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
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

//! Whisper audio encoder: Conv1d stem + transformer encoder layers.
//!
//! Implements the encoder portion of OpenAI's Whisper model, which transforms
//! log-mel spectrograms into a sequence of hidden representations.

use crate::error::{Error, Result};
use crate::model::config::AudioConfig;
use crate::nn::VarBuilder;
use crate::nn::conv1d::Conv1d;
use crate::nn::layernorm::LayerNorm;
use crate::nn::linear::Linear;
use numr::autograd::Var;
use numr::ops::{
    ActivationOps, BinaryOps, ConvOps, MatmulOps, NormalizationOps, ReduceOps, ScalarOps, ShapeOps,
    TensorOps,
};
use numr::runtime::{Runtime, RuntimeClient};
use numr::tensor::Tensor;

/// A single Whisper encoder transformer layer.
///
/// Standard transformer encoder with bidirectional (non-causal) self-attention.
pub struct WhisperEncoderLayer<R: Runtime> {
    ln1: LayerNorm<R>,
    q_proj: Linear<R>,
    k_proj: Linear<R>,
    v_proj: Linear<R>,
    out_proj: Linear<R>,
    ln2: LayerNorm<R>,
    fc1: Linear<R>,
    fc2: Linear<R>,
    num_heads: usize,
    head_dim: usize,
}

impl<R: Runtime> WhisperEncoderLayer<R> {
    /// Load a single encoder layer from a VarBuilder scoped to this layer's prefix.
    pub fn from_varbuilder(
        vb: &mut VarBuilder<'_, R>,
        hidden_size: usize,
        num_heads: usize,
    ) -> Result<Self> {
        let head_dim = hidden_size / num_heads;

        let mut self_attn = vb.pp("self_attn");

        let q_w = self_attn.take_tensor("q_proj.weight")?;
        let q_b = self_attn.take_tensor_optional("q_proj.bias")?;
        let q_proj = Linear::new(q_w, q_b, false);

        let k_w = self_attn.take_tensor("k_proj.weight")?;
        let k_b = self_attn.take_tensor_optional("k_proj.bias")?;
        let k_proj = Linear::new(k_w, k_b, false);

        let v_w = self_attn.take_tensor("v_proj.weight")?;
        let v_b = self_attn.take_tensor_optional("v_proj.bias")?;
        let v_proj = Linear::new(v_w, v_b, false);

        let out_w = self_attn.take_tensor("out_proj.weight")?;
        let out_b = self_attn.take_tensor_optional("out_proj.bias")?;
        let out_proj = Linear::new(out_w, out_b, false);

        drop(self_attn);

        let mut ln1_vb = vb.pp("self_attn_layer_norm");
        let ln1_w = ln1_vb.take_tensor("weight")?;
        let ln1_b = ln1_vb.take_tensor("bias")?;
        let ln1 = LayerNorm::new(ln1_w, ln1_b, 1e-5, false);
        drop(ln1_vb);

        let mut fc_vb = vb.pp("fc1");
        let fc1_w = fc_vb.take_tensor("weight")?;
        let fc1_b = fc_vb.take_tensor_optional("bias")?;
        let fc1 = Linear::new(fc1_w, fc1_b, false);
        drop(fc_vb);

        let mut fc2_vb = vb.pp("fc2");
        let fc2_w = fc2_vb.take_tensor("weight")?;
        let fc2_b = fc2_vb.take_tensor_optional("bias")?;
        let fc2 = Linear::new(fc2_w, fc2_b, false);
        drop(fc2_vb);

        let mut ln2_vb = vb.pp("final_layer_norm");
        let ln2_w = ln2_vb.take_tensor("weight")?;
        let ln2_b = ln2_vb.take_tensor("bias")?;
        let ln2 = LayerNorm::new(ln2_w, ln2_b, 1e-5, false);

        Ok(Self {
            ln1,
            q_proj,
            k_proj,
            v_proj,
            out_proj,
            ln2,
            fc1,
            fc2,
            num_heads,
            head_dim,
        })
    }

    /// Forward pass (inference, no autograd).
    ///
    /// Input/output: `[B, seq_len, hidden]`
    /// Uses bidirectional attention (no causal mask).
    pub fn forward_inference<C>(&self, client: &C, x: &Tensor<R>) -> Result<Tensor<R>>
    where
        C: RuntimeClient<R>
            + TensorOps<R>
            + ScalarOps<R>
            + MatmulOps<R>
            + BinaryOps<R>
            + ActivationOps<R>
            + NormalizationOps<R>
            + ReduceOps<R>
            + ShapeOps<R>,
        R::Client: TensorOps<R> + ScalarOps<R>,
    {
        let shape = x.shape();
        let batch = shape[0];
        let seq_len = shape[1];
        let hidden = shape[2];

        // Pre-norm
        let x_var = Var::new(x.clone(), false);
        let normed = self.ln1.forward(client, &x_var)?;
        let normed_t = normed.tensor().clone();

        // Q, K, V projections
        let normed_var = Var::new(normed_t.clone(), false);
        let q = self.q_proj.forward(client, &normed_var)?;
        let k = self.k_proj.forward(client, &normed_var)?;
        let v = self.v_proj.forward(client, &normed_var)?;

        // Reshape to [B, num_heads, seq_len, head_dim]
        let q = q
            .tensor()
            .reshape(&[batch, seq_len, self.num_heads, self.head_dim])
            .map_err(Error::Numr)?
            .transpose(1, 2)
            .map_err(Error::Numr)?
            .contiguous();
        let k = k
            .tensor()
            .reshape(&[batch, seq_len, self.num_heads, self.head_dim])
            .map_err(Error::Numr)?
            .transpose(1, 2)
            .map_err(Error::Numr)?
            .contiguous();
        let v = v
            .tensor()
            .reshape(&[batch, seq_len, self.num_heads, self.head_dim])
            .map_err(Error::Numr)?
            .transpose(1, 2)
            .map_err(Error::Numr)?
            .contiguous();

        // Attention: softmax(Q @ K^T / sqrt(d)) @ V
        let k_t = k.transpose(2, 3).map_err(Error::Numr)?.contiguous();
        let scale = (self.head_dim as f32).sqrt();
        let scores = client.matmul(&q, &k_t).map_err(Error::Numr)?;
        let scores = client
            .div_scalar(&scores, scale as f64)
            .map_err(Error::Numr)?;
        // Bidirectional: no causal mask
        let attn_weights = client.softmax(&scores, -1).map_err(Error::Numr)?;
        let attn_out = client.matmul(&attn_weights, &v).map_err(Error::Numr)?;

        // Reshape back to [B, seq_len, hidden]
        let attn_out = attn_out
            .transpose(1, 2)
            .map_err(Error::Numr)?
            .contiguous()
            .reshape(&[batch, seq_len, hidden])
            .map_err(Error::Numr)?;

        // Project out
        let attn_out_var = Var::new(attn_out, false);
        let projected = self.out_proj.forward(client, &attn_out_var)?;

        // Residual
        let after_attn = client.add(x, projected.tensor()).map_err(Error::Numr)?;

        // FFN with pre-norm
        let after_attn_var = Var::new(after_attn.clone(), false);
        let normed2 = self.ln2.forward(client, &after_attn_var)?;
        let normed2_var = Var::new(normed2.tensor().clone(), false);

        let fc1_out = self.fc1.forward(client, &normed2_var)?;
        let activated = client.gelu(fc1_out.tensor()).map_err(Error::Numr)?;
        let activated_var = Var::new(activated, false);
        let fc2_out = self.fc2.forward(client, &activated_var)?;

        // Residual
        client
            .add(&after_attn, fc2_out.tensor())
            .map_err(Error::Numr)
    }
}

/// Whisper audio encoder.
///
/// Architecture:
/// 1. Two Conv1d layers (stem) to downsample mel spectrogram
/// 2. Learned positional embedding
/// 3. Stack of transformer encoder layers (bidirectional attention)
/// 4. Final layer norm
pub struct WhisperEncoder<R: Runtime> {
    conv1: Conv1d<R>,
    conv2: Conv1d<R>,
    position_embedding: Var<R>,
    layers: Vec<WhisperEncoderLayer<R>>,
    ln_post: LayerNorm<R>,
}

impl<R: Runtime> WhisperEncoder<R> {
    /// Load from a VarBuilder and AudioConfig.
    pub fn from_varbuilder(vb: &mut VarBuilder<'_, R>, config: &AudioConfig) -> Result<Self> {
        let hidden = config.hidden_size;
        // Conv stem
        let mut conv1_vb = vb.pp("conv1");
        let conv1_w = conv1_vb.take_tensor("weight")?;
        let conv1_b = conv1_vb.take_tensor_optional("bias")?;
        drop(conv1_vb);
        let conv1 = Conv1d::new(
            conv1_w,
            conv1_b,
            1,                                    // stride=1
            numr::ops::PaddingMode::conv1d(1, 1), // padding=1
            1,                                    // dilation=1
            1,                                    // groups=1
            false,
        );

        let mut conv2_vb = vb.pp("conv2");
        let conv2_w = conv2_vb.take_tensor("weight")?;
        let conv2_b = conv2_vb.take_tensor_optional("bias")?;
        drop(conv2_vb);
        let conv2 = Conv1d::new(
            conv2_w,
            conv2_b,
            2,                                    // stride=2 (downsample)
            numr::ops::PaddingMode::conv1d(1, 1), // padding=1
            1,                                    // dilation=1
            1,                                    // groups=1
            false,
        );

        // Learned positional embedding: [1, max_audio_len, hidden]
        let pos_emb = vb.take_tensor("embed_positions.weight")?;
        let position_embedding = Var::new(pos_emb, false);

        // Transformer layers
        let mut layers = Vec::with_capacity(config.num_layers);
        for i in 0..config.num_layers {
            let mut layer_vb = vb.pp(&format!("layers.{i}"));
            layers.push(WhisperEncoderLayer::from_varbuilder(
                &mut layer_vb,
                hidden,
                config.num_heads,
            )?);
        }

        // Final layer norm
        let mut ln_vb = vb.pp("layer_norm");
        let ln_w = ln_vb.take_tensor("weight")?;
        let ln_b = ln_vb.take_tensor("bias")?;
        let ln_post = LayerNorm::new(ln_w, ln_b, 1e-5, false);

        Ok(Self {
            conv1,
            conv2,
            position_embedding,
            layers,
            ln_post,
        })
    }

    /// Forward pass (inference only).
    ///
    /// Input: `mel` with shape `[B, num_mel_bins, audio_len]`
    /// Output: `[B, seq_len, hidden]` where `seq_len ≈ audio_len / 2`
    pub fn forward_inference<C>(&self, client: &C, mel: &Tensor<R>) -> Result<Tensor<R>>
    where
        C: RuntimeClient<R>
            + TensorOps<R>
            + ScalarOps<R>
            + MatmulOps<R>
            + BinaryOps<R>
            + ActivationOps<R>
            + NormalizationOps<R>
            + ConvOps<R>
            + ReduceOps<R>
            + ShapeOps<R>,
        R::Client:
            TensorOps<R> + ScalarOps<R> + ConvOps<R> + ReduceOps<R> + BinaryOps<R> + ScalarOps<R>,
    {
        // Conv stem: [B, num_mel_bins, audio_len] -> [B, hidden, seq_len]
        let x = self.conv1.forward_inference(client, mel)?;
        let x = client.gelu(&x).map_err(Error::Numr)?;
        let x = self.conv2.forward_inference(client, &x)?;
        let x = client.gelu(&x).map_err(Error::Numr)?;

        // Transpose to [B, seq_len, hidden]
        let x = x.transpose(1, 2).map_err(Error::Numr)?.contiguous();

        let seq_len = x.shape()[1];

        // Add positional embedding (truncate if audio is shorter than max)
        let pos = self.position_embedding.tensor();
        let pos_seq_len = pos.shape()[0]; // [max_len, hidden]
        let pos_slice = if seq_len <= pos_seq_len {
            pos.narrow(0, 0, seq_len).map_err(Error::Numr)?
        } else {
            // Pad with zeros if audio exceeds max length (unusual but handle gracefully)
            pos.clone()
        };

        // Broadcast add: [B, seq_len, hidden] + [seq_len, hidden]
        let x = client.add(&x, &pos_slice).map_err(Error::Numr)?;

        // Transformer layers
        let mut hidden = x;
        for layer in &self.layers {
            hidden = layer.forward_inference(client, &hidden)?;
        }

        // Final layer norm
        let hidden_var = Var::new(hidden, false);
        let output = self.ln_post.forward(client, &hidden_var)?;

        Ok(output.tensor().clone())
    }
}