VidChain-exercise
/
VTimeLLM
/flash-attention
/csrc
/composable_kernel
/docs
/conceptual
/Composable-Kernel-math.rst
| .. meta:: | |
| :description: Composable Kernel mathematical basis | |
| :keywords: composable kernel, CK, ROCm, API, mathematics, algorithm | |
| .. _supported-primitives: | |
| ******************************************************************** | |
| Composable Kernel mathematical basis | |
| ******************************************************************** | |
| This is an introduction to the math which underpins the algorithms implemented in Composable Kernel. | |
| For vectors :math:`x^{(1)}, x^{(2)}, \ldots, x^{(T)}` of size :math:`B` you can decompose the | |
| softmax of concatenated :math:`x = [ x^{(1)}\ | \ \ldots \ | \ x^{(T)} ]` as, | |
| .. math:: | |
| :nowrap: | |
| \begin{align} | |
| m(x) & = m( [ x^{(1)}\ | \ \ldots \ | \ x^{(T)} ] ) = \max( m(x^{(1)}),\ldots, m(x^{(T)}) ) \\ | |
| f(x) & = [\exp( m(x^{(1)}) - m(x) ) f( x^{(1)} )\ | \ \ldots \ | \ \exp( m(x^{(T)}) - m(x) ) f( x^{(T)} )] \\ | |
| z(x) & = \exp( m(x^{(1)}) - m(x) )\ z(x^{(1)}) + \ldots + \exp( m(x^{(T)}) - m(x) )\ z(x^{(1)}) \\ | |
| \operatorname{softmax}(x) &= f(x)\ / \ z(x) | |
| \end{align} | |
| where :math:`f(x^{(j)}) = \exp( x^{(j)} - m(x^{(j)}) )` is of size :math:`B` and | |
| :math:`z(x^{(j)}) = f(x_1^{(j)})+ \ldots+ f(x_B^{(j)})` is a scalar. | |
| For a matrix :math:`X` composed of :math:`T_r \times T_c` tiles, :math:`X_{ij}`, of size | |
| :math:`B_r \times B_c` you can compute the row-wise softmax as follows. | |
| For :math:`j` from :math:`1` to :math:`T_c`, and :math:`i` from :math:`1` to :math:`T_r` calculate, | |
| .. math:: | |
| :nowrap: | |
| \begin{align} | |
| \tilde{m}_{ij} &= \operatorname{rowmax}( X_{ij} ) \\ | |
| \tilde{P}_{ij} &= \exp(X_{ij} - \tilde{m}_{ij} ) \\ | |
| \tilde{z}_{ij} &= \operatorname{rowsum}( P_{ij} ) \\ | |
| \end{align} | |
| If :math:`j=1`, initialize running max, running sum, and the first column block of the output, | |
| .. math:: | |
| :nowrap: | |
| \begin{align} | |
| m_i &= \tilde{m}_{i1} \\ | |
| z_i &= \tilde{z}_{i1} \\ | |
| \tilde{Y}_{i1} &= \diag(\tilde{z}_{ij})^{-1} \tilde{P}_{i1} | |
| \end{align} | |
| Else if :math:`j>1`, | |
| 1. Update running max, running sum and column blocks :math:`k=1` to :math:`k=j-1` | |
| .. math:: | |
| :nowrap: | |
| \begin{align} | |
| m^{new}_i &= \max(m_i, \tilde{m}_{ij} ) \\ | |
| z^{new}_i &= \exp(m_i - m^{new}_i)\ z_i + \exp( \tilde{m}_{ij} - m^{new}_i )\ \tilde{z}_{ij} \\ | |
| Y_{ik} &= \diag(z^{new}_{i})^{-1} \diag(z_{i}) \exp(m_i - m^{new}_i)\ Y_{ik} | |
| \end{align} | |
| 2. Initialize column block :math:`j` of output and reset running max and running sum variables: | |
| .. math:: | |
| :nowrap: | |
| \begin{align} | |
| \tilde{Y}_{ij} &= \diag(z^{new}_{i})^{-1} \exp(\tilde{m}_{ij} - m^{new}_i ) \tilde{P}_{ij} \\ | |
| z_i &= z^{new}_i \\ | |
| m_i &= m^{new}_i \\ | |
| \end{align} | |