hasktorch-gradually-typed-0.2.0.0: experimental project for hasktorch
Safe HaskellSafe-Inferred
LanguageHaskell2010

Torch.GraduallyTyped.NN.Transformer.T5.ThreeB

Synopsis

Documentation

type T5ThreeBNumLayers = 24 Source #

T5-3B number of layers. 'num_layers = 24'

t5ThreeBNumLayers :: SNat T5ThreeBNumLayers Source #

T5-3B number of layers singleton.

type T5ThreeBHeadDim = 'Dim ('Name "*") ('Size 32) Source #

T5-3B number of attention heads. 'n_heads = 32'

type T5ThreeBHeadEmbedDim = 'Dim ('Name "*") ('Size 128) Source #

T5-3B head embedding dimension. 'd_kv = 128'

type T5ThreeBEmbedDim = 'Dim ('Name "*") ('Size 4096) Source #

T5-3B embedding dimension. 'inner_dim = n_heads * d_kv = 4096'

type T5ThreeBInputEmbedDim = 'Dim ('Name "*") ('Size 1024) Source #

T5-3B model dimension. 'd_model = 1024'

type T5ThreeBFFNDim = 'Dim ('Name "*") ('Size 16384) Source #

T5-3B feed-forward network dimension. 'd_ff = 16384'

type T5ThreeBVocabDim = 'Dim ('Name "*") ('Size 32128) Source #

T5-3B vocabulary dimension. 'vocab_size = 32128'

t5ThreeBSpec :: STransformerHead transformerHead -> SGradient gradient -> SDevice device -> SHasDropout hasDropout -> ModelSpec (T5ThreeB transformerHead gradient device hasDropout) Source #

T5-3B model specification.