File size: 5,796 Bytes
b2cb4a0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
#
# For licensing see accompanying LICENSE file.
# Copyright (c) 2025 Apple Inc. Licensed under MIT License.
#

import mlx.core as mx
import mlx.nn as nn
import math
from einops.array_api import rearrange
import torch
import numpy as np


class AbsolutePositionEncoding(nn.Module):
    def __init__(self, in_dim, embed_dim, include_input=False):
        super().__init__()
        self.in_dim = in_dim
        self.hidden_dim = embed_dim
        self.include_input = include_input
        assert embed_dim % in_dim == 0, "embed_dim must be divisible by in_dim"
        self.embed_dim = embed_dim + in_dim if include_input else embed_dim

    def __call__(self, pos):
        pos_embs = []
        for i in range(self.in_dim):
            pe = self.get_1d_pos_embed(pos[..., i])
            pos_embs.append(pe)
        if self.include_input:
            pos_embs.append(pos)
        pos_embs = mx.concatenate(pos_embs, axis=-1)
        return pos_embs

    def get_1d_pos_embed(self, pos):
        """
        https://github.com/facebookresearch/DiT/blob/main/models.py#L303
        """
        embed_dim = self.hidden_dim // (self.in_dim * 2)
        omega = 2 ** mx.linspace(0, math.log(224, 2) - 1, embed_dim).astype(mx.float32)
        omega *= math.pi

        if len(pos.shape) == 1:
            out = mx.einsum("m,d->md", pos, omega)  # (M, D/2), outer product
        elif len(pos.shape) == 2:
            out = mx.einsum("nm,d->nmd", pos, omega)

        emb_sin = mx.sin(out)
        emb_cos = mx.cos(out)  # (*, M, D/2)
        emb = mx.concatenate([emb_sin, emb_cos], axis=-1)  # (*, M, D)
        return emb


class FourierPositionEncoding(nn.Module):
    def __init__(
        self,
        in_dim: int,
        include_input: bool = False,
        min_freq_log2: float = 0,
        max_freq_log2: float = 12,
        num_freqs: int = 32,
        log_sampling: bool = True,
    ):
        super().__init__()
        self.in_dim = in_dim
        self.include_input = include_input
        self.min_freq_log2 = min_freq_log2
        self.max_freq_log2 = max_freq_log2
        self.num_freqs = num_freqs
        self.log_sampling = log_sampling
        self.create_embedding_fn()

    def create_embedding_fn(self):
        d = self.in_dim
        dim_out = 0
        if self.include_input:
            dim_out += d

        min_freq = self.min_freq_log2
        max_freq = self.max_freq_log2
        N_freqs = self.num_freqs

        if self.log_sampling:
            freq_bands = 2.0 ** mx.linspace(min_freq, max_freq, num=N_freqs)  # (nf,)
        else:
            freq_bands = mx.linspace(2.0**min_freq, 2.0**max_freq, num=N_freqs)  # (nf,)

        assert mx.isfinite(
            freq_bands
        ).all(), f"nan: {mx.isnan(freq_bands).any()} inf: {mx.isinf(freq_bands).any()}"

        self.freq_bands = freq_bands
        self.embed_dim = dim_out + d * self.freq_bands.size * 2

    def __call__(
        self,
        pos,
    ):
        """
        Get the positional encoding for each coordinate.
        Args:
            pos:
                (*, in_dim)
        Returns:
            out:
                (*, in_dimitional_encoding)
        """

        out = []
        if self.include_input:
            out = [pos]  # (*, in_dim)

        pos = pos[..., None] * self.freq_bands  # (*b, d, nf)

        out += [
            mx.sin(pos).flatten(start_axis=-2),  # (*b, d*nf)
            mx.cos(pos).flatten(start_axis=-2),  # (*b, d*nf)
        ]

        out = mx.concatenate(out, axis=-1)  # (*b, 2 * in_dim * nf (+ in_dim))
        return out


def compute_axial_cis(
    ts,
    in_dim: int,
    dim: int,
    theta: float = 100.0,
):
    B, N, D = ts.shape
    freqs_all = []
    interval = 2 * in_dim
    for i in range(in_dim):
        freq = 1.0 / (
            theta
            ** (
                mx.arange(0, dim, interval)[: (dim // interval)].astype(mx.float32)
                / dim
            )
        )
        t = ts[..., i].flatten()
        freq_i = mx.outer(t, freq)
        freq_cis_i = polar(mx.ones_like(freq_i), freq_i)
        freq_cis_i = freq_cis_i.reshape((B, N, -1))
        freqs_all.append(freq_cis_i)
    freqs_cis = mx.concatenate(freqs_all, axis=-1)
    return freqs_cis


def polar(a, b):
    return a * mx.exp(1j * b)


def view_as_complex(x):
    x = x.astype(mx.float32)
    x = x.reshape(*x.shape[:-1], -1, 2)  # (..., dim/2, 2)
    return x[..., 0] + 1j * x[..., 1]  # real, imag


def view_as_real(input):
    return mx.stack([input.real, input.imag], axis=-1)  # (..., dim)


def apply_rotary_emb(xq: mx.array, xk: mx.array, freqs_cis: mx.array):
    # xq, xk: shape (..., dim)
    # freqs_cis: shape (..., dim // 2, 2) where last dim = [cos, sin]
    xq_ = view_as_complex(xq)
    xk_ = view_as_complex(xk)

    # Apply complex multiplication
    xq_out = xq_ * freqs_cis
    xk_out = xk_ * freqs_cis

    # Reconstruct to original shape
    xq_out = view_as_real(xq_out).flatten(start_axis=3)
    xk_out = view_as_real(xk_out).flatten(start_axis=3)

    return xq_out.astype(xq.dtype), xk_out.astype(xk.dtype)


class AxialRotaryPositionEncoding(nn.Module):
    def __init__(
        self,
        in_dim,
        embed_dim,
        num_heads,
        base=100.0,
    ):
        super().__init__()
        self.in_dim = in_dim
        self.num_heads = num_heads
        self.embed_dim = embed_dim // num_heads
        self.base = base

    def __call__(self, xq, xk, pos):
        """
        xq: [B, H, N, D]
        xk: [B, H, N, D]
        pos: [B, N, in_dim]
        """
        if pos.ndim == 2:
            pos = pos[..., None]
        freqs_cis = compute_axial_cis(pos, self.in_dim, self.embed_dim, self.base)
        freqs_cis = mx.expand_dims(freqs_cis, axis=1)

        return apply_rotary_emb(xq, xk, freqs_cis)