Function apply_rope

lit_gpt/model.py:469–475

(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor)

def apply_rope(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> torch.Tensor:
    head_size = x.size(-1)
    x1 = x[..., : head_size // 2]  # (B, nh, T, hs/2)
    x2 = x[..., head_size // 2 :]  # (B, nh, T, hs/2)
    rotated = torch.cat((-x2, x1), dim=-1)  # (B, nh, T, hs)
    roped = (x * cos) + (rotated * sin)
    return roped.type_as(x)