tribev2 0.0.4

TRIBE v2 — multimodal fMRI brain encoding model inference in Rust
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

use burn::prelude::*;
use burn::nn::{Linear, LinearConfig};
use burn::tensor::activation::softmax;

/// Multi-head self-attention matching x_transformers' non-causal encoder.
///
/// 1. **Manual QK^T + softmax + V** — the cubecl flash-attention kernel
///    hardcodes `causal: true` (wrong for a bidirectional encoder) and is
///    slower than plain matmuls for the small T=100 sequence used here.
///
/// 2. **Fused QKV projection + `narrow` split** — one [D→3D] matmul,
///    then view-level splits; avoids 3 separate GPU buffer copies.
///
/// 3. **Fused RoPE** (when `wgpu-kernels-metal`) — a single CubeCL kernel
///    replaces the slice×3 + mul×4 + sub + add + cat chain (~10 dispatches,
///    4 intermediate allocations per apply).
#[derive(Module, Debug)]
pub struct Attention<B: Backend> {
    pub to_qkv: Linear<B>,
    pub to_out:  Linear<B>,
    pub heads:    usize,
    pub dim_head: usize,
    pub scale:    f32,
}

impl<B: Backend> Attention<B> {
    pub fn new(dim: usize, heads: usize, device: &B::Device) -> Self {
        let dim_head = dim / heads;
        Self {
            to_qkv: LinearConfig::new(dim, 3 * dim).with_bias(false).init(device),
            to_out:  LinearConfig::new(dim,     dim).with_bias(false).init(device),
            heads,
            dim_head,
            scale: (dim_head as f32).powf(-0.5),
        }
    }
}

// ── forward — standard burn-ops path ──────────────────────────────────────
#[cfg(not(feature = "wgpu-kernels-metal"))]
impl<B: Backend> Attention<B> {
    pub fn forward(
        &self,
        x:          Tensor<B, 3>,
        rotary_cos: Option<&Tensor<B, 2>>,
        rotary_sin: Option<&Tensor<B, 2>>,
    ) -> Tensor<B, 3> {
        attention_forward(self, x, rotary_cos, rotary_sin)
    }
}

// ── forward — fused-kernel path ────────────────────────────────────────────
#[cfg(feature = "wgpu-kernels-metal")]
impl<B: Backend + crate::model_burn::FusedOps> Attention<B> {
    pub fn forward(
        &self,
        x:          Tensor<B, 3>,
        rotary_cos: Option<&Tensor<B, 2>>,
        rotary_sin: Option<&Tensor<B, 2>>,
    ) -> Tensor<B, 3> {
        attention_forward(self, x, rotary_cos, rotary_sin)
    }
}

// ── Shared implementation ──────────────────────────────────────────────────

/// Inner attention logic shared between both cfg branches.
/// The `where` bound differs per cfg but the body is identical.
#[cfg(not(feature = "wgpu-kernels-metal"))]
fn attention_forward<B: Backend>(
    this:       &Attention<B>,
    x:          Tensor<B, 3>,
    rotary_cos: Option<&Tensor<B, 2>>,
    rotary_sin: Option<&Tensor<B, 2>>,
) -> Tensor<B, 3> {
    _attention_forward_body(this, x, rotary_cos, rotary_sin)
}

#[cfg(feature = "wgpu-kernels-metal")]
fn attention_forward<B: Backend + crate::model_burn::FusedOps>(
    this:       &Attention<B>,
    x:          Tensor<B, 3>,
    rotary_cos: Option<&Tensor<B, 2>>,
    rotary_sin: Option<&Tensor<B, 2>>,
) -> Tensor<B, 3> {
    _attention_forward_body(this, x, rotary_cos, rotary_sin)
}

// Both forward-body variants are written out in full below.
// (They're identical — the only difference is the where-bound, which
//  Rust requires to be part of the function signature.)

#[cfg(not(feature = "wgpu-kernels-metal"))]
fn _attention_forward_body<B: Backend>(
    this:       &Attention<B>,
    x:          Tensor<B, 3>,
    rotary_cos: Option<&Tensor<B, 2>>,
    rotary_sin: Option<&Tensor<B, 2>>,
) -> Tensor<B, 3> {
    let [b, n, _d] = x.dims();
    let (h, dh) = (this.heads, this.dim_head);
    let dim     = h * dh;

    let qkv = this.to_qkv.forward(x);
    let q = qkv.clone().narrow(2, 0,     dim)
                .reshape([b, n, h, dh]).swap_dims(1, 2);
    let k = qkv.clone().narrow(2, dim,   dim)
                .reshape([b, n, h, dh]).swap_dims(1, 2);
    let v = qkv       .narrow(2, 2*dim,  dim)
                .reshape([b, n, h, dh]).swap_dims(1, 2);

    let (q, k) = match (rotary_cos, rotary_sin) {
        (Some(cos), Some(sin)) => (
            apply_rotary_precomputed(q, cos, sin),
            apply_rotary_precomputed(k, cos, sin),
        ),
        _ => (q, k),
    };

    let q    = q.mul_scalar(this.scale);
    let attn = softmax(q.matmul(k.swap_dims(2, 3)), 3);
    let out  = attn.matmul(v);
    let out  = out.swap_dims(1, 2).flatten(2, 3);
    this.to_out.forward(out)
}

#[cfg(feature = "wgpu-kernels-metal")]
fn _attention_forward_body<B: Backend + crate::model_burn::FusedOps>(
    this:       &Attention<B>,
    x:          Tensor<B, 3>,
    rotary_cos: Option<&Tensor<B, 2>>,
    rotary_sin: Option<&Tensor<B, 2>>,
) -> Tensor<B, 3> {
    let [b, n, _d] = x.dims();
    let (h, dh) = (this.heads, this.dim_head);
    let dim     = h * dh;

    let qkv = this.to_qkv.forward(x);
    let q = qkv.clone().narrow(2, 0,     dim)
                .reshape([b, n, h, dh]).swap_dims(1, 2);
    let k = qkv.clone().narrow(2, dim,   dim)
                .reshape([b, n, h, dh]).swap_dims(1, 2);
    let v = qkv       .narrow(2, 2*dim,  dim)
                .reshape([b, n, h, dh]).swap_dims(1, 2);

    let (q, k) = match (rotary_cos, rotary_sin) {
        (Some(cos), Some(sin)) => (
            // Fused CubeCL kernel: 1 dispatch vs ~10, no intermediate allocs
            rotate_qk::<B>(q, cos, sin),
            rotate_qk::<B>(k, cos, sin),
        ),
        _ => (q, k),
    };

    let q    = q.mul_scalar(this.scale);
    let attn = softmax(q.matmul(k.swap_dims(2, 3)), 3);
    let out  = attn.matmul(v);
    let out  = out.swap_dims(1, 2).flatten(2, 3);
    this.to_out.forward(out)
}

// ── RoPE dispatch ─────────────────────────────────────────────────────────

/// Standard burn-ops RoPE (used when wgpu-kernels-metal is off).
#[cfg(not(feature = "wgpu-kernels-metal"))]
fn apply_rotary_precomputed<B: Backend>(
    x:   Tensor<B, 4>,
    cos: &Tensor<B, 2>,
    sin: &Tensor<B, 2>,
) -> Tensor<B, 4> {
    let [b, h, n, d] = x.dims();
    let half    = cos.dims()[1];
    let rot_dim = half * 2;

    let x_rot = x.clone().slice([0..b, 0..h, 0..n, 0..rot_dim]);
    let x1    = x_rot.clone().slice([0..b, 0..h, 0..n, 0..half]);
    let x2    = x_rot        .slice([0..b, 0..h, 0..n, half..rot_dim]);

    let c = cos.clone().unsqueeze_dim::<3>(0).unsqueeze_dim::<4>(0);
    let s = sin.clone().unsqueeze_dim::<3>(0).unsqueeze_dim::<4>(0);

    let r1 = x1.clone() * c.clone() - x2.clone() * s.clone();
    let r2 = x2          * c         + x1          * s;
    let rotated = Tensor::cat(vec![r1, r2], 3);

    if rot_dim < d {
        let x_pass = x.slice([0..b, 0..h, 0..n, rot_dim..d]);
        Tensor::cat(vec![rotated, x_pass], 3)
    } else {
        rotated
    }
}

/// Fused CubeCL RoPE (used when wgpu-kernels-metal is on).
#[cfg(feature = "wgpu-kernels-metal")]
fn rotate_qk<B: Backend + crate::model_burn::FusedOps>(
    x:   Tensor<B, 4>,
    cos: &Tensor<B, 2>,
    sin: &Tensor<B, 2>,
) -> Tensor<B, 4> {
    B::rope_rotate(x, cos.clone(), sin.clone())
}