laya 0.1.0

Rust inference for the Laya non-autoregressive typed-decision model (ModernBERT-large + RL decision head)
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

//! The decision model: a frozen-architecture ModernBERT backbone plus the from-scratch
//! decision head (`DecisionModel` in `rl_common.py`).

use anyhow::Result;
use candle_core::{DType, Device, IndexOp, Tensor, D};
use candle_nn::ops::softmax;
use candle_nn::{embedding, layer_norm, linear, Embedding, LayerNorm, Linear, Module, VarBuilder};
use candle_transformers::models::modernbert::ModernBert;

use crate::config::EncoderConfig;

/// PyTorch's `nn.LayerNorm` default.
const LN_EPS: f64 = 1e-5;
/// `nn.MultiheadAttention` head size implied by `nhead = d // 64` in the reference.
const HEAD_DIM: usize = 64;
/// The reference masks unused option slots with this value before the softmax.
const MASK_FILL: f64 = -1e4;

/// One `nn.TransformerEncoderLayer(..., norm_first=True)`.
///
/// Note the activation: the reference constructs the layer positionally and never passes
/// `activation`, so it keeps PyTorch's default ReLU, not GELU.
struct HeadLayer {
    norm1: LayerNorm,
    norm2: LayerNorm,
    in_proj_w: Tensor,
    in_proj_b: Tensor,
    out_proj: Linear,
    linear1: Linear,
    linear2: Linear,
    n_heads: usize,
}

impl HeadLayer {
    fn load(vb: VarBuilder, d: usize, ff: usize) -> Result<Self> {
        let n_heads = std::cmp::max(1, d / HEAD_DIM);
        Ok(Self {
            norm1: layer_norm(d, LN_EPS, vb.pp("norm1"))?,
            norm2: layer_norm(d, LN_EPS, vb.pp("norm2"))?,
            in_proj_w: vb.get((3 * d, d), "self_attn.in_proj_weight")?,
            in_proj_b: vb.get(3 * d, "self_attn.in_proj_bias")?,
            out_proj: linear(d, d, vb.pp("self_attn.out_proj"))?,
            linear1: linear(d, ff, vb.pp("linear1"))?,
            linear2: linear(ff, d, vb.pp("linear2"))?,
            n_heads,
        })
    }

    /// `pad_bias` is additive, shaped (b, 1, 1, L): 0 for real tokens, -inf for padding.
    fn attn(&self, x: &Tensor, pad_bias: &Tensor) -> Result<Tensor> {
        let (b, l, d) = x.dims3()?;
        let hd = d / self.n_heads;

        let qkv = x
            .broadcast_matmul(&self.in_proj_w.t()?)?
            .broadcast_add(&self.in_proj_b)?;
        let qkv = qkv
            .reshape((b, l, 3, self.n_heads, hd))?
            .permute((2, 0, 3, 1, 4))?
            .contiguous()?;
        let q = qkv.i(0)?;
        let k = qkv.i(1)?;
        let v = qkv.i(2)?;

        let scale = (hd as f64).powf(-0.5);
        let att =
            (q.contiguous()? * scale)?.matmul(&k.transpose(D::Minus2, D::Minus1)?.contiguous()?)?;
        let att = att.broadcast_add(pad_bias)?;
        let att = softmax(&att, D::Minus1)?;

        let out = att.matmul(&v.contiguous()?)?;
        let out = out.transpose(1, 2)?.reshape((b, l, d))?;
        Ok(self.out_proj.forward(&out)?)
    }

    fn forward(&self, x: &Tensor, pad_bias: &Tensor) -> Result<Tensor> {
        let h = self.attn(&self.norm1.forward(x)?, pad_bias)?;
        let x = (x + h)?;
        let f = self
            .linear2
            .forward(&self.linear1.forward(&self.norm2.forward(&x)?)?.relu()?)?;
        Ok((x + f)?)
    }
}

/// `nn.Sequential(LayerNorm, Linear, GELU, Linear)` — indices 0, 1, 3 in the checkpoint.
struct Scorer {
    norm: LayerNorm,
    fc1: Linear,
    fc2: Linear,
}

impl Scorer {
    fn load(vb: VarBuilder, d: usize) -> Result<Self> {
        Ok(Self {
            norm: layer_norm(d, LN_EPS, vb.pp("0"))?,
            fc1: linear(d, d, vb.pp("1"))?,
            fc2: linear(d, 1, vb.pp("3"))?,
        })
    }

    fn forward(&self, x: &Tensor) -> Result<Tensor> {
        let x = self.fc1.forward(&self.norm.forward(x)?)?.gelu_erf()?;
        Ok(self.fc2.forward(&x)?)
    }
}

/// `nn.Sequential(Linear, GELU, Linear)` over the pooled state plus 4 distribution features.
struct ActHead {
    fc1: Linear,
    fc2: Linear,
}

impl ActHead {
    fn load(vb: VarBuilder, d: usize, n_act: usize) -> Result<Self> {
        Ok(Self {
            fc1: linear(d + 4, 256, vb.pp("0"))?,
            fc2: linear(256, n_act, vb.pp("2"))?,
        })
    }

    fn forward(&self, x: &Tensor) -> Result<Tensor> {
        Ok(self.fc2.forward(&self.fc1.forward(x)?.gelu_erf()?)?)
    }
}

pub struct DecisionModel {
    encoder: ModernBert,
    type_emb: Embedding,
    layers: Vec<HeadLayer>,
    scorer: Scorer,
    act_head: ActHead,
    pub device: Device,
    pub dtype: DType,
}

/// Raw model outputs for one batch: per-option logits (uncalibrated) and act-head probabilities.
pub struct Outputs {
    /// (batch, kmax) — masked slots hold `MASK_FILL`.
    pub logits: Vec<Vec<f32>>,
    /// (batch, n_act) after softmax. Index 0 is "answer", 1 is "escalate".
    pub act: Vec<Vec<f32>>,
}

impl DecisionModel {
    pub fn load(
        vb: VarBuilder,
        enc_cfg: &EncoderConfig,
        head_layers: usize,
        n_act: usize,
        device: Device,
        dtype: DType,
    ) -> Result<Self> {
        let d = enc_cfg.hidden_size;
        let encoder = ModernBert::load(vb.clone(), &enc_cfg.to_candle())?;
        let type_emb = embedding(3, d, vb.pp("type_emb"))?;
        let mut layers = Vec::with_capacity(head_layers);
        for i in 0..head_layers {
            layers.push(HeadLayer::load(
                vb.pp(format!("head.layers.{i}")),
                d,
                4 * d,
            )?);
        }
        Ok(Self {
            encoder,
            type_emb,
            layers,
            scorer: Scorer::load(vb.pp("scorer"), d)?,
            act_head: ActHead::load(vb.pp("act_head"), d, n_act)?,
            device,
            dtype,
        })
    }

    /// `input_ids`/`attention_mask`: (b, L). `marker_pos`/`marker_mask`: (b, kmax).
    pub fn forward(
        &self,
        input_ids: &Tensor,
        attention_mask: &Tensor,
        marker_pos: &Tensor,
        marker_mask: &[Vec<bool>],
        qtype: &Tensor,
    ) -> Result<Outputs> {
        let h = self.encoder.forward(input_ids, attention_mask)?;
        let (b, _l, d) = h.dims3()?;

        // Question-type embedding, broadcast over the sequence.
        let te = self.type_emb.forward(qtype)?.unsqueeze(1)?;
        let mut h = h.broadcast_add(&te.to_dtype(h.dtype())?)?;

        // The head attends over the full sequence, with padding masked out.
        // Large but finite: -inf here would make fully-padded rows produce NaN.
        let pad_bias = ((attention_mask.to_dtype(h.dtype())? - 1.0)? * 1e30)?
            .unsqueeze(1)?
            .unsqueeze(1)?;
        for layer in &self.layers {
            h = layer.forward(&h, &pad_bias)?;
        }

        // Gather the per-option marker positions.
        let kmax = marker_pos.dim(1)?;
        let idx = marker_pos
            .unsqueeze(2)?
            .expand((b, kmax, d))?
            .contiguous()?;
        let m = h.gather(&idx, 1)?;

        let logits = self
            .scorer
            .forward(&m)?
            .squeeze(D::Minus1)?
            .to_dtype(DType::F32)?;
        let mut logits: Vec<Vec<f32>> = logits.to_vec2()?;
        for (row, mask) in logits.iter_mut().zip(marker_mask) {
            for (v, keep) in row.iter_mut().zip(mask) {
                if !keep {
                    *v = MASK_FILL as f32;
                }
            }
        }

        // The act head sees the pooled sequence plus a summary of the answer distribution.
        let feats: Vec<f32> = logits
            .iter()
            .zip(marker_mask)
            .flat_map(|(row, mask)| {
                let p = stable_softmax(row);
                let k = mask.iter().filter(|m| **m).count().max(2) as f32;
                let ent: f32 = -p.iter().map(|x| x * x.max(1e-9).ln()).sum::<f32>() / k.ln();
                let mut sorted = p.clone();
                sorted.sort_by(|a, b| b.partial_cmp(a).unwrap());
                let (t1, t2) = (sorted[0], sorted.get(1).copied().unwrap_or(0.0));
                [t1, t1 - t2, ent, k / 255.0]
            })
            .collect();
        let feats = Tensor::from_vec(feats, (b, 4), &self.device)?.to_dtype(h.dtype())?;
        let pooled = h.i((.., 0, ..))?;
        let act = self
            .act_head
            .forward(&Tensor::cat(&[&pooled, &feats], 1)?)?;
        let act = softmax(&act.to_dtype(DType::F32)?, D::Minus1)?.to_vec2()?;

        Ok(Outputs { logits, act })
    }
}

/// Softmax in f32 with the max subtracted, matching the reference's numpy path.
pub fn stable_softmax(z: &[f32]) -> Vec<f32> {
    let max = z.iter().copied().fold(f32::NEG_INFINITY, f32::max);
    let e: Vec<f32> = z.iter().map(|v| (v - max).exp()).collect();
    let s: f32 = e.iter().sum();
    e.into_iter().map(|v| v / s).collect()
}