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

//! Paged KV cache serialization for disaggregated prefill/decode transfer.
//!
//! # Wire format — `LayeredPagedKvCache`
//!
//! ```text
//! [num_layers:  u32 LE]
//! [block_size:  u32 LE]
//! [seq_len:     u32 LE]
//! For each layer:
//!   [num_blocks:    u32 LE]
//!   [num_heads:     u32 LE]
//!   [head_dim:      u32 LE]
//!   [k_data:        num_blocks * block_size * num_heads * head_dim * 4 bytes (f32 LE)]
//!   [v_data:        num_blocks * block_size * num_heads * head_dim * 4 bytes (f32 LE)]
//!   [block_table_len: u32 LE]
//!   [block_ids:     block_table_len * 4 bytes (u32 LE)]   — BlockId = u32
//! ```

use crate::inference::{BlockTable, LayeredPagedKvCache};
use crate::{DType, Runtime};
use anyhow::{Result, anyhow};

/// Per-layer deserialized K/V data that could not be directly loaded into the
/// paged cache (the public API does not expose mutable raw-block writes).
///
/// The caller is responsible for feeding this data into the paged cache via
/// whatever mechanism is available (e.g. by using the data in model forward
/// passes directly, or by writing via `update()` with proper slot mappings).
#[derive(Debug)]
pub struct PagedLayerData {
    pub k_data: Vec<f32>,
    pub v_data: Vec<f32>,
    pub block_ids: Vec<u32>,
}

/// Serialize a `LayeredPagedKvCache` and its associated `BlockTable` into bytes.
pub fn serialize_paged_kv_cache<R>(
    cache: &LayeredPagedKvCache<R>,
    _block_table: &BlockTable,
) -> Vec<u8>
where
    R: Runtime<DType = DType>,
{
    let num_layers = cache.num_layers() as u32;
    let block_size = cache.block_size() as u32;
    let seq_len = cache.seq_len() as u32;

    let mut buf: Vec<u8> = Vec::new();

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

    for layer_idx in 0..num_layers as usize {
        let layer = cache.layer(layer_idx);
        let bt = cache.block_table(layer_idx);

        let num_blocks = layer.num_blocks() as u32;
        let num_heads = layer.num_heads() as u32;
        let head_dim = layer.head_dim() as u32;

        buf.extend_from_slice(&num_blocks.to_le_bytes());
        buf.extend_from_slice(&num_heads.to_le_bytes());
        buf.extend_from_slice(&head_dim.to_le_bytes());

        let k_data: Vec<f32> = layer.k_cache().to_vec::<f32>();
        let v_data: Vec<f32> = layer.v_cache().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));

        let block_ids = &bt.blocks;
        let bt_len = block_ids.len() as u32;
        buf.extend_from_slice(&bt_len.to_le_bytes());
        for &id in block_ids {
            buf.extend_from_slice(&id.to_le_bytes());
        }
    }

    buf
}

/// Deserialize bytes into a `LayeredPagedKvCache`, per-layer K/V data, and
/// per-layer block tables.
///
/// Returns `(cache, layer_data, block_tables)`. The `layer_data` contains the
/// deserialized K/V float data for each layer — the caller must write this into
/// the cache's backing tensors (the paged cache API does not expose raw block
/// writes, so this data is returned separately).
pub fn deserialize_paged_kv_cache<R>(
    bytes: &[u8],
    device: &R::Device,
) -> Result<(LayeredPagedKvCache<R>, Vec<PagedLayerData>, Vec<BlockTable>)>
where
    R: Runtime<DType = DType>,
{
    if bytes.len() < 12 {
        return Err(anyhow!(
            "Paged KV cache buffer too short: need 12 bytes, got {}",
            bytes.len()
        ));
    }

    let num_layers = u32::from_le_bytes(bytes[0..4].try_into().unwrap()) as usize;
    let block_size = u32::from_le_bytes(bytes[4..8].try_into().unwrap()) as usize;
    let seq_len = u32::from_le_bytes(bytes[8..12].try_into().unwrap()) as usize;

    let mut cursor = 12usize;

    struct RawLayerParams {
        num_blocks: usize,
        num_heads: usize,
        head_dim: usize,
        k_data: Vec<f32>,
        v_data: Vec<f32>,
        block_ids: Vec<u32>,
    }

    let mut raw_layers: Vec<RawLayerParams> = Vec::with_capacity(num_layers);

    for layer_idx in 0..num_layers {
        if cursor + 12 > bytes.len() {
            return Err(anyhow!(
                "Paged KV cache buffer truncated at layer {} header",
                layer_idx
            ));
        }
        let num_blocks = u32::from_le_bytes(bytes[cursor..cursor + 4].try_into().unwrap()) as usize;
        let num_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;
        cursor += 12;

        let numel = num_blocks * block_size * num_heads * head_dim;
        let data_bytes = numel * 4;

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

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

        if cursor + 4 > bytes.len() {
            return Err(anyhow!(
                "Paged KV cache buffer truncated at layer {} block table length",
                layer_idx
            ));
        }
        let bt_len = u32::from_le_bytes(bytes[cursor..cursor + 4].try_into().unwrap()) as usize;
        cursor += 4;

        if cursor + bt_len * 4 > bytes.len() {
            return Err(anyhow!(
                "Paged KV cache buffer truncated at layer {} block table data",
                layer_idx
            ));
        }
        let mut block_ids = Vec::with_capacity(bt_len);
        for i in 0..bt_len {
            let offset = cursor + i * 4;
            let id = u32::from_le_bytes(bytes[offset..offset + 4].try_into().unwrap());
            block_ids.push(id);
        }
        cursor += bt_len * 4;

        raw_layers.push(RawLayerParams {
            num_blocks,
            num_heads,
            head_dim,
            k_data,
            v_data,
            block_ids,
        });
    }

    if raw_layers.is_empty() {
        let cache = LayeredPagedKvCache::<R>::new(0, 0, block_size, 1, 64, DType::F32, device);
        return Ok((cache, Vec::new(), Vec::new()));
    }

    let first = &raw_layers[0];
    let mut paged_cache = LayeredPagedKvCache::<R>::new(
        num_layers,
        first.num_blocks,
        block_size,
        first.num_heads,
        first.head_dim,
        DType::F32,
        device,
    );
    paged_cache.set_seq_len(seq_len);

    let mut block_tables: Vec<BlockTable> = Vec::with_capacity(num_layers);
    let mut layer_data: Vec<PagedLayerData> = Vec::with_capacity(num_layers);

    for params in &raw_layers {
        let mut bt = BlockTable::new(block_size);
        bt.blocks = params.block_ids.clone();
        bt.num_tokens = seq_len;
        block_tables.push(bt);

        layer_data.push(PagedLayerData {
            k_data: params.k_data.clone(),
            v_data: params.v_data.clone(),
            block_ids: params.block_ids.clone(),
        });
    }

    Ok((paged_cache, layer_data, block_tables))
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_deserialize_too_short_buffer() {
        let bytes = [0u8; 4];
        let result =
            deserialize_paged_kv_cache::<crate::CpuRuntime>(&bytes, &crate::CpuDevice::new());
        assert!(result.is_err());
    }

    #[test]
    fn test_deserialize_zero_layers() {
        let mut bytes = Vec::new();
        bytes.extend_from_slice(&0u32.to_le_bytes()); // num_layers = 0
        bytes.extend_from_slice(&16u32.to_le_bytes()); // block_size
        bytes.extend_from_slice(&0u32.to_le_bytes()); // seq_len

        let (cache, layer_data, block_tables) =
            deserialize_paged_kv_cache::<crate::CpuRuntime>(&bytes, &crate::CpuDevice::new())
                .unwrap();
        assert_eq!(cache.num_layers(), 0);
        assert!(layer_data.is_empty());
        assert!(block_tables.is_empty());
    }

    #[test]
    fn test_deserialize_truncated_layer_header() {
        let mut bytes = Vec::new();
        bytes.extend_from_slice(&1u32.to_le_bytes()); // num_layers = 1
        bytes.extend_from_slice(&16u32.to_le_bytes()); // block_size
        bytes.extend_from_slice(&0u32.to_le_bytes()); // seq_len
        // Missing layer header → should error

        let result =
            deserialize_paged_kv_cache::<crate::CpuRuntime>(&bytes, &crate::CpuDevice::new());
        assert!(result.is_err());
    }
}