WhisperDecoder class - whisper_decoder library

Constructors

WhisperDecoder({required int vocabSize, required int embedDim, required int numHeads, required int numLayers, int nCtx = 448, Device device = Device.CPU})

blocks → List<WhisperDecoderBlock>: final
device → Device: final
embedDim → int: final
hashCode → int: The hash code for this object.
no setterinherited
ln → LayerNorm: final
nCtx → int: final
numHeads → int: final
numLayers → int: final
positionalEmbedding → Tensor: final
runtimeType → Type: A representation of the runtime type of the object.
no setterinherited
tokenEmbedding → Embedding: final
training ↔ bool: Whether this module is in training mode. Layers that behave differently between training and inference (e.g. Dropout) read this flag in their call method. Defaults to training mode.
getter/setter pairinherited
vocabSize → int: final

eval() → void: Put this module (and any registered submodules) into evaluation mode.
inherited
forward(Tensor tokens) → Tensor: Run the decoder on the full token prefix [B, T]. Returns hidden states [B, T, embedDim].
greedyDecode({required List<int> startTokens, required int eot, int maxLen = 100, List<int> initialSuppress = const []}) → List<int>: Greedy decode from startTokens up to maxLen, stopping when eot is emitted. Caller must have called primeCrossAttn first.
logitsLastToken(Tensor hidden) → Tensor: Given decoder hidden [B, T, C], project only the last position into vocab logits [B, V] using the tied token-embedding matrix.
noSuchMethod(Invocation invocation) → dynamic: Invoked when a nonexistent method or property is accessed.
inherited
parameters() → List<Tensor>: Trainable tensors owned by this module (and its submodules).
override
primeCrossAttn(Tensor encoderMemory) → void
submodules() → List<Module>: Submodules owned by this module. Subclasses that compose other modules should override this so train() / eval() propagate. Default: empty.
inherited
toString() → String: A string representation of this object.
inherited
train() → void: Put this module (and any registered submodules) into training mode.
inherited
zeroGrad() → void: Zero every parameter's gradient. Safe to call before each backward.
inherited