boostr 0.1.0

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

//! AttentionBlock and SsmBlock sub-modules for the hybrid model.

use crate::error::{Error, Result};
use crate::inference::{KvCache, SsmState};
use crate::model::mamba::mamba2::Mamba2;
use crate::model::traits::ModelClient;
use crate::nn::{Linear, RmsNorm, RoPE};
use crate::ops::impl_generic::attention::multi_head_attention_impl;
use crate::ops::impl_generic::attention::rope::apply_rope_impl;
use numr::autograd::{Var, var_add, var_mul, var_narrow, var_reshape, var_silu};
use numr::dtype::DType;
use numr::ops::{
    ActivationOps, BinaryOps, CompareOps, ConditionalOps, ConvOps, IndexingOps, NormalizationOps,
    ReduceOps, ScalarOps, ShapeOps, TensorOps, UnaryOps,
};
use numr::runtime::Runtime;

/// Attention block: pre-norm → multi-head attention → residual + pre-norm → MLP → residual
pub(super) struct AttentionBlock<R: Runtime> {
    pub(super) input_layernorm: RmsNorm<R>,
    pub(super) q_proj: Linear<R>,
    pub(super) k_proj: Linear<R>,
    pub(super) v_proj: Linear<R>,
    pub(super) o_proj: Linear<R>,
    pub(super) post_attention_layernorm: RmsNorm<R>,
    pub(super) gate_proj: Linear<R>,
    pub(super) up_proj: Linear<R>,
    pub(super) down_proj: Linear<R>,
    pub(super) num_heads: usize,
    pub(super) num_kv_heads: usize,
    pub(super) head_dim: usize,
}

/// SSM block: pre-norm → Mamba2 → residual
pub(super) struct SsmBlock<R: Runtime> {
    pub(super) norm: RmsNorm<R>,
    pub(super) mamba: Mamba2<R>,
}

// ── AttentionBlock forward ──────────────────────────────────────────

impl<R: Runtime<DType = DType>> AttentionBlock<R> {
    pub(super) fn forward_with_kv_cache<C>(
        &self,
        client: &C,
        x: &Var<R>,
        rope: &RoPE<R>,
        kv_cache: &mut KvCache<R>,
        position: usize,
    ) -> Result<Var<R>>
    where
        C: ModelClient<R>,
        R::Client: TensorOps<R>
            + ScalarOps<R>
            + ReduceOps<R>
            + IndexingOps<R>
            + ShapeOps<R>
            + ActivationOps<R>
            + BinaryOps<R>
            + UnaryOps<R>
            + CompareOps<R>
            + ConditionalOps<R>,
    {
        // Pre-norm attention + residual
        let normed = self.input_layernorm.forward(client, x)?;
        let attn_out = self.attention_forward(client, &normed, rope, kv_cache, position)?;
        let h = var_add(x, &attn_out, client).map_err(Error::Numr)?;

        // Pre-norm MLP + residual
        let normed = self.post_attention_layernorm.forward(client, &h)?;
        let mlp_out = self.mlp_forward(client, &normed)?;
        var_add(&h, &mlp_out, client).map_err(Error::Numr)
    }

    fn attention_forward<C>(
        &self,
        client: &C,
        x: &Var<R>,
        rope: &RoPE<R>,
        kv_cache: &mut KvCache<R>,
        position: usize,
    ) -> Result<Var<R>>
    where
        C: ModelClient<R>,
        R::Client: TensorOps<R>
            + ScalarOps<R>
            + ReduceOps<R>
            + IndexingOps<R>
            + ShapeOps<R>
            + ActivationOps<R>
            + BinaryOps<R>
            + UnaryOps<R>
            + CompareOps<R>
            + ConditionalOps<R>,
    {
        let shape = x.shape().to_vec();
        let batch = shape[0];
        let seq_len = shape[1];

        let q = self.q_proj.forward(client, x)?;
        let k = self.k_proj.forward(client, x)?;
        let v = self.v_proj.forward(client, x)?;

        let q = var_reshape(&q, &[batch, seq_len, self.num_heads, self.head_dim])
            .map_err(Error::Numr)?;
        let k = var_reshape(&k, &[batch, seq_len, self.num_kv_heads, self.head_dim])
            .map_err(Error::Numr)?;
        let v = var_reshape(&v, &[batch, seq_len, self.num_kv_heads, self.head_dim])
            .map_err(Error::Numr)?;

        let q = numr::autograd::var_permute(&q, &[0, 2, 1, 3]).map_err(Error::Numr)?;
        let q = var_contiguous(&q);
        let k = numr::autograd::var_permute(&k, &[0, 2, 1, 3]).map_err(Error::Numr)?;
        let k = var_contiguous(&k);
        let v = numr::autograd::var_permute(&v, &[0, 2, 1, 3]).map_err(Error::Numr)?;
        let v = var_contiguous(&v);

        // Apply RoPE with position offset
        let cos_offset = var_narrow(rope.cos_cache(), 0, position, seq_len).map_err(Error::Numr)?;
        let sin_offset = var_narrow(rope.sin_cache(), 0, position, seq_len).map_err(Error::Numr)?;

        let q = apply_rope_impl(client, &q, &cos_offset, &sin_offset)?;
        let k = apply_rope_impl(client, &k, &cos_offset, &sin_offset)?;

        // Update KV cache with new K/V tensors [B, H_kv, S, D]
        kv_cache.update(k.tensor(), v.tensor())?;

        // Get full cached K/V for attention
        let (cached_k, cached_v) = kv_cache.get_kv()?;
        let cached_k = Var::new(cached_k.contiguous(), false);
        let cached_v = Var::new(cached_v.contiguous(), false);

        // GQA: repeat K/V heads to match Q heads if needed
        let (cached_k, cached_v) = if self.num_kv_heads < self.num_heads {
            let repeat = self.num_heads / self.num_kv_heads;
            let k_rep = repeat_kv(&cached_k, repeat).map_err(Error::Numr)?;
            let v_rep = repeat_kv(&cached_v, repeat).map_err(Error::Numr)?;
            (k_rep, v_rep)
        } else {
            (cached_k, cached_v)
        };

        // Multi-head attention (Q attends to full cached K/V)
        let attn_out =
            multi_head_attention_impl(client, &q, &cached_k, &cached_v, None, self.num_heads)?;

        // [B, H, S, D] -> [B, S, H, D] -> [B, S, H*D]
        let attn_out =
            numr::autograd::var_permute(&attn_out, &[0, 2, 1, 3]).map_err(Error::Numr)?;
        let attn_out = var_contiguous(&attn_out);
        let attn_out = var_reshape(&attn_out, &[batch, seq_len, self.num_heads * self.head_dim])
            .map_err(Error::Numr)?;

        // Output projection
        self.o_proj.forward(client, &attn_out)
    }

    fn mlp_forward<C>(&self, client: &C, x: &Var<R>) -> Result<Var<R>>
    where
        C: ModelClient<R>,
        R::Client: TensorOps<R>
            + ScalarOps<R>
            + ReduceOps<R>
            + IndexingOps<R>
            + ShapeOps<R>
            + ActivationOps<R>
            + BinaryOps<R>
            + UnaryOps<R>
            + CompareOps<R>
            + ConditionalOps<R>,
    {
        let gate = self.gate_proj.forward(client, x)?;
        let up = self.up_proj.forward(client, x)?;
        let gate_silu = var_silu(&gate, client).map_err(Error::Numr)?;
        let hidden = var_mul(&gate_silu, &up, client).map_err(Error::Numr)?;
        self.down_proj.forward(client, &hidden)
    }
}

// ── SsmBlock forward ────────────────────────────────────────────────

impl<R: Runtime<DType = DType>> SsmBlock<R> {
    pub(super) fn forward_inference<C>(
        &self,
        client: &C,
        x: &Var<R>,
        state: &mut SsmState<R>,
    ) -> Result<Var<R>>
    where
        C: ModelClient<R> + ConvOps<R> + NormalizationOps<R> + UnaryOps<R> + ActivationOps<R>,
        R::Client: TensorOps<R>
            + ScalarOps<R>
            + ActivationOps<R>
            + ConvOps<R>
            + ReduceOps<R>
            + BinaryOps<R>
            + IndexingOps<R>,
    {
        let normed = self.norm.forward(client, x)?;
        let out_tensor = self
            .mamba
            .forward_inference(client, normed.tensor(), state)?;
        let out = Var::new(out_tensor, false);
        numr::autograd::var_add(x, &out, client).map_err(Error::Numr)
    }
}

// ── Helpers ─────────────────────────────────────────────────────────

pub(super) fn var_contiguous<R: Runtime>(v: &Var<R>) -> Var<R> {
    Var::new(v.tensor().contiguous(), v.requires_grad())
}

pub(super) fn repeat_kv<R: Runtime>(x: &Var<R>, repeat: usize) -> numr::error::Result<Var<R>> {
    if repeat == 1 {
        return Ok(x.clone());
    }
    let shape = x.shape();
    let [b, h_kv, s, d] = [shape[0], shape[1], shape[2], shape[3]];

    // Reshape to [B, H_kv, 1, S, D] then broadcast to [B, H_kv, repeat, S, D]
    let expanded = x.tensor().reshape(&[b, h_kv, 1, s, d])?;
    let expanded = expanded.broadcast_to(&[b, h_kv, repeat, s, d])?;
    // Reshape to [B, H_kv * repeat, S, D] — need contiguous for reshape after broadcast
    let result = expanded.contiguous().reshape(&[b, h_kv * repeat, s, d])?;
    Ok(Var::new(result, x.requires_grad()))
}