oa::TransformerBlock

TransformerBlock class in the OA Ml public surface.

ClC++ public headers

Inheritance

public Module

Public Methods

oa::AttentionMode oa::TransformerBlock::attentionMode()

oa::I32 oa::TransformerBlock::dFF()

oa::I32 oa::TransformerBlock::dModel()

void oa::TransformerBlock::enableAdaptiveConditioning(oa::I32 inConditionDim)

oa::Matrix oa::TransformerBlock::forward(const oa::Matrix & inX)

oa::Matrix oa::TransformerBlock::forwardConditioned(const oa::Matrix & inX, const oa::Matrix & inCondition, const oa::Matrix & inAdditiveMask = = {})

oa::Matrix oa::TransformerBlock::forwardMasked(const oa::Matrix & inX, const oa::Matrix & inAdditiveMask)

void oa::TransformerBlock::init(oa::I32 inDModel, oa::I32 inDFF, oa::I32 inSeqLen, oa::F32 inEps = 1e-5f)

void oa::TransformerBlock::init(oa::I32 inDModel, oa::I32 inDFF, oa::I32 inSeqLen, oa::I32 inNumHeads, oa::F32 inEps)

void oa::TransformerBlock::initMoe(oa::I32 inDModel, oa::I32 inExpertDFF, oa::I32 inSeqLen, oa::I32 inNumExperts, oa::I32 inExpertsPerToken, oa::F32 inEps = 1e-5f)

void oa::TransformerBlock::initMoe(oa::I32 inDModel, oa::I32 inExpertDFF, oa::I32 inSeqLen, oa::I32 inNumHeads, oa::I32 inNumExperts, oa::I32 inExpertsPerToken, oa::F32 inEps)

bool oa::TransformerBlock::isMoe()

Moe * oa::TransformerBlock::moe()

const Moe * oa::TransformerBlock::moe()

oa::I32 oa::TransformerBlock::numHeads()

oa::I32 oa::TransformerBlock::seqLen()

void oa::TransformerBlock::setAttentionMode(oa::AttentionMode inMode)

void oa::TransformerBlock::setSeqLen(oa::I32 inSeqLen)

Constructor & Destructor Documentation

oa::TransformerBlock::TransformerBlock()

No public source comment is attached to this declaration.

oa::TransformerBlock::TransformerBlock(
	oa::I32 inDModel,
	oa::I32 inDFF,
	oa::I32 inSeqLen,
	oa::F32 inEps = 1e-5f
)

No public source comment is attached to this declaration.

Parameters

inDModel

oa::I32

—

inDFF

oa::I32

—

inSeqLen

oa::I32

—

inEps

oa::F32

Default: 1e-5f

oa::TransformerBlock::TransformerBlock(
	oa::I32 inDModel,
	oa::I32 inDFF,
	oa::I32 inSeqLen,
	oa::I32 inNumHeads,
	oa::F32 inEps
)

No public source comment is attached to this declaration.

Parameters

inDModel

oa::I32

—

inDFF

oa::I32

—

inSeqLen

oa::I32

—

inNumHeads

oa::I32

—

inEps

oa::F32

—

oa::TransformerBlock::TransformerBlock(
	oa::I32 inDModel,
	oa::I32 inExpertDFF,
	oa::I32 inSeqLen,
	oa::I32 inNumExperts,
	oa::I32 inExpertsPerToken,
	oa::F32 inEps = 1e-5f
)

No public source comment is attached to this declaration.

Parameters

inDModel

oa::I32

—

inExpertDFF

oa::I32

—

inSeqLen

oa::I32

—

inNumExperts

oa::I32

—

inExpertsPerToken

oa::I32

—

inEps

oa::F32

Default: 1e-5f

oa::TransformerBlock::TransformerBlock(
	oa::I32 inDModel,
	oa::I32 inExpertDFF,
	oa::I32 inSeqLen,
	oa::I32 inNumHeads,
	oa::I32 inNumExperts,
	oa::I32 inExpertsPerToken,
	oa::F32 inEps
)

No public source comment is attached to this declaration.

Parameters

inDModel

oa::I32

—

inExpertDFF

oa::I32

—

inSeqLen

oa::I32

—

inNumHeads

oa::I32

—

inNumExperts

oa::I32

—

inExpertsPerToken

oa::I32

—

inEps

oa::F32

—

Public Method Documentation

oa::AttentionMode oa::TransformerBlock::attentionMode()

No public source comment is attached to this declaration.

Returns

oa::AttentionMode

The declared return value.

oa::I32 oa::TransformerBlock::dFF()

No public source comment is attached to this declaration.

Returns

oa::I32

The declared return value.

oa::I32 oa::TransformerBlock::dModel()

No public source comment is attached to this declaration.

Returns

oa::I32

The declared return value.

void oa::TransformerBlock::enableAdaptiveConditioning(
	oa::I32 inConditionDim
)

Enables DiT-style AdaLN-Zero modulation. The ordinary language-model path remains unchanged until this is explicitly enabled by a conditioned model.

Parameters

inConditionDim

oa::I32

—

Returns

void

The declared return value.

oa::Matrix oa::TransformerBlock::forward(
	const oa::Matrix & inX
)

forward: x [B*S, D] → [B*S, D]. inSeqLen separates sequences within the flattened batch. Causal visibility remains the language-model default; flow/denoising encoders explicitly select Bidirectional.

Parameters

inX

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

oa::Matrix oa::TransformerBlock::forwardConditioned(
	const oa::Matrix & inX,
	const oa::Matrix & inCondition,
	const oa::Matrix & inAdditiveMask = = {}
)

No public source comment is attached to this declaration.

Parameters

inX

const oa::Matrix &

—

inCondition

const oa::Matrix &

—

inAdditiveMask

const oa::Matrix &

Default: = {}

Returns

oa::Matrix

The declared return value.

oa::Matrix oa::TransformerBlock::forwardMasked(
	const oa::Matrix & inX,
	const oa::Matrix & inAdditiveMask
)

Same block with an explicit additive attention mask [B*H*S,S]. The mask is shared by language padding and bidirectional denoising; Flash attention is intentionally bypassed because its current kernel is causal-only.

Parameters

inX

const oa::Matrix &

—

inAdditiveMask

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

void oa::TransformerBlock::init(
	oa::I32 inDModel,
	oa::I32 inDFF,
	oa::I32 inSeqLen,
	oa::F32 inEps = 1e-5f
)

No public source comment is attached to this declaration.

Parameters

inDModel

oa::I32

—

inDFF

oa::I32

—

inSeqLen

oa::I32

—

inEps

oa::F32

Default: 1e-5f

Returns

void

The declared return value.

void oa::TransformerBlock::init(
	oa::I32 inDModel,
	oa::I32 inDFF,
	oa::I32 inSeqLen,
	oa::I32 inNumHeads,
	oa::F32 inEps
)

No public source comment is attached to this declaration.

Parameters

inDModel

oa::I32

—

inDFF

oa::I32

—

inSeqLen

oa::I32

—

inNumHeads

oa::I32

—

inEps

oa::F32

—

Returns

void

The declared return value.

void oa::TransformerBlock::initMoe(
	oa::I32 inDModel,
	oa::I32 inExpertDFF,
	oa::I32 inSeqLen,
	oa::I32 inNumExperts,
	oa::I32 inExpertsPerToken,
	oa::F32 inEps = 1e-5f
)

No public source comment is attached to this declaration.

Parameters

inDModel

oa::I32

—

inExpertDFF

oa::I32

—

inSeqLen

oa::I32

—

inNumExperts

oa::I32

—

inExpertsPerToken

oa::I32

—

inEps

oa::F32

Default: 1e-5f

Returns

void

The declared return value.

void oa::TransformerBlock::initMoe(
	oa::I32 inDModel,
	oa::I32 inExpertDFF,
	oa::I32 inSeqLen,
	oa::I32 inNumHeads,
	oa::I32 inNumExperts,
	oa::I32 inExpertsPerToken,
	oa::F32 inEps
)

No public source comment is attached to this declaration.

Parameters

inDModel

oa::I32

—

inExpertDFF

oa::I32

—

inSeqLen

oa::I32

—

inNumHeads

oa::I32

—

inNumExperts

oa::I32

—

inExpertsPerToken

oa::I32

—

inEps

oa::F32

—

Returns

void

The declared return value.

bool oa::TransformerBlock::isMoe()

No public source comment is attached to this declaration.

Returns

bool

The declared return value.

Moe * oa::TransformerBlock::moe()

No public source comment is attached to this declaration.

Returns

Moe *

The declared return value.

const Moe * oa::TransformerBlock::moe()

No public source comment is attached to this declaration.

Returns

const Moe *

The declared return value.

oa::I32 oa::TransformerBlock::numHeads()

No public source comment is attached to this declaration.

Returns

oa::I32

The declared return value.

oa::I32 oa::TransformerBlock::seqLen()

No public source comment is attached to this declaration.

Returns

oa::I32

The declared return value.

void oa::TransformerBlock::setAttentionMode(
	oa::AttentionMode inMode
)

No public source comment is attached to this declaration.

Parameters

inMode

oa::AttentionMode

—

Returns

void

The declared return value.

void oa::TransformerBlock::setSeqLen(
	oa::I32 inSeqLen
)

The weights are sequence-length independent. Updating the runtime length only changes the B/S view and its causal mask, which lets one block serve fixed-length training and growing-prefix autoregressive generation.

Parameters

inSeqLen

oa::I32

—

Returns

void

The declared return value.