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

//! CPU fallback for variable-length (ragged) attention
//!
//! Unpacks sequences from the packed buffer, runs standard attention per-sequence,
//! and writes results back to the packed output buffer.

use crate::error::Result;
use crate::ops::traits::VarLenAttentionOps;
use numr::runtime::cpu::{CpuClient, CpuRuntime};
use numr::tensor::Tensor;

impl VarLenAttentionOps<CpuRuntime> for CpuClient {
    fn varlen_attention_fwd(
        &self,
        q: &Tensor<CpuRuntime>,
        k: &Tensor<CpuRuntime>,
        v: &Tensor<CpuRuntime>,
        cu_seqlens_q: &Tensor<CpuRuntime>,
        cu_seqlens_k: &Tensor<CpuRuntime>,
        batch_size: usize,
        num_heads: usize,
        num_kv_heads: usize,
        _max_seqlen_q: usize,
        _max_seqlen_k: usize,
        head_dim: usize,
        causal: bool,
    ) -> Result<(Tensor<CpuRuntime>, Tensor<CpuRuntime>)> {
        // GQA ratio: each kv head serves (num_heads / num_kv_heads) query heads.
        // For MHA num_kv_heads == num_heads so the ratio is 1 and the mapping is
        // the identity: kv_h = q_h / 1 = q_h.
        let gqa_ratio = num_heads / num_kv_heads;

        let total_tokens_q = q.shape()[0];
        let device = q.device();

        let q_data = q.to_vec::<f32>();
        let k_data = k.to_vec::<f32>();
        let v_data = v.to_vec::<f32>();
        let cu_q = cu_seqlens_q.to_vec::<i32>();
        let cu_k = cu_seqlens_k.to_vec::<i32>();

        let mut out = vec![0.0f32; total_tokens_q * num_heads * head_dim];
        let mut lse = vec![0.0f32; total_tokens_q * num_heads];

        let scale = (head_dim as f32).sqrt().recip();

        for b in 0..batch_size {
            let sq_start = cu_q[b] as usize;
            let sq_end = cu_q[b + 1] as usize;
            let sk_start = cu_k[b] as usize;
            let sk_end = cu_k[b + 1] as usize;
            let seq_len_q = sq_end - sq_start;
            let seq_len_k = sk_end - sk_start;

            for h in 0..num_heads {
                // GQA: map query head h to the corresponding kv head.
                let kv_h = h / gqa_ratio;

                for qi in 0..seq_len_q {
                    let q_offset = ((sq_start + qi) * num_heads + h) * head_dim;

                    let mut max_score = f32::NEG_INFINITY;
                    let mut scores = Vec::with_capacity(seq_len_k);

                    for ki in 0..seq_len_k {
                        if causal && qi < ki {
                            scores.push(f32::NEG_INFINITY);
                            continue;
                        }
                        // K/V row stride uses num_kv_heads (GQA layout).
                        let k_offset = ((sk_start + ki) * num_kv_heads + kv_h) * head_dim;
                        let mut dot = 0.0f32;
                        for d in 0..head_dim {
                            dot += q_data[q_offset + d] * k_data[k_offset + d];
                        }
                        let s = dot * scale;
                        max_score = max_score.max(s);
                        scores.push(s);
                    }

                    let mut sum_exp = 0.0f32;
                    let mut exp_scores = Vec::with_capacity(seq_len_k);
                    for &s in &scores {
                        let e = (s - max_score).exp();
                        sum_exp += e;
                        exp_scores.push(e);
                    }

                    let o_offset = ((sq_start + qi) * num_heads + h) * head_dim;
                    for (ki, &exp_s) in exp_scores.iter().enumerate() {
                        let weight = exp_s / sum_exp;
                        // V row stride also uses num_kv_heads.
                        let v_offset = ((sk_start + ki) * num_kv_heads + kv_h) * head_dim;
                        for d in 0..head_dim {
                            out[o_offset + d] += weight * v_data[v_offset + d];
                        }
                    }

                    lse[(sq_start + qi) * num_heads + h] = max_score + sum_exp.ln();
                }
            }
        }

        let output =
            Tensor::<CpuRuntime>::from_slice(&out, &[total_tokens_q, num_heads, head_dim], device);
        let lse_tensor =
            Tensor::<CpuRuntime>::from_slice(&lse, &[total_tokens_q, num_heads], device);
        Ok((output, lse_tensor))
    }

    fn varlen_attention_bwd(
        &self,
        dout: &Tensor<CpuRuntime>,
        q: &Tensor<CpuRuntime>,
        k: &Tensor<CpuRuntime>,
        v: &Tensor<CpuRuntime>,
        output: &Tensor<CpuRuntime>,
        lse: &Tensor<CpuRuntime>,
        cu_seqlens_q: &Tensor<CpuRuntime>,
        cu_seqlens_k: &Tensor<CpuRuntime>,
        batch_size: usize,
        num_heads: usize,
        num_kv_heads: usize,
        _max_seqlen_q: usize,
        _max_seqlen_k: usize,
        head_dim: usize,
        causal: bool,
    ) -> Result<(Tensor<CpuRuntime>, Tensor<CpuRuntime>, Tensor<CpuRuntime>)> {
        // GQA backward: multiple Q heads share one KV head.
        // dK and dV accumulate contributions from all Q heads mapping to the same
        // KV head — the serial += across the h-loop is the scalar analog of atomicAdd.
        let gqa_ratio = num_heads / num_kv_heads;

        let total_tokens_q = q.shape()[0];
        let total_tokens_k = k.shape()[0];
        let device = q.device();

        let q_data = q.to_vec::<f32>();
        let k_data = k.to_vec::<f32>();
        let v_data = v.to_vec::<f32>();
        let o_data = output.to_vec::<f32>();
        let lse_data = lse.to_vec::<f32>();
        let do_data = dout.to_vec::<f32>();
        let cu_q = cu_seqlens_q.to_vec::<i32>();
        let cu_k = cu_seqlens_k.to_vec::<i32>();

        let mut dq = vec![0.0f32; total_tokens_q * num_heads * head_dim];
        let mut dk = vec![0.0f32; total_tokens_k * num_kv_heads * head_dim];
        let mut dv = vec![0.0f32; total_tokens_k * num_kv_heads * head_dim];

        let scale = (head_dim as f32).sqrt().recip();

        for b in 0..batch_size {
            let sq_start = cu_q[b] as usize;
            let sq_end = cu_q[b + 1] as usize;
            let sk_start = cu_k[b] as usize;
            let sk_end = cu_k[b + 1] as usize;
            let seq_len_q = sq_end - sq_start;
            let seq_len_k = sk_end - sk_start;

            for h in 0..num_heads {
                // GQA: map query head h to the corresponding kv head.
                let kv_h = h / gqa_ratio;

                for qi in 0..seq_len_q {
                    let q_off = ((sq_start + qi) * num_heads + h) * head_dim;
                    let l = lse_data[(sq_start + qi) * num_heads + h];

                    // D = sum(dO * O)
                    let o_off = ((sq_start + qi) * num_heads + h) * head_dim;
                    let mut d_val = 0.0f32;
                    for d in 0..head_dim {
                        d_val += do_data[o_off + d] * o_data[o_off + d];
                    }

                    for ki in 0..seq_len_k {
                        if causal && qi < ki {
                            continue;
                        }
                        // K/V row stride uses num_kv_heads (GQA layout)
                        let k_off = ((sk_start + ki) * num_kv_heads + kv_h) * head_dim;

                        // Recompute score and prob
                        let mut score = 0.0f32;
                        for d in 0..head_dim {
                            score += q_data[q_off + d] * k_data[k_off + d];
                        }
                        score *= scale;
                        let prob = (score - l).exp();

                        // grad_prob = V @ dO
                        let v_off = ((sk_start + ki) * num_kv_heads + kv_h) * head_dim;
                        let mut grad_prob = 0.0f32;
                        for d in 0..head_dim {
                            grad_prob += v_data[v_off + d] * do_data[o_off + d];
                        }

                        let grad_score = prob * (grad_prob - d_val);

                        // Accumulate gradients — dk/dv indexed by kv_h (GQA scatter)
                        for d in 0..head_dim {
                            dq[q_off + d] += scale * grad_score * k_data[k_off + d];
                            dk[k_off + d] += scale * grad_score * q_data[q_off + d];
                            dv[v_off + d] += prob * do_data[o_off + d];
                        }
                    }
                }
            }
        }

        let dq_t =
            Tensor::<CpuRuntime>::from_slice(&dq, &[total_tokens_q, num_heads, head_dim], device);
        let dk_t = Tensor::<CpuRuntime>::from_slice(
            &dk,
            &[total_tokens_k, num_kv_heads, head_dim],
            device,
        );
        let dv_t = Tensor::<CpuRuntime>::from_slice(
            &dv,
            &[total_tokens_k, num_kv_heads, head_dim],
            device,
        );
        Ok((dq_t, dk_t, dv_t))
    }
}

#[cfg(test)]
#[path = "varlen_attention_tests.rs"]
mod tests;