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

//! KV cache serialization for disaggregated prefill/decode transfer.
//!
//! Converts a `LayeredKvCache` into a flat byte buffer suitable for sending
//! over nexar, and reconstructs it on the other side.
//!
//! # Wire format — `LayeredKvCache`
//!
//! ```text
//! [num_layers:  u32 LE]
//! [seq_len:     u32 LE]       — used token count (same for all layers)
//! For each layer:
//!   [batch_size:    u32 LE]
//!   [num_kv_heads:  u32 LE]
//!   [head_dim:      u32 LE]
//!   [k_data:        seq_len * batch_size * num_kv_heads * head_dim * 4 bytes (f32 LE)]
//!   [v_data:        seq_len * batch_size * num_kv_heads * head_dim * 4 bytes (f32 LE)]
//! ```

use crate::inference::LayeredKvCache;
use crate::{DType, IndexingOps, Runtime, Tensor};
use anyhow::{Result, anyhow};

/// Serialize a `LayeredKvCache` into bytes for network transfer.
///
/// Extracts only the *used* portion of each layer's K/V tensors (up to
/// `seq_len` tokens). The resulting bytes contain enough metadata for the
/// receiving side to reconstruct a fresh cache with the same dimensions.
///
/// Note: This function calls `to_vec::<f32>()` on each tensor, which copies
/// data from the device to CPU memory. For GPU tensors this is an intentional
/// transfer — disaggregated inference requires moving the KV cache over the
/// network, so a CPU copy is unavoidable.
pub fn serialize_kv_cache<R>(cache: &LayeredKvCache<R>) -> Vec<u8>
where
    R: Runtime<DType = DType>,
    R::Client: IndexingOps<R>,
{
    let num_layers = cache.num_layers() as u32;
    let seq_len = cache.seq_len() as u32;

    let mut buf: Vec<u8> =
        Vec::with_capacity(8 + num_layers as usize * (12 + seq_len as usize * 4 * 2 * 64 * 32));

    buf.extend_from_slice(&num_layers.to_le_bytes());
    buf.extend_from_slice(&seq_len.to_le_bytes());

    for layer_idx in 0..num_layers as usize {
        let layer = match cache.layer(layer_idx) {
            Some(l) => l,
            None => {
                buf.extend_from_slice(&0u32.to_le_bytes());
                buf.extend_from_slice(&0u32.to_le_bytes());
                buf.extend_from_slice(&0u32.to_le_bytes());
                continue;
            }
        };

        let batch_size = layer.batch_size() as u32;
        let num_kv_heads = layer.num_kv_heads() as u32;
        let head_dim = layer.head_dim() as u32;

        buf.extend_from_slice(&batch_size.to_le_bytes());
        buf.extend_from_slice(&num_kv_heads.to_le_bytes());
        buf.extend_from_slice(&head_dim.to_le_bytes());

        if seq_len == 0 {
            continue;
        }

        match layer.get_kv() {
            Ok((k, v)) => {
                let k_data: Vec<f32> = k.contiguous().to_vec::<f32>();
                let v_data: Vec<f32> = v.contiguous().to_vec::<f32>();
                buf.extend_from_slice(bytemuck::cast_slice::<f32, u8>(&k_data));
                buf.extend_from_slice(bytemuck::cast_slice::<f32, u8>(&v_data));
            }
            Err(_) => {
                let numel = batch_size as usize
                    * num_kv_heads as usize
                    * seq_len as usize
                    * head_dim as usize;
                let zeros = vec![0u8; numel * 4 * 2];
                buf.extend_from_slice(&zeros);
            }
        }
    }

    buf
}

/// Deserialize bytes (produced by [`serialize_kv_cache`]) into a fresh
/// `LayeredKvCache` on the given device.
pub fn deserialize_kv_cache<R>(bytes: &[u8], device: &R::Device) -> Result<LayeredKvCache<R>>
where
    R: Runtime<DType = DType>,
    R::Client: IndexingOps<R>,
{
    if bytes.len() < 8 {
        return Err(anyhow!(
            "KV cache buffer too short: need at least 8 bytes, got {}",
            bytes.len()
        ));
    }

    let num_layers = u32::from_le_bytes(bytes[0..4].try_into().unwrap()) as usize;
    let seq_len = u32::from_le_bytes(bytes[4..8].try_into().unwrap()) as usize;

    let mut cursor = 8usize;

    if num_layers == 0 {
        let cache = LayeredKvCache::<R>::new_positional(0, 1, 1, 1, 64, 1, DType::F32, device)?;
        return Ok(cache);
    }

    if cursor + 12 > bytes.len() {
        return Err(anyhow!("KV cache buffer truncated in layer 0 header"));
    }

    let batch_size = u32::from_le_bytes(bytes[cursor..cursor + 4].try_into().unwrap()) as usize;
    let num_kv_heads =
        u32::from_le_bytes(bytes[cursor + 4..cursor + 8].try_into().unwrap()) as usize;
    let head_dim = u32::from_le_bytes(bytes[cursor + 8..cursor + 12].try_into().unwrap()) as usize;

    let initial_capacity = seq_len.max(1);
    let max_seq_len = (seq_len * 2).max(32768);

    let mut cache = LayeredKvCache::<R>::new_positional(
        num_layers,
        batch_size,
        num_kv_heads,
        initial_capacity,
        max_seq_len,
        head_dim,
        DType::F32,
        device,
    )?;

    for layer_idx in 0..num_layers {
        if cursor + 12 > bytes.len() {
            return Err(anyhow!(
                "KV cache buffer truncated at layer {} header (offset {})",
                layer_idx,
                cursor
            ));
        }

        let layer_batch =
            u32::from_le_bytes(bytes[cursor..cursor + 4].try_into().unwrap()) as usize;
        let layer_heads =
            u32::from_le_bytes(bytes[cursor + 4..cursor + 8].try_into().unwrap()) as usize;
        let layer_dim =
            u32::from_le_bytes(bytes[cursor + 8..cursor + 12].try_into().unwrap()) as usize;
        cursor += 12;

        if seq_len == 0 {
            continue;
        }

        let numel = layer_batch * layer_heads * seq_len * layer_dim;
        let data_bytes = numel * 4;

        if cursor + data_bytes * 2 > bytes.len() {
            return Err(anyhow!(
                "KV cache buffer truncated at layer {} data (need {} bytes, have {})",
                layer_idx,
                data_bytes * 2,
                bytes.len() - cursor
            ));
        }

        let k_f32: Vec<f32> =
            bytemuck::cast_slice::<u8, f32>(&bytes[cursor..cursor + data_bytes]).to_vec();
        cursor += data_bytes;

        let v_f32: Vec<f32> =
            bytemuck::cast_slice::<u8, f32>(&bytes[cursor..cursor + data_bytes]).to_vec();
        cursor += data_bytes;

        let k_tensor = Tensor::<R>::from_slice(
            &k_f32,
            &[layer_batch, layer_heads, seq_len, layer_dim],
            device,
        );
        let v_tensor = Tensor::<R>::from_slice(
            &v_f32,
            &[layer_batch, layer_heads, seq_len, layer_dim],
            device,
        );

        if let Some(layer) = cache.layer_mut(layer_idx) {
            layer
                .update(&k_tensor, &v_tensor)
                .map_err(|e| anyhow!("Failed to write K/V into layer {}: {}", layer_idx, e))?;
        }
    }

    Ok(cache)
}

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

    fn cpu_device() -> CpuDevice {
        CpuDevice::new()
    }

    #[test]
    fn test_serialize_empty_cache() {
        let device = cpu_device();
        let cache =
            LayeredKvCache::<CpuRuntime>::new_positional(2, 1, 2, 4, 64, 32, DType::F32, &device)
                .unwrap();

        let bytes = serialize_kv_cache(&cache);
        assert!(bytes.len() >= 8 + 2 * 12);
    }

    #[test]
    fn test_roundtrip_empty_cache() {
        let device = cpu_device();
        let cache =
            LayeredKvCache::<CpuRuntime>::new_positional(2, 1, 2, 4, 64, 32, DType::F32, &device)
                .unwrap();

        let bytes = serialize_kv_cache(&cache);
        let restored = deserialize_kv_cache::<CpuRuntime>(&bytes, &device).unwrap();

        assert_eq!(restored.num_layers(), 2);
        assert_eq!(restored.seq_len(), 0);
    }

    #[test]
    fn test_roundtrip_with_data() {
        let device = cpu_device();
        let mut cache =
            LayeredKvCache::<CpuRuntime>::new_positional(1, 1, 2, 16, 64, 4, DType::F32, &device)
                .unwrap();

        let k_data: Vec<f32> = (0..24).map(|i| i as f32 * 0.1).collect();
        let v_data: Vec<f32> = (0..24).map(|i| i as f32 * 0.2).collect();
        let k = Tensor::<CpuRuntime>::from_slice(&k_data, &[1, 2, 3, 4], &device);
        let v = Tensor::<CpuRuntime>::from_slice(&v_data, &[1, 2, 3, 4], &device);

        cache.layer_mut(0).unwrap().update(&k, &v).unwrap();
        assert_eq!(cache.seq_len(), 3);

        let bytes = serialize_kv_cache(&cache);
        let restored = deserialize_kv_cache::<CpuRuntime>(&bytes, &device).unwrap();

        assert_eq!(restored.num_layers(), 1);
        assert_eq!(restored.seq_len(), 3);

        let (rk, rv) = restored.layer(0).unwrap().get_kv().unwrap();
        let rk_data: Vec<f32> = rk.contiguous().to_vec::<f32>();
        let rv_data: Vec<f32> = rv.contiguous().to_vec::<f32>();

        for (orig, got) in k_data.iter().zip(rk_data.iter()) {
            assert!((orig - got).abs() < 1e-6, "K mismatch: {} vs {}", orig, got);
        }
        for (orig, got) in v_data.iter().zip(rv_data.iter()) {
            assert!((orig - got).abs() < 1e-6, "V mismatch: {} vs {}", orig, got);
        }
    }
}