burn-msa 0.2.4

MiniMax Sparse Attention (MSA) - block-sparse attention with per-GQA-group top-k block selection for Burn
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
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
use burn::module::Module;
use burn::nn::{Initializer, Linear, LinearConfig};
use burn::tensor::activation;
use burn::tensor::backend::Backend;
use burn::tensor::{Int, Tensor};

use crate::config::MsaConfig;

#[derive(Module, Debug)]
pub struct SparseAttention<B: Backend> {
    pub q_proj: Linear<B>,
    pub k_proj: Linear<B>,
    pub v_proj: Linear<B>,
    pub out_proj: Linear<B>,
    pub cfg: MsaConfig,
}

impl<B: Backend> SparseAttention<B> {
    pub fn new(cfg: &MsaConfig, device: &B::Device) -> Self {
        Self {
            q_proj: LinearConfig::new(cfg.d_model, cfg.n_heads_q * cfg.d_head)
                .with_bias(false)
                .with_initializer(Initializer::XavierUniform { gain: 1.0 })
                .init(device),
            k_proj: LinearConfig::new(cfg.d_model, cfg.n_heads_kv * cfg.d_head)
                .with_bias(false)
                .with_initializer(Initializer::XavierUniform { gain: 1.0 })
                .init(device),
            v_proj: LinearConfig::new(cfg.d_model, cfg.n_heads_kv * cfg.d_head)
                .with_bias(false)
                .with_initializer(Initializer::XavierUniform { gain: 1.0 })
                .init(device),
            out_proj: LinearConfig::new(cfg.n_heads_q * cfg.d_head, cfg.d_model)
                .with_bias(false)
                .with_initializer(Initializer::XavierUniform { gain: 1.0 })
                .init(device),
            cfg: cfg.clone(),
        }
    }

    fn to_4d(t: Tensor<B, 3>, n_heads: usize, d_head: usize) -> Tensor<B, 4> {
        let [batch, seq, _] = t.dims();
        t.reshape::<4, _>([batch, seq, n_heads, d_head])
            .swap_dims(1, 2)
    }

    pub fn forward_sparse(
        &self,
        q: Tensor<B, 3>,
        k: Tensor<B, 3>,
        v: Tensor<B, 3>,
        block_indices: Tensor<B, 4, Int>,
    ) -> Tensor<B, 3> {
        self.forward_sparse_with_weights(q, k, v, block_indices).0
    }

    pub fn forward_sparse_with_weights(
        &self,
        q: Tensor<B, 3>,
        k: Tensor<B, 3>,
        v: Tensor<B, 3>,
        block_indices: Tensor<B, 4, Int>,
    ) -> (Tensor<B, 3>, Tensor<B, 4>) {
        let (out_4d, block_attn) = sparse_attn_batched_gqa(
            Self::to_4d(q, self.cfg.n_heads_q, self.cfg.d_head),
            Self::to_4d(k, self.cfg.n_heads_kv, self.cfg.d_head),
            Self::to_4d(v, self.cfg.n_heads_kv, self.cfg.d_head),
            block_indices,
            (self.cfg.d_head as f64).sqrt(),
            self.cfg.block_size,
            self.cfg.n_heads_kv,
            self.cfg.n_heads_q,
            self.cfg.causal,
        );
        let [b, _, s, _] = out_4d.dims();
        let out_3d =
            out_4d
                .swap_dims(1, 2)
                .reshape::<3, _>([b, s, self.cfg.n_heads_q * self.cfg.d_head]);
        (self.out_proj.forward(out_3d), block_attn)
    }

    pub fn forward_dense(&self, q: Tensor<B, 3>, k: Tensor<B, 3>, v: Tensor<B, 3>) -> Tensor<B, 3> {
        let cfg = &self.cfg;
        let d_head = cfg.d_head;
        let n_heads_q = cfg.n_heads_q;
        let n_heads_kv = cfg.n_heads_kv;
        let scale = (d_head as f64).sqrt();
        let q_per_kv = n_heads_q / n_heads_kv;

        let q_4d = Self::to_4d(q, n_heads_q, d_head);
        let k_4d = Self::to_4d(k, n_heads_kv, d_head);
        let v_4d = Self::to_4d(v, n_heads_kv, d_head);

        let [_, _, seq_q, _] = q_4d.dims();
        let seq_kv = k_4d.dims()[2];

        if q_per_kv > 1 {
            let k_exp = k_4d.repeat(&[1, q_per_kv, 1, 1]);
            let s = q_4d.matmul(k_exp.swap_dims(2, 3)).div_scalar(scale);
            let v_exp = v_4d.repeat(&[1, q_per_kv, 1, 1]);
            let scores = causal_mask_4d::<B>(s, seq_q, seq_kv, cfg.causal);
            let attn = softmax(scores, 3);
            let out_4d = attn.matmul(v_exp);
            let [b, _, s, _] = out_4d.dims();
            let out_3d = out_4d
                .swap_dims(1, 2)
                .reshape::<3, _>([b, s, n_heads_q * d_head]);
            self.out_proj.forward(out_3d)
        } else {
            let s = q_4d.matmul(k_4d.swap_dims(2, 3)).div_scalar(scale);
            let scores = causal_mask_4d::<B>(s, seq_q, seq_kv, cfg.causal);
            let attn = softmax(scores, 3);
            let out_4d = attn.matmul(v_4d);
            let [b, _, s, _] = out_4d.dims();
            let out_3d = out_4d
                .swap_dims(1, 2)
                .reshape::<3, _>([b, s, n_heads_q * d_head]);
            self.out_proj.forward(out_3d)
        }
    }
}

fn causal_mask_4d<B: Backend>(
    scores: Tensor<B, 4>,
    seq_q: usize,
    seq_kv: usize,
    causal: bool,
) -> Tensor<B, 4> {
    if !causal || seq_q <= 1 {
        return scores;
    }
    let device = scores.device();
    let q_pos =
        Tensor::<B, 1, Int>::arange(0..seq_q as i64, &device).reshape::<4, _>([1, 1, seq_q, 1]);
    let kv_pos =
        Tensor::<B, 1, Int>::arange(0..seq_kv as i64, &device).reshape::<4, _>([1, 1, 1, seq_kv]);
    scores.add(
        kv_pos
            .greater(q_pos)
            .float()
            .mul_scalar(crate::NEG_INF_SAFE),
    )
}

/// 5D batched GQA sparse attention - single gather pass.
///
/// Shapes:
/// - q: [batch, n_heads_q, seq_q, d_head]
/// - k, v: [batch, n_heads_kv, seq_kv, d_head]
/// - block_indices: [batch, n_heads_kv, seq_q, topk]
///
/// Returns (output [batch, n_heads_q, seq_q, d_head], block_attn [batch, n_heads_kv, seq_q, topk]).
#[allow(clippy::too_many_arguments)]
pub fn sparse_attn_batched_gqa<B: Backend>(
    q: Tensor<B, 4>,
    k: Tensor<B, 4>,
    v: Tensor<B, 4>,
    block_indices: Tensor<B, 4, Int>,
    scale: f64,
    block_size: usize,
    n_heads_kv: usize,
    n_heads_q: usize,
    causal: bool,
) -> (Tensor<B, 4>, Tensor<B, 4>) {
    let batch = q.dims()[0];
    let seq_q = q.dims()[2];
    let d_head = q.dims()[3];
    let seq_kv = k.dims()[2];
    let qpkv = n_heads_q / n_heads_kv;
    let topk = block_indices.dims()[3];
    let attended = topk * block_size;
    let device = q.device();

    let offsets = Tensor::<B, 1, Int>::arange(0..block_size as i64, &device)
        .reshape::<5, _>([1, 1, 1, 1, block_size]);

    let all_idx = block_indices.clone()
        .mul_scalar(block_size as i64)
        .unsqueeze_dim::<5>(4)
        .add(offsets)
        .reshape::<4, _>([batch, n_heads_kv, seq_q, attended]);

    // Masked full-attention: scores [B, kv, qpkv, S, S] are cheap (16MB at
    // b=8, kv=2, qpkv=4, S=256), the attended-token mask zeros everything
    // outside the selected blocks, and the output reuses the full V directly.
    // This avoids materializing 5D [B, kv, S, attended, d_head] gather tensors
    // (134MB each in the autodiff graph per iteration) — ~8× less memory than
    // gather-based sparse attention.
    // Attended-token mask: mark the selected *blocks* (always valid indices
    // into n_blocks), then expand to tokens. Block-level scatter avoids the
    // OOB/duplicate problem of token-level indices when seq_kv is not a
    // multiple of block_size.
    let n_blocks = seq_kv.div_ceil(block_size);
    let block_mask = Tensor::<B, 4, Int>::zeros([batch, n_heads_kv, seq_q, n_blocks], &device)
        .scatter(
            3,
            block_indices,
            Tensor::<B, 4, Int>::ones([batch, n_heads_kv, seq_q, topk], &device),
            burn::tensor::IndexingUpdateOp::Add,
        ); // [B, kv, S, n_blocks] — 1 at selected blocks
    let block_of_token = Tensor::<B, 1, Int>::arange(0..seq_kv as i64, &device)
        .div_scalar(block_size as i64)
        .reshape::<3, _>([1, 1, seq_kv])
        .expand([batch, n_heads_kv, seq_q, seq_kv]); // [B, kv, S, seq]
    let token_mask = block_mask.gather(3, block_of_token); // 1 iff token's block selected
    let mask = token_mask
        .unsqueeze_dim::<5>(2)
        .expand([batch, n_heads_kv, qpkv, seq_q, seq_kv]);

    // All tensors in [B, kv, qpkv, S, seq] layout.
    let q_gqa = q.reshape::<5, _>([batch, n_heads_kv, qpkv, seq_q, d_head]);
    let k_t = k
        .reshape::<5, _>([batch, n_heads_kv, 1, seq_kv, d_head])
        .swap_dims(3, 4);

    let mut scores = q_gqa.matmul(k_t).div_scalar(scale); // [B, kv, qpkv, S, S]

    if causal {
        let q_pos =
            Tensor::<B, 1, Int>::arange(0..seq_q as i64, &device).reshape::<4, _>([1, 1, seq_q, 1]);
        scores = scores.add(
            all_idx
                .clone()
                .unsqueeze_dim::<5>(2)
                .greater(q_pos.unsqueeze_dim::<5>(2))
                .float()
                .mul_scalar(crate::NEG_INF_SAFE),
        );
    }

    // Mask non-attended blocks: scores × attended + (1-attended)·(-inf).
    let m = mask.float();
    scores = scores.mul(m.clone()).add(m.neg().add_scalar(1.0).mul_scalar(crate::NEG_INF_SAFE));

    let attn = softmax::<B, 5>(scores, 4); // [B, kv, qpkv, S, S]

    let v_rep = v
        .reshape::<5, _>([batch, n_heads_kv, 1, seq_kv, d_head])
        .repeat(&[1, 1, qpkv, 1, 1]);
    let out_5d = attn
        .clone()
        .matmul(v_rep) // [B, kv, qpkv, S, d]
        .reshape::<4, _>([batch, n_heads_q, seq_q, d_head]);

    // Block attention for the KL-alignment loss: per selected block, the sum
    // of attention over its tokens. Gather only the selected slots (small
    // [B, kv, qpkv, S, attended] tensor) — avoids reshaping over seq_kv which
    // may not be a multiple of block_size.
    let attn_sel = attn.gather(
        4,
        all_idx
            .clone()
            .clamp_max((seq_kv - 1) as i64)
            .unsqueeze_dim::<5>(2)
            .expand([batch, n_heads_kv, qpkv, seq_q, attended]),
    ); // [B, kv, qpkv, S, topk*block]
    let ba = attn_sel
        .reshape::<6, _>([batch, n_heads_kv, qpkv, seq_q, topk, block_size])
        .sum_dim(5)
        .squeeze_dim::<5>(5)
        .mean_dim(2)
        .squeeze_dim::<4>(2); // [B, kv, S, topk]

    (out_5d, ba)
}

#[inline]
pub fn softmax<B: Backend, const D: usize>(t: Tensor<B, D>, dim: usize) -> Tensor<B, D> {
    activation::softmax(t, dim)
}

#[cfg(test)]
mod tests {
    use super::*;
    use burn::tensor::{Distribution, TensorData};
    use burn_ndarray::{NdArray, NdArrayDevice};

    fn device() -> NdArrayDevice {
        NdArrayDevice::default()
    }

    fn extract_f32(t: Tensor<NdArray, 4>) -> Vec<f32> {
        t.into_data()
            .bytes
            .chunks_exact(4)
            .map(|b| f32::from_le_bytes(b.try_into().unwrap()))
            .collect()
    }

    #[allow(clippy::too_many_arguments)]
    fn sparse_attn_per_token<B: Backend>(
        q: Tensor<B, 4>,
        k: Tensor<B, 4>,
        v: Tensor<B, 4>,
        block_indices: Tensor<B, 4, Int>,
        scale: f64,
        block_size: usize,
        n_heads_kv: usize,
        n_heads_q: usize,
    ) -> (Tensor<B, 4>, Tensor<B, 4>) {
        let batch = q.dims()[0];
        let seq_q = q.dims()[2];
        let d_head = q.dims()[3];
        let seq_kv = k.dims()[2];
        let q_per_kv = n_heads_q / n_heads_kv;
        let topk = block_indices.dims()[3];
        let attended = topk * block_size;

        let offsets_5d = Tensor::<B, 1, Int>::arange(0..block_size as i64, &q.device())
            .reshape::<5, _>([1, 1, 1, 1, block_size]);

        let mut out_rows = Vec::with_capacity(n_heads_kv);
        let mut aw_rows = Vec::with_capacity(n_heads_kv);

        for h_kv in 0..n_heads_kv {
            let k_h = k.clone().slice([0..batch, h_kv..h_kv + 1]);
            let v_h = v.clone().slice([0..batch, h_kv..h_kv + 1]);
            let mut attn_outputs = Vec::with_capacity(seq_q);
            let mut attn_weights = Vec::with_capacity(seq_q);

            for t in 0..seq_q {
                let blk = block_indices
                    .clone()
                    .slice([0..batch, h_kv..h_kv + 1, t..t + 1]);
                let blk_scaled = blk.mul_scalar(block_size as i64);
                let blk_5d = blk_scaled
                    .unsqueeze_dim::<5>(4)
                    .repeat(&[1, 1, 1, 1, block_size]);
                let idx = blk_5d
                    .add(offsets_5d.clone())
                    .reshape::<4, _>([batch, 1, attended, 1])
                    .repeat(&[1, 1, 1, d_head])
                    .clamp_max((seq_kv - 1) as i64);

                let k_sel = k_h.clone().gather(2, idx.clone());
                let v_sel = v_h.clone().gather(2, idx);

                let q_t = q.clone().slice([
                    0..batch,
                    (h_kv * q_per_kv)..((h_kv + 1) * q_per_kv),
                    t..t + 1,
                ]);

                let scores = q_t.matmul(k_sel.swap_dims(2, 3)).div_scalar(scale);
                let attn_n = softmax_4d(scores, 3);

                let aw = attn_n
                    .clone()
                    .reshape::<5, _>([batch, q_per_kv, 1, topk, block_size])
                    .sum_dim(4)
                    .squeeze_dim::<4>(4);

                attn_outputs.push(attn_n.matmul(v_sel));
                attn_weights.push(aw);
            }
            out_rows.push(Tensor::cat(attn_outputs, 2));
            let aw_seq = Tensor::cat(attn_weights, 2);
            aw_rows.push(aw_seq.mean_dim(1));
        }

        let block_attn = Tensor::cat(aw_rows, 1);
        (Tensor::cat(out_rows, 1), block_attn)
    }

    fn softmax_4d<B: Backend>(t: Tensor<B, 4>, dim: usize) -> Tensor<B, 4> {
        let mx = t.clone().max_dim(dim);
        let shifted = t.sub(mx);
        let ex = shifted.exp();
        let sum = ex.clone().sum_dim(dim);
        ex.div(sum)
    }

    #[test]
    fn batched_matches_per_token() {
        let d_head = 8;

        for &seq_q in [1usize, 2, 4, 7].iter() {
            for &seq_kv in [4usize, 8, 12, 16].iter() {// multiples of block_size=4
                for &n_heads_kv in [1usize, 2].iter() {
                    for &topk in [1usize, 3].iter() {
                        let n_heads_q = n_heads_kv * 2;
                        let block_size = 4;
                        let scale = (d_head as f64).sqrt();
                        let batch = 1;

                        let q = Tensor::<NdArray, 4>::random(
                            [batch, n_heads_q, seq_q, d_head],
                            Distribution::Normal(0.0, 1.0),
                            &device(),
                        );
                        let k = Tensor::<NdArray, 4>::random(
                            [batch, n_heads_kv, seq_kv, d_head],
                            Distribution::Normal(0.0, 1.0),
                            &device(),
                        );
                        let v = Tensor::<NdArray, 4>::random(
                            [batch, n_heads_kv, seq_kv, d_head],
                            Distribution::Normal(0.0, 1.0),
                            &device(),
                        );

                        let n_blocks = seq_kv.div_ceil(block_size);
                        let topk_actual = topk.min(n_blocks);
                        let per_group: Vec<i64> = (0..seq_q as i64)
                            .flat_map(|_| (0..topk_actual as i64).collect::<Vec<_>>())
                            .collect();
                        let idx_data: Vec<i64> = per_group.repeat(n_heads_kv);
                        let bi = Tensor::<NdArray, 4, Int>::from_data(
                            TensorData::new(idx_data, [batch, n_heads_kv, seq_q, topk_actual]),
                            &device(),
                        );

                        let (ref_out, _) = sparse_attn_per_token(
                            q.clone(),
                            k.clone(),
                            v.clone(),
                            bi.clone(),
                            scale,
                            block_size,
                            n_heads_kv,
                            n_heads_q,
                        );
                        let (batched_out, _) = sparse_attn_batched_gqa(
                            q, k, v, bi, scale, block_size, n_heads_kv, n_heads_q, false,
                        );

                        let ref_vals = extract_f32(ref_out);
                        let batched_vals = extract_f32(batched_out);

                        assert_eq!(ref_vals.len(), batched_vals.len());

                        for (i, (r, b)) in ref_vals.iter().zip(batched_vals.iter()).enumerate() {
                            if r.is_nan() && b.is_nan() {
                                continue;
                            }
                            if *r == 0.0 && *b == 0.0 {
                                continue;
                            }
                            let abs_diff = (r - b).abs();
                            let rel_diff = abs_diff / r.abs().max(b.abs()).max(f32::EPSILON);
                            assert!(
                                abs_diff <= 1e-5 || rel_diff <= 1e-4,
                                "Mismatch idx={} seq_q={} seq_kv={} kh={} tk={}: ref={} got={} abs={:.2e} rel={:.2e}",
                                i, seq_q, seq_kv, n_heads_kv, topk, r, b, abs_diff, rel_diff
                            );
                        }
                    }
                }
            }
        }
    }
}