emmi.functional.rope¶
Functions¶
|
Applies Rotary Position Embeddings (RoPE) |
Module Contents¶
- emmi.functional.rope.rope(x, freqs)¶
Applies Rotary Position Embeddings (RoPE)
- Parameters:
x (torch.Tensor) – Vector to rotate (e.g., queries or keys of a transformer). Shape=(batch_size, num_heads, seqlen, head_dim).
freqs (tuple) – Complex tensor of frequencies for rotating x.
freqs – Sine/cosine frequencies for rotating x. For 1D, freqs is a tuple with length 1 with shape
(batch_size
num_heads
rotate. (num_dim_to_rotate) where num_dim_to_rotate is the number of dimensions to)
dimensional (If positions are higher)
rotation. (corresponds to frequencies of the nth axis for)
- Returns:
Rotated x.
- Return type:
torch.Tensor