINALIZE TaskType = 2
)

type Tensor 

type Tensor struct {
	Type DType

	Reusable bool // this tensor Data buffer might be reused with pooling

	Dims uint32

	NE [MAX_DIMS]uint32 // number of elements
	NB [MAX_DIMS]uint32 // stride in bytes

	TasksCount int

	Data []float32
	// contains filtered or unexported fields
}

func Add 

func Add(ctx *Context, a, b *Tensor) *Tensor

func AddImpl 

func AddImpl(ctx *Context, a, b *Tensor, inplace bool) *Tensor

ggml_add

func AddInplace 

func AddInplace(ctx *Context, a, b *Tensor) *Tensor

func Copy 

func Copy(ctx *Context, a, b *Tensor) *Tensor

func CopyImpl 

func CopyImpl(ctx *Context, a, b *Tensor, inplace bool) *Tensor

ggml_cpy

func CopyInplace 

func CopyInplace(ctx *Context, a, b *Tensor) *Tensor

func DiagMaskInf 

func DiagMaskInf(ctx *Context, a *Tensor, past uint32) *Tensor

ggml_diag_mask_inf

func Div 

func Div(ctx *Context, a, b *Tensor) *Tensor

func DivImpl 

func DivImpl(ctx *Context, a, b *Tensor, inplace bool) *Tensor

ggml_div

func DivInplace 

func DivInplace(ctx *Context, a, b *Tensor, inplace bool) *Tensor

func DupTensor 

func DupTensor(ctx *Context, src *Tensor) *Tensor

ggml_dup_tensor

func GetRows 

func GetRows(ctx *Context, a, b *Tensor) *Tensor

ggml_get_rows

func Mul 

func Mul(ctx *Context, a, b *Tensor) *Tensor

struct ggml_tensor * Mul(

func MulImpl 

func MulImpl(ctx *Context, a, b *Tensor, inplace bool) *Tensor

struct ggml_tensor * Mul_impl(

func MulInplace 

func MulInplace(ctx *Context, a, b *Tensor) *Tensor

struct ggml_tensor * Mul_inplace(

func MulMat 

func MulMat(ctx *Context, a, b *Tensor) *Tensor

ggml_mul_mat

func NewFP32 

func NewFP32(ctx *Context, value float32) *Tensor

ggml_new_f32

func NewTensor 

func NewTensor(ctx *Context, dt DType, dims uint32, ne0, ne1, ne2, ne3 uint32, data []float32) *Tensor

ggml_new_tensor_impl

func NewTensor1D 

func NewTensor1D(ctx *Context, dt DType, ne0 uint32) *Tensor

ggml_new_tensor_1d

func NewTensor2D 

func NewTensor2D(ctx *Context, dt DType, ne0, ne1 uint32) *Tensor

ggml_new_tensor_2d

func NewTensor3D 

func NewTensor3D(ctx *Context, dt DType, ne0, ne1, ne2 uint32) *Tensor

func NewTensor4D 

func NewTensor4D(ctx *Context, dt DType, ne0, ne1, ne2, ne3 uint32) *Tensor

func Permute 

func Permute(ctx *Context, a *Tensor, axis0, axis1, axis2, axis3 uint32) *Tensor

ggml_permute

func RMSNorm 

func RMSNorm(ctx *Context, a *Tensor) *Tensor

func RMSNormImpl 

func RMSNormImpl(ctx *Context, a *Tensor, inplace bool) *Tensor

ggml_rms_norm_impl

func RMSNormInplace 

func RMSNormInplace(ctx *Context, a *Tensor) *Tensor

func Repeat 

func Repeat(ctx *Context, a, b *Tensor) *Tensor

struct ggml_tensor * Repeat(

func Reshape3D 

func Reshape3D(ctx *Context, a *Tensor, ne0, ne1, ne2 uint32) *Tensor

func Rope 

func Rope(ctx *Context, a *Tensor, past, dims, mode uint32) *Tensor

ggml_rope

func Scale 

func Scale(ctx *Context, a, b *Tensor) *Tensor

func ScaleImpl 

func ScaleImpl(ctx *Context, a, b *Tensor, inplace bool) *Tensor

ggml_scale

func ScaleInplace 

func ScaleInplace(ctx *Context, a, b *Tensor) *Tensor

func SetFP32 

func SetFP32(tensor *Tensor, value float32) *Tensor

ggml_set_f32

func Sgn 

func Sgn(ctx *Context, a *Tensor) *Tensor

func SgnImpl 

func SgnImpl(ctx *Context, a *Tensor, inplace bool) *Tensor

ggml_sgn

func SgnInplace 

func SgnInplace(ctx *Context, a *Tensor) *Tensor

func Silu 

func Silu(ctx *Context, a *Tensor) *Tensor

func SiluImpl 

func SiluImpl(ctx *Context, a *Tensor, inplace bool) *Tensor

func SiluInplace 

func SiluInplace(ctx *Context, a *Tensor) *Tensor

func SoftMax 

func SoftMax(ctx *Context, a *Tensor) *Tensor

ggml_soft_max

func Step 

func Step(ctx *Context, a *Tensor) *Tensor

func StepImpl 

func StepImpl(ctx *Context, a *Tensor, inplace bool) *Tensor

ggml_step

func StepInplace 

func StepInplace(ctx *Context, a *Tensor) *Tensor

func Sub 

func Sub(ctx *Context, a, b *Tensor) *Tensor

func SubImpl 

func SubImpl(ctx *Context, a, b *Tensor, inplace bool) *Tensor

ggml_sub

func SubInplace 

func SubInplace(ctx *Context, a, b *Tensor) *Tensor

func Sum 

func Sum(ctx *Context, a *Tensor) *Tensor

ggml_sum

func Transpose 

func Transpose(ctx *Context, a *Tensor) *Tensor

func View1D 

func View1D(ctx *Context, a *Tensor, ne0 uint32, offset uint32) *Tensor

ggml_view_1d NB! Originally offset in bytes, but here in floats (4-bytes)

func ViewTensor 

func ViewTensor(ctx *Context, src *Tensor) *Tensor

ggml_view_tensor

func (*Tensor) IsContiguous 

func (tensor *Tensor) IsContiguous() bool

ggml_is_contiguous

func (*Tensor) Nbytes 

func (t *Tensor) Nbytes() uint32

ggml_nbytes

func (*Tensor) Nelements 

func (t *Tensor) Nelements() uint32

func (*Tensor) Nrows 

func (t *Tensor) Nrows() uint32

type TokenScore 

type TokenScore struct {
	Token string
	Score float32
}

type Vocab 

type Vocab struct {
	Size     uint32
	Token2ID map[string]uint32
	ID2Token []TokenScore
}

func NewVocab 

func NewVocab(size uint32) *Vocab

Source Files

View all Source files

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL
go.dev uses cookies from Google to deliver and enhance the quality of its services and to analyze traffic. Learn more.