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

use super::*;
use crate::trainer::checkpoint::load_checkpoint;
use crate::trainer::test_helpers::*;
use numr::runtime::cpu::CpuRuntime;
use numr::tensor::Tensor;
use std::collections::HashMap;
use tempfile::TempDir;

#[test]
fn test_distributed_save_and_load() {
    let dir = TempDir::new().unwrap();
    let device = make_device();

    // Rank 0 state
    let mut model_r0 = HashMap::new();
    model_r0.insert(
        "embed.weight".to_string(),
        Tensor::<CpuRuntime>::from_slice(&[1.0f32, 2.0, 3.0], &[3], &device),
    );

    // Rank 1 state
    let mut model_r1 = HashMap::new();
    model_r1.insert(
        "head.weight".to_string(),
        Tensor::<CpuRuntime>::from_slice(&[4.0f32, 5.0], &[2], &device),
    );

    let state = make_training_state(100);

    // Save both ranks
    save_distributed_checkpoint(
        dir.path(),
        0,
        2,
        &model_r0,
        None,
        &state,
        ShardingConfig {
            world_size: 2,
            rank: 0,
            owned_params: vec!["embed.weight".to_string()],
            strategy: ShardingStrategy::ZeroPartitioned { stage: 3 },
            split_dims: HashMap::new(),
        },
    )
    .unwrap();

    save_distributed_checkpoint(
        dir.path(),
        1,
        2,
        &model_r1,
        None,
        &state,
        ShardingConfig {
            world_size: 2,
            rank: 1,
            owned_params: vec!["head.weight".to_string()],
            strategy: ShardingStrategy::ZeroPartitioned { stage: 3 },
            split_dims: HashMap::new(),
        },
    )
    .unwrap();

    // Verify directory structure
    assert!(dir.path().join("rank_0/model.safetensors").exists());
    assert!(dir.path().join("rank_1/model.safetensors").exists());
    assert!(dir.path().join("sharding_meta.json").exists());

    // Load each rank
    let (loaded_r0, _, _) =
        load_distributed_checkpoint::<CpuRuntime, _>(dir.path(), 0, &device).unwrap();
    assert!(loaded_r0.contains_key("embed.weight"));

    let (loaded_r1, _, _) =
        load_distributed_checkpoint::<CpuRuntime, _>(dir.path(), 1, &device).unwrap();
    assert!(loaded_r1.contains_key("head.weight"));
}

#[test]
fn test_consolidate_zero_partitioned() {
    let dir = TempDir::new().unwrap();
    let output_dir = TempDir::new().unwrap();
    let device = make_device();

    // Rank 0 owns embed.weight
    let mut model_r0 = HashMap::new();
    model_r0.insert(
        "embed.weight".to_string(),
        Tensor::<CpuRuntime>::from_slice(&[1.0f32, 2.0], &[2], &device),
    );

    // Rank 1 owns head.weight
    let mut model_r1 = HashMap::new();
    model_r1.insert(
        "head.weight".to_string(),
        Tensor::<CpuRuntime>::from_slice(&[3.0f32, 4.0], &[2], &device),
    );

    let state = make_training_state(200);

    save_distributed_checkpoint(
        dir.path(),
        0,
        2,
        &model_r0,
        None,
        &state,
        ShardingConfig {
            world_size: 2,
            rank: 0,
            owned_params: vec!["embed.weight".to_string()],
            strategy: ShardingStrategy::ZeroPartitioned { stage: 3 },
            split_dims: HashMap::new(),
        },
    )
    .unwrap();

    save_distributed_checkpoint(
        dir.path(),
        1,
        2,
        &model_r1,
        None,
        &state,
        ShardingConfig {
            world_size: 2,
            rank: 1,
            owned_params: vec!["head.weight".to_string()],
            strategy: ShardingStrategy::ZeroPartitioned { stage: 3 },
            split_dims: HashMap::new(),
        },
    )
    .unwrap();

    // Update sharding_meta.json with rank 1's info
    let meta_json = std::fs::read_to_string(dir.path().join("sharding_meta.json")).unwrap();
    let mut meta: ShardingMeta = serde_json::from_str(&meta_json).unwrap();
    meta.shards[1].owned_params = vec!["head.weight".to_string()];
    meta.shards[1].split_dims = HashMap::new();
    std::fs::write(
        dir.path().join("sharding_meta.json"),
        serde_json::to_string_pretty(&meta).unwrap(),
    )
    .unwrap();

    // Consolidate
    consolidate_checkpoint(dir.path(), output_dir.path()).unwrap();

    // Verify merged checkpoint has both params
    let (merged, _, merged_state) =
        load_checkpoint::<CpuRuntime, _>(output_dir.path(), &device).unwrap();
    assert_eq!(merged.len(), 2);
    assert!(merged.contains_key("embed.weight"));
    assert!(merged.contains_key("head.weight"));
    assert_eq!(merged_state.step, 200);

    let embed: Vec<f32> = merged["embed.weight"].to_vec();
    assert!((embed[0] - 1.0).abs() < 1e-6);
    let head: Vec<f32> = merged["head.weight"].to_vec();
    assert!((head[0] - 3.0).abs() < 1e-6);
}

#[test]
fn test_distributed_topology_mismatch() {
    let dir = TempDir::new().unwrap();
    let device = make_device();

    let mut model = HashMap::new();
    model.insert(
        "w".to_string(),
        Tensor::<CpuRuntime>::from_slice(&[1.0f32], &[1], &device),
    );
    let state = make_training_state(1);

    save_distributed_checkpoint(
        dir.path(),
        0,
        2,
        &model,
        None,
        &state,
        ShardingConfig {
            world_size: 2,
            rank: 0,
            owned_params: vec!["w".to_string()],
            strategy: ShardingStrategy::Replicated,
            split_dims: HashMap::new(),
        },
    )
    .unwrap();

    // Try loading as rank 3 (out of range for world_size=2)
    let err = load_distributed_checkpoint::<CpuRuntime, _>(dir.path(), 3, &device).unwrap_err();
    let msg = format!("{err}");
    assert!(msg.contains("out of range"), "unexpected error: {msg}");
}

#[test]
fn test_consolidate_tensor_parallel() {
    let dir = TempDir::new().unwrap();
    let output_dir = TempDir::new().unwrap();
    let device = make_device();

    // Rank 0: first half of wq weight [2, 4]
    let mut model_r0 = HashMap::new();
    model_r0.insert(
        "attn.wq".to_string(),
        Tensor::<CpuRuntime>::from_slice(
            &[1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0],
            &[2, 4],
            &device,
        ),
    );
    model_r0.insert(
        "norm.weight".to_string(),
        Tensor::<CpuRuntime>::from_slice(&[1.0f32, 1.0], &[2], &device),
    );

    // Rank 1: second half of wq weight [2, 4]
    let mut model_r1 = HashMap::new();
    model_r1.insert(
        "attn.wq".to_string(),
        Tensor::<CpuRuntime>::from_slice(
            &[9.0f32, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0],
            &[2, 4],
            &device,
        ),
    );
    model_r1.insert(
        "norm.weight".to_string(),
        Tensor::<CpuRuntime>::from_slice(&[1.0f32, 1.0], &[2], &device),
    );

    let state = make_training_state(300);

    let mut split_dims = HashMap::new();
    split_dims.insert("attn.wq".to_string(), 0usize);

    save_distributed_checkpoint(
        dir.path(),
        0,
        2,
        &model_r0,
        None,
        &state,
        ShardingConfig {
            world_size: 2,
            rank: 0,
            owned_params: vec!["attn.wq".to_string(), "norm.weight".to_string()],
            strategy: ShardingStrategy::TensorParallel,
            split_dims: split_dims.clone(),
        },
    )
    .unwrap();

    save_distributed_checkpoint(
        dir.path(),
        1,
        2,
        &model_r1,
        None,
        &state,
        ShardingConfig {
            world_size: 2,
            rank: 1,
            owned_params: vec!["attn.wq".to_string(), "norm.weight".to_string()],
            strategy: ShardingStrategy::TensorParallel,
            split_dims: split_dims.clone(),
        },
    )
    .unwrap();

    // Update sharding_meta.json with rank 1's info
    let meta_json = std::fs::read_to_string(dir.path().join("sharding_meta.json")).unwrap();
    let mut meta: ShardingMeta = serde_json::from_str(&meta_json).unwrap();
    meta.shards[1].owned_params = vec!["attn.wq".to_string(), "norm.weight".to_string()];
    meta.shards[1].split_dims = split_dims;
    std::fs::write(
        dir.path().join("sharding_meta.json"),
        serde_json::to_string_pretty(&meta).unwrap(),
    )
    .unwrap();

    // Consolidate
    consolidate_checkpoint(dir.path(), output_dir.path()).unwrap();

    let (merged, _, merged_state) =
        load_checkpoint::<CpuRuntime, _>(output_dir.path(), &device).unwrap();

    // attn.wq should be [4, 4] (concatenated along dim 0)
    let wq = &merged["attn.wq"];
    assert_eq!(wq.shape(), &[4, 4]);
    let wq_data: Vec<f32> = wq.to_vec();
    // First row from rank 0 should be [1, 2, 3, 4]
    assert!((wq_data[0] - 1.0).abs() < 1e-6);
    assert!((wq_data[1] - 2.0).abs() < 1e-6);
    // First row from rank 1 (at index 8) should be [9, 10, 11, 12]
    assert!((wq_data[8] - 9.0).abs() < 1e-6);
    assert!((wq_data[9] - 10.0).abs() < 1e-6);

    // norm.weight should be [2] (replicated, taken from rank 0)
    let norm = &merged["norm.weight"];
    assert_eq!(norm.shape(), &[2]);
    let norm_data: Vec<f32> = norm.to_vec();
    assert!((norm_data[0] - 1.0).abs() < 1e-6);

    // Training state should be preserved
    assert_eq!(merged_state.step, 300);
}