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

//! RoPE (Rotary Position Embedding) module
//!
//! Wraps the RoPEOps trait as a reusable module with precomputed frequency caches.

use crate::error::Result;
use crate::model::config::RopeScalingConfig;
use crate::ops::RoPEOps;
use numr::autograd::Var;
use numr::ops::TypeConversionOps;
use numr::runtime::{Runtime, RuntimeClient};
use numr::tensor::Tensor;

/// Rotary Position Embedding module
///
/// Stores precomputed cos/sin caches for each position.
/// cos_cache, sin_cache: `[max_seq_len, head_dim/2]`
pub struct RoPE<R: Runtime> {
    cos_cache: Var<R>,
    sin_cache: Var<R>,
}

impl<R: Runtime> RoPE<R> {
    /// Create from precomputed cos/sin caches.
    ///
    /// cos_cache, sin_cache: `[max_seq_len, head_dim/2]`
    pub fn new(cos_cache: Tensor<R>, sin_cache: Tensor<R>) -> Self {
        Self {
            cos_cache: Var::new(cos_cache, false),
            sin_cache: Var::new(sin_cache, false),
        }
    }

    /// Compute frequency caches: cos(pos * freq), sin(pos * freq)
    /// where freq_i = 1 / (base^(2i/dim)), optionally with scaling.
    ///
    /// Supports:
    /// - No scaling (standard RoPE)
    /// - Linear scaling: `freq /= factor`
    /// - Llama3 (NTK-aware): frequency-dependent scaling with low/high freq factors
    ///
    /// Returns `RoPE` with cos/sin caches `[max_seq_len, dim/2]`.
    pub fn precompute_freqs(
        max_seq_len: usize,
        head_dim: usize,
        base: f32,
        scaling: Option<&RopeScalingConfig>,
        device: &<R as Runtime>::Device,
    ) -> Self
    where
        R: Runtime<DType = numr::dtype::DType>,
    {
        let half_dim = head_dim / 2;

        // Compute base frequencies
        let mut freqs: Vec<f32> = (0..half_dim)
            .map(|i| 1.0 / base.powf(2.0 * i as f32 / head_dim as f32))
            .collect();

        // Apply scaling
        if let Some(cfg) = scaling {
            match cfg.scaling_type.as_str() {
                "linear" => {
                    for f in &mut freqs {
                        *f /= cfg.factor;
                    }
                }
                "llama3" => {
                    let old_context_len =
                        cfg.original_max_position_embeddings.unwrap_or(8192) as f32;
                    let low_freq_factor = cfg.low_freq_factor.unwrap_or(1.0);
                    let high_freq_factor = cfg.high_freq_factor.unwrap_or(4.0);
                    let low_freq_wavelen = old_context_len / low_freq_factor;
                    let high_freq_wavelen = old_context_len / high_freq_factor;

                    for f in &mut freqs {
                        let wavelen = 2.0 * std::f32::consts::PI / *f;
                        if wavelen < high_freq_wavelen {
                            // High frequency: no scaling
                        } else if wavelen > low_freq_wavelen {
                            // Low frequency: full linear scaling
                            *f /= cfg.factor;
                        } else {
                            // Middle: smooth interpolation
                            let smooth = (old_context_len / wavelen - low_freq_factor)
                                / (high_freq_factor - low_freq_factor);
                            *f = (1.0 - smooth) * (*f / cfg.factor) + smooth * *f;
                        }
                    }
                }
                _ => {
                    // Unknown scaling type — fall through with unscaled frequencies
                }
            }
        }

        // Build caches
        let mut cos_data = vec![0.0f32; max_seq_len * half_dim];
        let mut sin_data = vec![0.0f32; max_seq_len * half_dim];

        for pos in 0..max_seq_len {
            for (i, &freq) in freqs.iter().enumerate() {
                let angle = pos as f32 * freq;
                cos_data[pos * half_dim + i] = angle.cos();
                sin_data[pos * half_dim + i] = angle.sin();
            }
        }

        let cos_cache = Tensor::<R>::from_slice(&cos_data, &[max_seq_len, half_dim], device);
        let sin_cache = Tensor::<R>::from_slice(&sin_data, &[max_seq_len, half_dim], device);

        Self::new(cos_cache, sin_cache)
    }

    /// Apply RoPE to input tensor `x: [B, H, S, D]`
    ///
    /// Uses the cached cos/sin values for positions 0..S.
    pub fn forward<C>(&self, client: &C, x: &Var<R>) -> Result<Var<R>>
    where
        C: RuntimeClient<R> + RoPEOps<R>,
    {
        client.apply_rope(x, &self.cos_cache, &self.sin_cache)
    }

    /// Cast cos/sin caches to the given dtype (e.g. BF16) so that
    /// per-token casts are avoided during inference.
    pub fn cast_caches(&mut self, dtype: numr::dtype::DType)
    where
        R: Runtime<DType = numr::dtype::DType>,
        R::Client: numr::ops::TypeConversionOps<R>,
    {
        if self.cos_cache.tensor().dtype() != dtype {
            let device = self.cos_cache.tensor().device().clone();
            let client = R::default_client(&device);
            if let Ok(cos) = client.cast(self.cos_cache.tensor(), dtype) {
                self.cos_cache = Var::new(cos, false);
            }
            if let Ok(sin) = client.cast(self.sin_cache.tensor(), dtype) {
                self.sin_cache = Var::new(sin, false);
            }
        }
    }

    pub fn cos_cache(&self) -> &Var<R> {
        &self.cos_cache
    }

    pub fn sin_cache(&self) -> &Var<R> {
        &self.sin_cache
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use numr::runtime::cpu::{CpuDevice, CpuRuntime};

    #[test]
    fn test_rope_precompute_shape() {
        let device = CpuDevice::new();
        let rope = RoPE::<CpuRuntime>::precompute_freqs(128, 64, 10000.0, None, &device);
        assert_eq!(rope.cos_cache().shape(), &[128, 32]);
        assert_eq!(rope.sin_cache().shape(), &[128, 32]);
    }

    #[test]
    fn test_rope_precompute_values() {
        let device = CpuDevice::new();
        let rope = RoPE::<CpuRuntime>::precompute_freqs(4, 8, 10000.0, None, &device);

        let cos: Vec<f32> = rope.cos_cache().tensor().to_vec();
        let sin: Vec<f32> = rope.sin_cache().tensor().to_vec();

        // pos=0: all cos=1, sin=0
        for i in 0..4 {
            assert!((cos[i] - 1.0).abs() < 1e-6, "cos[0,{i}]={}", cos[i]);
            assert!(sin[i].abs() < 1e-6, "sin[0,{i}]={}", sin[i]);
        }
    }

    #[test]
    fn test_rope_forward() {
        let device = CpuDevice::new();
        let client = CpuRuntime::default_client(&device);
        let rope = RoPE::<CpuRuntime>::precompute_freqs(8, 16, 10000.0, None, &device);

        let x = Var::new(
            Tensor::<CpuRuntime>::from_slice(&[0.1f32; 2 * 4 * 16], &[1, 2, 4, 16], &device),
            false,
        );
        let out = rope.forward(&client, &x).unwrap();
        assert_eq!(out.shape(), &[1, 2, 4, 16]);
    }

    #[test]
    fn test_rope_linear_scaling() {
        let device = CpuDevice::new();
        let cfg = RopeScalingConfig {
            scaling_type: "linear".to_string(),
            factor: 2.0,
            original_max_position_embeddings: None,
            low_freq_factor: None,
            high_freq_factor: None,
            attention_factor: None,
            beta_fast: None,
            beta_slow: None,
        };

        let unscaled = RoPE::<CpuRuntime>::precompute_freqs(4, 8, 10000.0, None, &device);
        let scaled = RoPE::<CpuRuntime>::precompute_freqs(4, 8, 10000.0, Some(&cfg), &device);

        let cos_unscaled: Vec<f32> = unscaled.cos_cache().tensor().to_vec();
        let cos_scaled: Vec<f32> = scaled.cos_cache().tensor().to_vec();

        // At pos=0, both should be all 1s (cos(0)=1)
        assert!((cos_scaled[0] - 1.0).abs() < 1e-6);

        // At pos=2 scaled should match pos=1 unscaled (freq halved → angle halved)
        let half_dim = 4;
        for i in 0..half_dim {
            let expected = cos_unscaled[half_dim + i]; // pos=1 unscaled
            let actual = cos_scaled[2 * half_dim + i]; // pos=2 scaled
            assert!(
                (actual - expected).abs() < 1e-5,
                "dim {i}: expected {expected}, got {actual}"
            );
        }
    }

    #[test]
    fn test_rope_llama3_scaling() {
        let device = CpuDevice::new();
        let cfg = RopeScalingConfig {
            scaling_type: "llama3".to_string(),
            factor: 8.0,
            original_max_position_embeddings: Some(8192),
            low_freq_factor: Some(1.0),
            high_freq_factor: Some(4.0),
            attention_factor: None,
            beta_fast: None,
            beta_slow: None,
        };

        let rope = RoPE::<CpuRuntime>::precompute_freqs(128, 64, 500000.0, Some(&cfg), &device);
        assert_eq!(rope.cos_cache().shape(), &[128, 32]);
        // Verify it doesn't panic and produces valid values
        let cos: Vec<f32> = rope.cos_cache().tensor().to_vec();
        for &v in &cos {
            assert!(v.is_finite(), "non-finite cos value: {v}");
            assert!((-1.0..=1.0).contains(&v), "cos out of range: {v}");
        }
    }
}