Package io.github.kirstenali.deepj.tensor.metal

Class MetalBackend

java.lang.Object

io.github.kirstenali.deepj.tensor.metal.MetalBackend

All Implemented Interfaces:

TensorBackend

public final class MetalBackend
extends Object
implements TensorBackend

Constructor Summary

Constructors

Constructor	Description
MetalBackend()

Method Summary

Modifier and Type	Method	Description
void	adamWUpdate(Tensor w, Tensor g, Tensor mt, Tensor vt, double lr, double beta1, double beta2, double eps, double weightDecay, double bc1, double bc2)	In-place AdamW update.
Tensor	add(Tensor a, Tensor b)
Tensor	addBroadcastCols(Tensor a, Tensor v)
Tensor	addBroadcastRows(Tensor a, Tensor v)
Tensor	addRowVector(Tensor a, Tensor v)
Tensor	addScalar(Tensor a, double scalar)
Tensor	causalMask(int size)
Tensor	clamp(Tensor a, double min, double max)
Tensor	crossEntropyGradient(Tensor logits, int[] targets)
double	crossEntropyLoss(Tensor logits, int[] targets)
Tensor	divide(Tensor a, Tensor b)
Tensor	divideBroadcastCols(Tensor a, Tensor v)
Tensor	divideScalar(Tensor a, double scalar)
Tensor	exp(Tensor a)
double[]	flattenTensor(Tensor t)
Tensor	gelu(Tensor a)
Tensor	geluBackward(Tensor input, Tensor gradOutput)
double	get(Tensor t, int r, int c)
Tensor	getRow(Tensor t, int row)
Tensor	layerNormBackward(Tensor dXHat, Tensor xHat, Tensor std, int dim)	LayerNorm backward through normalization (given dXHat, xHat, std).
Tensor	log(Tensor a)
void	materializeTensor(Tensor t)	Materialize a tensor: flush any pending GPU computation and download the result to the tensor's CPU data[][].
Tensor	matmul(Tensor a, Tensor b)
Tensor	maxAlongRows(Tensor a)
Tensor	meanAlongRows(Tensor a)
Tensor	multiply(Tensor a, Tensor b)
Tensor	multiplyBroadcastCols(Tensor a, Tensor v)
Tensor	multiplyBroadcastRows(Tensor a, Tensor v)
Tensor	multiplyScalar(Tensor a, double scalar)
Tensor	neg(Tensor a)
Tensor	ones(int rows, int cols)
Tensor	pow(Tensor a, double exponent)
void	print(Tensor t, String label)
Tensor	random(int rows, int cols, Random rand)
void	releaseResources()	Release backend-owned resources (GPU buffers, native handles, etc.).
Tensor	relu(Tensor a)
Tensor	reluBackward(Tensor input, Tensor gradOutput)
Tensor	sampleRows(Tensor t, int n, Random rnd)	Sample random rows from t.
void	scatterAddRows(Tensor target, int[] indices, Tensor grad)	Scatter-add: target.data[indices[i]] += grad.data[i] (for embedding backward).
void	set(Tensor t, int r, int c, double value)
void	setRow(Tensor t, int row, Tensor source, int srcRow)
Tensor	sigmoid(Tensor a)
Tensor	sliceRows(Tensor t, int[] rowIndices, int cols)	Gather rows by index (for embedding lookup).
Tensor	softmaxBackward(Tensor gradOutput, Tensor softmaxOut)
Tensor	softmaxRows(Tensor logits)
Tensor	sqrt(Tensor a)
Tensor	subtract(Tensor a, Tensor b)
Tensor	subtractBroadcastCols(Tensor a, Tensor v)
double	sum(Tensor a)
double	sumAbs(Tensor a)
Tensor	sumAlongCols(Tensor a)
Tensor	sumAlongRows(Tensor a)
Tensor	sumRows(Tensor a)
Tensor	tanh(Tensor a)
Tensor	transpose(Tensor a)
Tensor	unflattenToTensor(double[] flat, int rows, int cols)
Tensor	varianceAlongRows(Tensor a)
Tensor	zeros(int rows, int cols)

Methods inherited from class java.lang.Object

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

Methods inherited from interface io.github.kirstenali.deepj.tensor.TensorBackend

addInPlace, addScalarInPlace, divideInPlace, divideScalarInPlace, expInPlace, geluInPlace, logInPlace, multiplyInPlace, multiplyScalarInPlace, negInPlace, reluInPlace, sigmoidInPlace, sqrtInPlace, subtractInPlace, tanhInPlace

Constructor Details

MetalBackend

public MetalBackend()

Method Details

materializeTensor

public void materializeTensor(Tensor t)

Description copied from interface: TensorBackend

Materialize a tensor: flush any pending GPU computation and download the result to the tensor's CPU data[][]. Default is a no-op (for CpuBackend).

Specified by:

materializeTensor in interface TensorBackend

zeros

public Tensor zeros(int rows,
 int cols)

Specified by:

zeros in interface TensorBackend

ones

public Tensor ones(int rows,
 int cols)

Specified by:

ones in interface TensorBackend

random

public Tensor random(int rows,
 int cols,
 Random rand)

Specified by:

random in interface TensorBackend

causalMask

public Tensor causalMask(int size)

Specified by:

causalMask in interface TensorBackend

unflattenToTensor

public Tensor unflattenToTensor(double[] flat,
 int rows,
 int cols)

Specified by:

unflattenToTensor in interface TensorBackend

flattenTensor

public double[] flattenTensor(Tensor t)

Specified by:

flattenTensor in interface TensorBackend

matmul

public Tensor matmul(Tensor a,
 Tensor b)

Specified by:

matmul in interface TensorBackend

add

public Tensor add(Tensor a,
 Tensor b)

Specified by:

add in interface TensorBackend

subtract

public Tensor subtract(Tensor a,
 Tensor b)

Specified by:

subtract in interface TensorBackend

multiply

public Tensor multiply(Tensor a,
 Tensor b)

Specified by:

multiply in interface TensorBackend

divide

public Tensor divide(Tensor a,
 Tensor b)

Specified by:

divide in interface TensorBackend

addRowVector

public Tensor addRowVector(Tensor a,
 Tensor v)

Specified by:

addRowVector in interface TensorBackend

addBroadcastCols

public Tensor addBroadcastCols(Tensor a,
 Tensor v)

Specified by:

addBroadcastCols in interface TensorBackend

divideBroadcastCols

public Tensor divideBroadcastCols(Tensor a,
 Tensor v)

Specified by:

divideBroadcastCols in interface TensorBackend

subtractBroadcastCols

public Tensor subtractBroadcastCols(Tensor a,
 Tensor v)

Specified by:

subtractBroadcastCols in interface TensorBackend

multiplyBroadcastCols

public Tensor multiplyBroadcastCols(Tensor a,
 Tensor v)

Specified by:

multiplyBroadcastCols in interface TensorBackend

addBroadcastRows

public Tensor addBroadcastRows(Tensor a,
 Tensor v)

Specified by:

addBroadcastRows in interface TensorBackend

multiplyBroadcastRows

public Tensor multiplyBroadcastRows(Tensor a,
 Tensor v)

Specified by:

multiplyBroadcastRows in interface TensorBackend

multiplyScalar

public Tensor multiplyScalar(Tensor a,
 double scalar)

Specified by:

multiplyScalar in interface TensorBackend

addScalar

public Tensor addScalar(Tensor a,
 double scalar)

Specified by:

addScalar in interface TensorBackend

divideScalar

public Tensor divideScalar(Tensor a,
 double scalar)

Specified by:

divideScalar in interface TensorBackend

sumRows

public Tensor sumRows(Tensor a)

Specified by:

sumRows in interface TensorBackend

sumAlongRows

public Tensor sumAlongRows(Tensor a)

Specified by:

sumAlongRows in interface TensorBackend

sumAlongCols

public Tensor sumAlongCols(Tensor a)

Specified by:

sumAlongCols in interface TensorBackend

meanAlongRows

public Tensor meanAlongRows(Tensor a)

Specified by:

meanAlongRows in interface TensorBackend

varianceAlongRows

public Tensor varianceAlongRows(Tensor a)

Specified by:

varianceAlongRows in interface TensorBackend

maxAlongRows

public Tensor maxAlongRows(Tensor a)

Specified by:

maxAlongRows in interface TensorBackend

sum

public double sum(Tensor a)

Specified by:

sum in interface TensorBackend

sumAbs

public double sumAbs(Tensor a)

Specified by:

sumAbs in interface TensorBackend

clamp

public Tensor clamp(Tensor a,
 double min,
 double max)

Specified by:

clamp in interface TensorBackend

transpose

public Tensor transpose(Tensor a)

Specified by:

transpose in interface TensorBackend

pow

public Tensor pow(Tensor a,
 double exponent)

Specified by:

pow in interface TensorBackend

sqrt

public Tensor sqrt(Tensor a)

Specified by:

sqrt in interface TensorBackend

neg

public Tensor neg(Tensor a)

Specified by:

neg in interface TensorBackend

exp

public Tensor exp(Tensor a)

Specified by:

exp in interface TensorBackend

log

public Tensor log(Tensor a)

Specified by:

log in interface TensorBackend

tanh

public Tensor tanh(Tensor a)

Specified by:

tanh in interface TensorBackend

sigmoid

public Tensor sigmoid(Tensor a)

Specified by:

sigmoid in interface TensorBackend

relu

public Tensor relu(Tensor a)

Specified by:

relu in interface TensorBackend

reluBackward

public Tensor reluBackward(Tensor input,
 Tensor gradOutput)

Specified by:

reluBackward in interface TensorBackend

gelu

public Tensor gelu(Tensor a)

Specified by:

gelu in interface TensorBackend

geluBackward

public Tensor geluBackward(Tensor input,
 Tensor gradOutput)

Specified by:

geluBackward in interface TensorBackend

softmaxRows

public Tensor softmaxRows(Tensor logits)

Specified by:

softmaxRows in interface TensorBackend

softmaxBackward

public Tensor softmaxBackward(Tensor gradOutput,
 Tensor softmaxOut)

Specified by:

softmaxBackward in interface TensorBackend

crossEntropyLoss

public double crossEntropyLoss(Tensor logits,
 int[] targets)

Specified by:

crossEntropyLoss in interface TensorBackend

crossEntropyGradient

public Tensor crossEntropyGradient(Tensor logits,
 int[] targets)

Specified by:

crossEntropyGradient in interface TensorBackend

adamWUpdate

public void adamWUpdate(Tensor w,
 Tensor g,
 Tensor mt,
 Tensor vt,
 double lr,
 double beta1,
 double beta2,
 double eps,
 double weightDecay,
 double bc1,
 double bc2)

Description copied from interface: TensorBackend

In-place AdamW update. Mutates w, mt, vt.

Specified by:

adamWUpdate in interface TensorBackend

layerNormBackward

public Tensor layerNormBackward(Tensor dXHat,
 Tensor xHat,
 Tensor std,
 int dim)

Description copied from interface: TensorBackend

LayerNorm backward through normalization (given dXHat, xHat, std).

Specified by:

layerNormBackward in interface TensorBackend

get

public double get(Tensor t,
 int r,
 int c)

Specified by:

get in interface TensorBackend

set

public void set(Tensor t,
 int r,
 int c,
 double value)

Specified by:

set in interface TensorBackend

getRow

public Tensor getRow(Tensor t,
 int row)

Specified by:

getRow in interface TensorBackend

setRow

public void setRow(Tensor t,
 int row,
 Tensor source,
 int srcRow)

Specified by:

setRow in interface TensorBackend

sliceRows

public Tensor sliceRows(Tensor t,
 int[] rowIndices,
 int cols)

Description copied from interface: TensorBackend

Gather rows by index (for embedding lookup).

Specified by:

sliceRows in interface TensorBackend

scatterAddRows

public void scatterAddRows(Tensor target,
 int[] indices,
 Tensor grad)

Description copied from interface: TensorBackend

Scatter-add: target.data[indices[i]] += grad.data[i] (for embedding backward).

Specified by:

scatterAddRows in interface TensorBackend

sampleRows

public Tensor sampleRows(Tensor t,
 int n,
 Random rnd)

Description copied from interface: TensorBackend

Sample random rows from t.

Specified by:

sampleRows in interface TensorBackend

print

public void print(Tensor t,
 String label)

Specified by:

print in interface TensorBackend

releaseResources

public void releaseResources()

Description copied from interface: TensorBackend

Release backend-owned resources (GPU buffers, native handles, etc.). Default is a no-op for backends without external resources.

Specified by:

releaseResources in interface TensorBackend