Open fabianlim opened 3 weeks ago
This could be very possible by just providing the correct sin and cos values adjusted according to position ids. This can be done outside of the kernel and then passed in:
sin
cos
def _rope_embedding( Q, Q_row_stride, cos, cos_row_stride, sin, sin_row_stride, seqlen, head_dim : tl.constexpr, n_heads : tl.constexpr, BACKWARD_PASS : tl.constexpr, BLOCK_SIZE : tl.constexpr, ):
This could be very possible by just providing the correct
sin
andcos
values adjusted according to position ids. This can be done outside of the kernel and then passed in: