boostr 0.2.0

ML framework built on numr - attention, quantization, model architectures
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

//! Generic paged-attention backward.
//!
//! Paged KV lives in a block pool with logical shape
//! `[num_blocks, block_size, num_kv_heads, head_dim]`; a `[B, max_num_blocks]`
//! block table maps each sequence's logical blocks to physical ones. The
//! backward gathers the per-sequence KV into a dense `[B, num_kv_heads, S_k, D]`
//! tensor, runs the shared standard-attention backward, then scatters the KV
//! gradients back into the block pool. Gather/scatter run on-device via numr
//! `index_select` / `scatter_reduce`; only the small `block_table` is read
//! host-side to build the row-index map.

use super::flash_standard::{StandardAttentionClient, StandardAttnConfig, standard_attention_bwd};
use crate::error::{Error, Result};
use numr::dtype::DType;
use numr::ops::{IndexingOps, ScatterReduceOp};
use numr::runtime::Runtime;
use numr::tensor::Tensor;

/// The paged KV pool plus its block table.
pub struct PagedKv<'a, R: Runtime> {
    /// Key block pool `[num_blocks, block_size, num_kv_heads, head_dim]`.
    pub k_blocks: &'a Tensor<R>,
    /// Value block pool, same shape as `k_blocks`.
    pub v_blocks: &'a Tensor<R>,
    /// Block table `[B, max_num_blocks]` (i32).
    pub block_table: &'a Tensor<R>,
}

/// Geometry for paged attention backward.
#[derive(Debug, Clone, Copy)]
pub struct PagedAttnConfig {
    /// Number of query heads.
    pub num_heads: usize,
    /// Number of key/value heads.
    pub num_kv_heads: usize,
    /// Total key/value sequence length.
    pub seq_len_k: usize,
    /// Per-head dimension.
    pub head_dim: usize,
    /// Tokens per KV block.
    pub block_size: usize,
    /// Apply a causal mask.
    pub causal: bool,
}

/// Row index, into the KV pool flattened to `[num_rows, head_dim]`, for every
/// `(b, h_kv, t)` in `[B, num_kv_heads, S_k]` order. A row is one `head_dim`
/// vector: `row = (physical_block · block_size + offset) · num_kv_heads + h_kv`.
fn paged_kv_row_indices(
    block_table: &[i32],
    batch_size: usize,
    num_kv_heads: usize,
    seq_len_k: usize,
    block_size: usize,
    max_num_blocks: usize,
) -> Vec<i64> {
    let mut rows = Vec::with_capacity(batch_size * num_kv_heads * seq_len_k);
    for b in 0..batch_size {
        for h_kv in 0..num_kv_heads {
            for t in 0..seq_len_k {
                let logical_block = t / block_size;
                let offset = t % block_size;
                let physical_block = block_table[b * max_num_blocks + logical_block] as usize;
                let row = (physical_block * block_size + offset) * num_kv_heads + h_kv;
                rows.push(row as i64);
            }
        }
    }
    rows
}

/// Paged attention backward. Returns `(dq, dk_blocks, dv_blocks)` where the KV
/// gradients have the same block-pool shape as `k_blocks` / `v_blocks`.
pub fn paged_attention_bwd_impl<R, C>(
    client: &C,
    dout: &Tensor<R>,
    q: &Tensor<R>,
    kv: &PagedKv<R>,
    output: &Tensor<R>,
    cfg: PagedAttnConfig,
) -> Result<(Tensor<R>, Tensor<R>, Tensor<R>)>
where
    R: Runtime<DType = DType>,
    C: StandardAttentionClient<R> + IndexingOps<R>,
{
    let batch_size = q.shape()[0];
    let max_num_blocks = kv.block_table.shape()[1];
    let d = cfg.head_dim;
    let nkv = cfg.num_kv_heads;
    let sk = cfg.seq_len_k;
    let device = q.device();

    let pool_rows: usize = kv.k_blocks.shape().iter().product::<usize>() / d;
    let n = batch_size * nkv * sk;

    // Build the gather/scatter row-index map from the (small) block table.
    let bt = kv.block_table.to_vec::<i32>();
    let rows = paged_kv_row_indices(&bt, batch_size, nkv, sk, cfg.block_size, max_num_blocks);
    let idx = Tensor::<R>::from_slice(&rows, &[n], device);

    // Gather KV blocks → dense [B, num_kv_heads, S_k, D].
    let k_pool = kv.k_blocks.reshape(&[pool_rows, d]).map_err(Error::Numr)?;
    let v_pool = kv.v_blocks.reshape(&[pool_rows, d]).map_err(Error::Numr)?;
    let k_dense = client
        .index_select(&k_pool, 0, &idx)
        .map_err(Error::Numr)?
        .reshape(&[batch_size, nkv, sk, d])
        .map_err(Error::Numr)?;
    let v_dense = client
        .index_select(&v_pool, 0, &idx)
        .map_err(Error::Numr)?
        .reshape(&[batch_size, nkv, sk, d])
        .map_err(Error::Numr)?;

    // Shared standard-attention backward (handles GQA expand/reduce).
    let (dq, dk_dense, dv_dense) = standard_attention_bwd(
        client,
        dout,
        q,
        &k_dense,
        &v_dense,
        output,
        StandardAttnConfig {
            num_heads: cfg.num_heads,
            num_kv_heads: nkv,
            causal: cfg.causal,
            window_size: 0,
        },
    )?;

    // Scatter KV gradients back into the block pool (accumulate onto zeros).
    let dk_blocks = scatter_kv_grad(
        client,
        &dk_dense,
        &idx,
        pool_rows,
        n,
        d,
        kv.k_blocks.shape(),
    )?;
    let dv_blocks = scatter_kv_grad(
        client,
        &dv_dense,
        &idx,
        pool_rows,
        n,
        d,
        kv.v_blocks.shape(),
    )?;

    Ok((dq, dk_blocks, dv_blocks))
}

/// Scatter-add a dense KV gradient `[B, num_kv_heads, S_k, D]` back into a
/// zeroed block pool shaped `pool_shape`, using the same row map as the gather.
fn scatter_kv_grad<R, C>(
    client: &C,
    grad_dense: &Tensor<R>,
    idx: &Tensor<R>,
    pool_rows: usize,
    n: usize,
    d: usize,
    pool_shape: &[usize],
) -> Result<Tensor<R>>
where
    R: Runtime<DType = DType>,
    C: StandardAttentionClient<R> + IndexingOps<R>,
{
    let grad_rows = grad_dense.reshape(&[n, d]).map_err(Error::Numr)?;
    // scatter_reduce is element-wise: index.shape == src.shape. Broadcast each
    // row's pool index across the `d` columns.
    let index2d = idx
        .reshape(&[n, 1])
        .map_err(Error::Numr)?
        .broadcast_to(&[n, d])
        .map_err(Error::Numr)?
        .contiguous()?;
    let dst = Tensor::<R>::zeros(&[pool_rows, d], DType::F32, grad_dense.device());
    let scattered = client
        .scatter_reduce(&dst, 0, &index2d, &grad_rows, ScatterReduceOp::Sum, true)
        .map_err(Error::Numr)?;
    scattered.reshape(pool_shape).map_err(Error::Numr)
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::ops::impl_generic::attention::flash_standard::standard_attention_fwd;
    use crate::test_utils::cpu_setup;
    use numr::ops::IndexingOps;
    use numr::runtime::cpu::CpuRuntime;

    #[test]
    fn row_indices_match_layout() {
        // 2 batches, block_size 2, 2 kv heads, seq 3, max_num_blocks 2.
        // block_table: b0 -> [0,1], b1 -> [2,3]
        let bt = [0, 1, 2, 3];
        let rows = paged_kv_row_indices(&bt, 2, 2, 3, 2, 2);
        // For b0,h0: t0 ->(0*2+0)*2+0=0, t1->(0*2+1)*2+0=2, t2->(1*2+0)*2+0=4
        assert_eq!(&rows[0..3], &[0, 2, 4]);
        // b0,h1: +1 each -> 1,3,5
        assert_eq!(&rows[3..6], &[1, 3, 5]);
        // b1,h0: physical blocks 2,3 -> t0->(2*2+0)*2=8, t1->(2*2+1)*2=10, t2->(3*2+0)*2=12
        assert_eq!(&rows[6..9], &[8, 10, 12]);
    }

    #[test]
    fn paged_bwd_matches_dense_reference() {
        let (client, device) = cpu_setup();
        let (b, h, nkv, d) = (2usize, 2usize, 1usize, 4usize);
        let block_size = 2usize;
        let seq_len_k = 4usize;
        let seq_len_q = 4usize;
        let max_num_blocks = 2usize;
        let num_blocks = 4usize; // b0 -> blocks 0,1 ; b1 -> blocks 2,3

        let nrows = num_blocks * block_size * nkv;
        let pool: Vec<f32> = (0..nrows * d).map(|i| (i as f32 * 0.017).sin()).collect();
        let k_blocks =
            Tensor::<CpuRuntime>::from_slice(&pool, &[num_blocks, block_size, nkv, d], &device);
        let vpool: Vec<f32> = (0..nrows * d).map(|i| (i as f32 * 0.023).cos()).collect();
        let v_blocks =
            Tensor::<CpuRuntime>::from_slice(&vpool, &[num_blocks, block_size, nkv, d], &device);
        let bt = Tensor::<CpuRuntime>::from_slice(&[0i32, 1, 2, 3], &[b, max_num_blocks], &device);

        let qd: Vec<f32> = (0..b * h * seq_len_q * d)
            .map(|i| (i as f32 * 0.01).cos())
            .collect();
        let q = Tensor::<CpuRuntime>::from_slice(&qd, &[b, h, seq_len_q, d], &device);
        let dod: Vec<f32> = (0..b * h * seq_len_q * d)
            .map(|i| (i as f32 * 0.03).sin())
            .collect();
        let dout = Tensor::<CpuRuntime>::from_slice(&dod, &[b, h, seq_len_q, d], &device);

        let cfg = PagedAttnConfig {
            num_heads: h,
            num_kv_heads: nkv,
            seq_len_k,
            head_dim: d,
            block_size,
            causal: true,
        };

        // Reference: gather KV densely with the same index map, run standard fwd/bwd.
        let rows =
            paged_kv_row_indices(&[0, 1, 2, 3], b, nkv, seq_len_k, block_size, max_num_blocks);
        let n = rows.len();
        let idx = Tensor::<CpuRuntime>::from_slice(&rows, &[n], &device);
        let k_pool = k_blocks.reshape(&[nrows, d]).unwrap();
        let v_pool = v_blocks.reshape(&[nrows, d]).unwrap();
        let k_dense = client
            .index_select(&k_pool, 0, &idx)
            .unwrap()
            .reshape(&[b, nkv, seq_len_k, d])
            .unwrap();
        let v_dense = client
            .index_select(&v_pool, 0, &idx)
            .unwrap()
            .reshape(&[b, nkv, seq_len_k, d])
            .unwrap();
        let scfg = StandardAttnConfig {
            num_heads: h,
            num_kv_heads: nkv,
            causal: true,
            window_size: 0,
        };
        let (output, _lse) = standard_attention_fwd(&client, &q, &k_dense, &v_dense, scfg).unwrap();
        let (dq_ref, dk_ref_dense, _dv_ref_dense) =
            standard_attention_bwd(&client, &dout, &q, &k_dense, &v_dense, &output, scfg).unwrap();

        // Paged backward under test.
        let kv = PagedKv {
            k_blocks: &k_blocks,
            v_blocks: &v_blocks,
            block_table: &bt,
        };
        let (dq, dk_blocks, _dv_blocks) =
            paged_attention_bwd_impl(&client, &dout, &q, &kv, &output, cfg).unwrap();

        // dq must match the dense reference exactly.
        let dq_a: Vec<f32> = dq.to_vec();
        let dq_b: Vec<f32> = dq_ref.to_vec();
        for (x, y) in dq_a.iter().zip(dq_b.iter()) {
            assert!((x - y).abs() < 1e-5, "dq mismatch: {x} vs {y}");
        }

        // Gathering dk_blocks back with the same index map must reproduce the
        // dense dk reference (the scatter is the gather's inverse here, since the
        // block table maps each token to a distinct physical slot).
        let dk_pool = dk_blocks.reshape(&[nrows, d]).unwrap();
        let dk_regathered = client.index_select(&dk_pool, 0, &idx).unwrap();
        let dk_re: Vec<f32> = dk_regathered.to_vec();
        let dk_ref: Vec<f32> = dk_ref_dense.reshape(&[n, d]).unwrap().to_vec();
        for (x, y) in dk_re.iter().zip(dk_ref.iter()) {
            assert!((x - y).abs() < 1e-5, "dk mismatch: {x} vs {y}");
        }
    }
}