from typing import Union, List
import numpy as np
import torch
from torch import nn
import torch_npu
class FluxPosEmbed(nn.Module):
def __init__(self, theta: int, axes_dim: List[int]):
super().__init__()
self.theta = theta
self.axes_dim = axes_dim
def forward(self, ids: torch.Tensor) -> torch.Tensor:
n_axes = ids.shape[-1]
cos_out = []
sin_out = []
pos = ids.float()
is_mps = ids.device.type == "mps"
freqs_dtype = torch.float32 if is_mps else torch.float64
for i in range(n_axes):
cos, sin = get_1d_rotary_pos_embed(
self.axes_dim[i],
pos[:, i],
theta=self.theta,
repeat_interleave_real=True,
use_real=True,
freqs_dtype=freqs_dtype,
)
cos_out.append(cos)
sin_out.append(sin)
freqs_cos = torch.cat(cos_out, dim=-1).to(ids.device)
freqs_sin = torch.cat(sin_out, dim=-1).to(ids.device)
return freqs_cos, freqs_sin
def get_1d_rotary_pos_embed(
dim: int,
pos: Union[np.ndarray, int],
theta: float = 10000.0,
use_real=False,
linear_factor=1.0,
ntk_factor=1.0,
repeat_interleave_real=True,
freqs_dtype=torch.float32,
):
"""
Precompute the frequency tensor for complex exponentials (cis) with given dimensions.
This function calculates a frequency tensor with complex exponentials using the given dimension 'dim' and the end
index 'end'. The 'theta' parameter scales the frequencies. The returned tensor contains complex values in complex64
data type.
Args:
dim (`int`): Dimension of the frequency tensor.
pos (`np.ndarray` or `int`): Position indices for the frequency tensor. [S] or scalar
theta (`float`, *optional*, defaults to 10000.0):
Scaling factor for frequency computation. Defaults to 10000.0.
use_real (`bool`, *optional*):
If True, return real part and imaginary part separately. Otherwise, return complex numbers.
linear_factor (`float`, *optional*, defaults to 1.0):
Scaling factor for the context extrapolation. Defaults to 1.0.
ntk_factor (`float`, *optional*, defaults to 1.0):
Scaling factor for the NTK-Aware RoPE. Defaults to 1.0.
repeat_interleave_real (`bool`, *optional*, defaults to `True`):
If `True` and `use_real`, real part and imaginary part are each interleaved with themselves to reach `dim`.
Otherwise, they are concateanted with themselves.
freqs_dtype (`torch.float32` or `torch.float64`, *optional*, defaults to `torch.float32`):
the dtype of the frequency tensor.
Returns:
`torch.Tensor`: Precomputed frequency tensor with complex exponentials. [S, D/2]
"""
assert dim % 2 == 0
if isinstance(pos, int):
pos = torch.arange(pos)
if isinstance(pos, np.ndarray):
pos = torch.from_numpy(pos)
theta = theta * ntk_factor
freqs = (
1.0
/ (theta ** (torch.arange(0, dim, 2, dtype=freqs_dtype, device=pos.device)[: (dim // 2)] / dim))
/ linear_factor
)
freqs = torch.outer(pos, freqs).float()
if use_real and repeat_interleave_real:
freqs_cos = freqs.cos().repeat_interleave(2, dim=1).float()
freqs_sin = freqs.sin().repeat_interleave(2, dim=1).float()
return freqs_cos, freqs_sin
elif use_real:
freqs_cos = torch.cat([freqs.cos(), freqs.cos()], dim=-1).float()
freqs_sin = torch.cat([freqs.sin(), freqs.sin()], dim=-1).float()
return freqs_cos, freqs_sin
else:
freqs_cis = torch.polar(torch.ones_like(freqs), freqs)
return freqs_cis