rlx-sam2 0.2.4

SAM 2 (Hiera) for RLX
Documentation 
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

// RLX — versatile ML compiler + runtime.
// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
//
// This program is free software: you can redistribute it and/or modify
// it under the terms of the GNU General Public License as published by
// the Free Software Foundation, version 3.
//
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU General Public License
// along with this program. If not, see <https://www.gnu.org/licenses/>.

//! SAM2 axial 2-D RoPE (host + cos/sin tables for [`Op::Rope`] IR).

/// Per-token cos/sin for x-axis and y-axis halves (`[seq, head_dim/4]` each).
pub fn axial_rope_cos_sin_for_seq(
    end_x: usize,
    end_y: usize,
    n_tokens: usize,
    head_dim: usize,
    theta: f32,
    repeat_factor: usize,
) -> (Vec<f32>, Vec<f32>, Vec<f32>, Vec<f32>) {
    debug_assert!(head_dim.is_multiple_of(4));
    let q4 = head_dim / 4;
    let spatial = end_x * end_y;
    let mut freqs = vec![0f32; q4];
    for i in 0..q4 {
        freqs[i] = 1.0 / theta.powf((4 * i) as f32 / head_dim as f32);
    }
    let mut cs_x = vec![0f32; spatial * q4];
    let mut sn_x = vec![0f32; spatial * q4];
    let mut cs_y = vec![0f32; spatial * q4];
    let mut sn_y = vec![0f32; spatial * q4];
    for pos in 0..spatial {
        let tx = (pos % end_x) as f32;
        let ty = (pos / end_x) as f32;
        for c in 0..q4 {
            let ax = tx * freqs[c];
            let ay = ty * freqs[c];
            cs_x[pos * q4 + c] = ax.cos();
            sn_x[pos * q4 + c] = ax.sin();
            cs_y[pos * q4 + c] = ay.cos();
            sn_y[pos * q4 + c] = ay.sin();
        }
    }
    let mut cos_x = vec![0f32; n_tokens * q4];
    let mut sin_x = vec![0f32; n_tokens * q4];
    let mut cos_y = vec![0f32; n_tokens * q4];
    let mut sin_y = vec![0f32; n_tokens * q4];
    for tok in 0..n_tokens {
        let pos = tok / repeat_factor;
        for c in 0..q4 {
            cos_x[tok * q4 + c] = cs_x[pos * q4 + c];
            sin_x[tok * q4 + c] = sn_x[pos * q4 + c];
            cos_y[tok * q4 + c] = cs_y[pos * q4 + c];
            sin_y[tok * q4 + c] = sn_y[pos * q4 + c];
        }
    }
    (cos_x, sin_x, cos_y, sin_y)
}

/// Apply axial 2-D RoPE on `x` shaped `[nh, n_tokens, head_dim]`.
pub fn apply_axial_rope_2d(
    x: &[f32],
    nh: usize,
    n_tokens: usize,
    dh: usize,
    end_x: usize,
    end_y: usize,
    theta: f32,
    repeat_factor: usize,
) -> Vec<f32> {
    debug_assert!(
        dh.is_multiple_of(4),
        "RoPE expects head_dim multiple of 4 (got {dh})"
    );
    let half = dh / 2;
    let q4 = dh / 4;
    let spatial = end_x * end_y;
    debug_assert_eq!(
        n_tokens,
        spatial * repeat_factor,
        "RoPE token count mismatch"
    );

    let (cs_x, sn_x, cs_y, sn_y) =
        axial_rope_cos_sin_for_seq(end_x, end_y, n_tokens, dh, theta, repeat_factor);

    let mut out = vec![0f32; nh * n_tokens * dh];
    for h in 0..nh {
        for tok in 0..n_tokens {
            let pos = tok / repeat_factor;
            let src_base = (h * n_tokens + tok) * dh;
            let dst_base = src_base;
            for c in 0..q4 {
                let ix0 = src_base + 2 * c;
                let ix1 = src_base + 2 * c + 1;
                let x0 = x[ix0];
                let x1 = x[ix1];
                let co = cs_x[pos * q4 + c];
                let si = sn_x[pos * q4 + c];
                out[dst_base + 2 * c] = x0 * co - x1 * si;
                out[dst_base + 2 * c + 1] = x0 * si + x1 * co;
            }
            for c in 0..q4 {
                let ix0 = src_base + half + 2 * c;
                let ix1 = src_base + half + 2 * c + 1;
                let x0 = x[ix0];
                let x1 = x[ix1];
                let co = cs_y[pos * q4 + c];
                let si = sn_y[pos * q4 + c];
                out[dst_base + half + 2 * c] = x0 * co - x1 * si;
                out[dst_base + half + 2 * c + 1] = x0 * si + x1 * co;
            }
        }
    }
    out
}

#[cfg(test)]
mod tests {
    use super::*;

    /// Interleaved pair rotation on `[nh, n, half]` (matches [`apply_axial_rope_2d`] x/y halves).
    fn rope_interleaved_half(
        x: &[f32],
        cos: &[f32],
        sin: &[f32],
        nh: usize,
        n: usize,
        half: usize,
    ) -> Vec<f32> {
        let q4 = half / 2;
        let mut out = x.to_vec();
        for h in 0..nh {
            for tok in 0..n {
                let base = (h * n + tok) * half;
                for c in 0..q4 {
                    let ix0 = base + 2 * c;
                    let ix1 = base + 2 * c + 1;
                    let co = cos[tok * q4 + c];
                    let si = sin[tok * q4 + c];
                    let x0 = out[ix0];
                    let x1 = out[ix1];
                    out[ix0] = x0 * co - x1 * si;
                    out[ix1] = x0 * si + x1 * co;
                }
            }
        }
        out
    }

    #[test]
    fn dual_rope_n_matches_apply_axial() {
        let nh = 1usize;
        let n = 64usize;
        let dh = 256usize;
        let end_x = 8usize;
        let end_y = 8usize;
        let x: Vec<f32> = (0..nh * n * dh).map(|i| i as f32 * 0.001).collect();
        let host = apply_axial_rope_2d(&x, nh, n, dh, end_x, end_y, 10000.0, 1);
        let half = dh / 2;
        let (cx, sx, cy, sy) = axial_rope_cos_sin_for_seq(end_x, end_y, n, dh, 10000.0, 1);
        let mut lo = vec![0f32; nh * n * half];
        let mut hi = vec![0f32; nh * n * half];
        for h in 0..nh {
            for tok in 0..n {
                let base = (h * n + tok) * dh;
                let off = (h * n + tok) * half;
                lo[off..off + half].copy_from_slice(&x[base..base + half]);
                hi[off..off + half].copy_from_slice(&x[base + half..base + dh]);
            }
        }
        let lo_r = rope_interleaved_half(&lo, &cx, &sx, nh, n, half);
        let hi_r = rope_interleaved_half(&hi, &cy, &sy, nh, n, half);
        let mut manual = vec![0f32; nh * n * dh];
        for h in 0..nh {
            for tok in 0..n {
                let base = (h * n + tok) * dh;
                let off = (h * n + tok) * half;
                manual[base..base + half].copy_from_slice(&lo_r[off..off + half]);
                manual[base + half..base + dh].copy_from_slice(&hi_r[off..off + half]);
            }
        }
        let fd = host
            .iter()
            .zip(&manual)
            .map(|(a, b)| (a - b).abs())
            .fold(0f32, f32::max);
        assert!(fd < 2e-2, "dual rope_n vs axial max |Δ| = {fd:.3e}");
    }
}