minillm 0.1.1

A mini inference engine for running transformer language models
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

use crate::tensor::Tensor;
use crate::weights::ModelWeights;

pub struct MultiHeadAttention {
    pub n_heads: usize,
    pub head_dim: usize,
    pub hidden_size: usize,
}

impl MultiHeadAttention {
    pub fn new(hidden_size: usize, n_heads: usize) -> Self {
        assert_eq!(
            hidden_size % n_heads,
            0,
            "hidden_size must be divisible by n_heads"
        );

        Self {
            n_heads,
            head_dim: hidden_size / n_heads,
            hidden_size,
        }
    }

    pub fn forward(
        &self,
        x: &Tensor,
        weights: &ModelWeights,
        layer_idx: usize,
        causal_mask: Option<&Tensor>,
    ) -> Tensor {
        let seq_len = x.shape()[1];
        let batch_size = x.shape()[0];

        // 1. Get combined Q, K, V from c_attn (GPT-2 style: one matrix for all Q,K,V)
        let qkv_weight = &weights.tensors[&format!("h.{}.attn.c_attn.weight", layer_idx)];
        let qkv_bias = &weights.tensors[&format!("h.{}.attn.c_attn.bias", layer_idx)];

        // Linear transformation: [batch, seq_len, hidden] -> [batch, seq_len, 3*hidden]
        let qkv = x.matmul(qkv_weight).add(qkv_bias);

        // 2. Split into Q, K, V: [batch, seq_len, 3*hidden] -> 3 x [batch, seq_len, hidden]
        let (q, k, v) = self.split_qkv(&qkv);

        // 3. Reshape for multi-head: [batch, seq_len, hidden] -> [batch, n_heads, seq_len, head_dim]
        let q = self.reshape_for_heads(&q, batch_size, seq_len);
        let k = self.reshape_for_heads(&k, batch_size, seq_len);
        let v = self.reshape_for_heads(&v, batch_size, seq_len);

        // 4. Compute attention scores: Q @ K^T
        let k_transposed = self.transpose_last_two_dims(&k);
        let scores = q.matmul(&k_transposed);

        // 5. Scale by sqrt(head_dim)
        let scale = 1.0 / (self.head_dim as f32).sqrt();
        let scaled_scores = self.scale_tensor(&scores, scale);

        // 6. Apply causal mask (prevent attending to future tokens)
        let masked_scores = if let Some(mask) = causal_mask {
            self.apply_mask(&scaled_scores, mask)
        } else {
            self.create_and_apply_causal_mask(&scaled_scores, seq_len)
        };

        // 7. Softmax to get attention weights
        let attn_weights = masked_scores.softmax(3); // softmax over last dimension (keys)

        // 8. Apply attention to values: attn_weights @ V
        let attn_output = attn_weights.matmul(&v);

        // 9. Reshape back: [batch, n_heads, seq_len, head_dim] -> [batch, seq_len, hidden]
        let reshaped_output = self.reshape_from_heads(&attn_output, batch_size, seq_len);

        // 10. Final linear projection
        let proj_weight = &weights.tensors[&format!("h.{}.attn.c_proj.weight", layer_idx)];
        let proj_bias = &weights.tensors[&format!("h.{}.attn.c_proj.bias", layer_idx)];

        reshaped_output.matmul(proj_weight).add(proj_bias)
    }

    fn split_qkv(&self, qkv: &Tensor) -> (Tensor, Tensor, Tensor) {
        let shape = qkv.shape();
        let batch_size = shape[0];
        let seq_len = shape[1];
        let total_dim = shape[2];

        assert_eq!(total_dim, 3 * self.hidden_size, "QKV dimension mismatch");

        // Split the last dimension into 3 equal parts
        let mut q_data = Vec::new();
        let mut k_data = Vec::new();
        let mut v_data = Vec::new();

        for batch in 0..batch_size {
            for seq in 0..seq_len {
                for dim in 0..self.hidden_size {
                    // Q: indices 0 to hidden_size-1
                    let q_idx = batch * seq_len * total_dim + seq * total_dim + dim;
                    q_data.push(qkv.data.as_slice().unwrap()[q_idx]);

                    // K: indices hidden_size to 2*hidden_size-1
                    let k_idx =
                        batch * seq_len * total_dim + seq * total_dim + (self.hidden_size + dim);
                    k_data.push(qkv.data.as_slice().unwrap()[k_idx]);

                    // V: indices 2*hidden_size to 3*hidden_size-1
                    let v_idx = batch * seq_len * total_dim
                        + seq * total_dim
                        + (2 * self.hidden_size + dim);
                    v_data.push(qkv.data.as_slice().unwrap()[v_idx]);
                }
            }
        }

        let qkv_shape = [batch_size, seq_len, self.hidden_size];
        (
            Tensor::from_shape(&qkv_shape, q_data),
            Tensor::from_shape(&qkv_shape, k_data),
            Tensor::from_shape(&qkv_shape, v_data),
        )
    }

    fn reshape_for_heads(&self, tensor: &Tensor, batch_size: usize, seq_len: usize) -> Tensor {
        // [batch, seq_len, hidden] -> [batch, seq_len, n_heads, head_dim] -> [batch, n_heads, seq_len, head_dim]
        let mut data = Vec::new();
        let src_data = tensor.data.as_slice().unwrap();

        for batch in 0..batch_size {
            for head in 0..self.n_heads {
                for seq in 0..seq_len {
                    for dim in 0..self.head_dim {
                        let src_idx = batch * seq_len * self.hidden_size
                            + seq * self.hidden_size
                            + head * self.head_dim
                            + dim;
                        data.push(src_data[src_idx]);
                    }
                }
            }
        }

        Tensor::from_shape(&[batch_size, self.n_heads, seq_len, self.head_dim], data)
    }

    fn reshape_from_heads(&self, tensor: &Tensor, batch_size: usize, seq_len: usize) -> Tensor {
        // [batch, n_heads, seq_len, head_dim] -> [batch, seq_len, hidden]
        let mut data = Vec::new();
        let src_data = tensor.data.as_slice().unwrap();

        for batch in 0..batch_size {
            for seq in 0..seq_len {
                for head in 0..self.n_heads {
                    for dim in 0..self.head_dim {
                        let src_idx = batch * self.n_heads * seq_len * self.head_dim
                            + head * seq_len * self.head_dim
                            + seq * self.head_dim
                            + dim;
                        data.push(src_data[src_idx]);
                    }
                }
            }
        }

        Tensor::from_shape(&[batch_size, seq_len, self.hidden_size], data)
    }

    fn transpose_last_two_dims(&self, tensor: &Tensor) -> Tensor {
        // [batch, n_heads, seq_len, head_dim] -> [batch, n_heads, head_dim, seq_len]
        let shape = tensor.shape();
        let batch_size = shape[0];
        let n_heads = shape[1];
        let seq_len = shape[2];
        let head_dim = shape[3];

        let mut data = Vec::new();
        let src_data = tensor.data.as_slice().unwrap();

        for batch in 0..batch_size {
            for head in 0..n_heads {
                for dim in 0..head_dim {
                    for seq in 0..seq_len {
                        let src_idx = batch * n_heads * seq_len * head_dim
                            + head * seq_len * head_dim
                            + seq * head_dim
                            + dim;
                        data.push(src_data[src_idx]);
                    }
                }
            }
        }

        Tensor::from_shape(&[batch_size, n_heads, head_dim, seq_len], data)
    }

    fn scale_tensor(&self, tensor: &Tensor, scale: f32) -> Tensor {
        let scaled_data: Vec<f32> = tensor
            .data
            .as_slice()
            .unwrap()
            .iter()
            .map(|&x| x * scale)
            .collect();

        Tensor::from_shape(&tensor.shape(), scaled_data)
    }

    fn create_and_apply_causal_mask(&self, scores: &Tensor, seq_len: usize) -> Tensor {
        let shape = scores.shape();
        let mut masked_data = scores.data.as_slice().unwrap().to_vec();

        let batch_size = shape[0];
        let n_heads = shape[1];

        // Apply causal mask: set future positions to -inf
        for batch in 0..batch_size {
            for head in 0..n_heads {
                for i in 0..seq_len {
                    for j in (i + 1)..seq_len {
                        // j > i means future position
                        let idx = batch * n_heads * seq_len * seq_len
                            + head * seq_len * seq_len
                            + i * seq_len
                            + j;
                        masked_data[idx] = f32::NEG_INFINITY;
                    }
                }
            }
        }

        Tensor::from_shape(&shape, masked_data)
    }

    fn apply_mask(&self, scores: &Tensor, mask: &Tensor) -> Tensor {
        // Apply custom mask (add mask to scores, where mask has -inf for masked positions)
        scores.add(mask)
    }
}