Application Scenarios

  • It is usually used before the attention mechanism and between the attention and the feedforward network ffn.
    Figure 1 Context of the RmsNormOperation operator
  • Application scenario when rstd is set to true
    • Input: xDim:[8,8,8]; gammaDim:[8]; Output: rstdDim:[8, 8, 1]

      i

      xDimNum - gammaDimNum

      condition(i >= (xDimNum - gammaDimNum))

      rstdDimNum

      0

      2

      false

      xDim[0]

      1

      2

      false

      xDim[1]

      2

      2

      true

      1

    • Input: xDim:[8,8,8]; gammaDim:[8,8]; Output: rstdDim:[8, 1, 1]

      i

      xDimNum - gammaDimNum

      condition(i >= (xDimNum - gammaDimNum))

      rstdDimNum

      0

      1

      false

      xDim[0]

      1

      1

      true

      1

      2

      1

      true

      1

    • Input: xDim: [8, 16, 32]; gammaDim: [8, 16, 32]; Output: rstdDim: [1, 1, 1]

      i

      xDimNum - gammaDimNum

      condition(i >= (xDimNum - gammaDimNum))

      rstdDimNum

      0

      0

      true

      1

      1

      0

      true

      1

      2

      0

      true

      1