boostr 0.2.0

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

use super::super::config::HiddenAct;
use super::*;
use crate::error::Error;
use crate::model::Pooling;
use crate::test_utils::cpu_setup;
use numr::runtime::cpu::CpuRuntime;

#[test]
fn cpu_prefers_f32_compute() {
    assert_eq!(preferred_compute_dtype::<CpuRuntime>(), DType::F32);
}

// from_gguf must select F16 on CUDA (WMMA) when the f16 feature is on, so the
// embed forward uses tensor cores instead of the ~50-100x slower F32 matmul.
// No GPU needed: R::name() is a pure identifier.
#[cfg(all(feature = "cuda", feature = "f16"))]
#[test]
fn cuda_prefers_f16_compute() {
    assert_eq!(
        preferred_compute_dtype::<numr::runtime::cuda::CudaRuntime>(),
        DType::F16
    );
}

fn make_test_pipeline() -> (EmbeddingPipeline<CpuRuntime>, numr::runtime::cpu::CpuClient) {
    let (client, device) = cpu_setup();

    // Use cl100k_base vocab size for a realistic tokenizer
    let tokenizer = splintr::from_pretrained("cl100k_base").unwrap();
    let vocab_size = tokenizer.vocab_size();

    let config = EncoderConfig {
        vocab_size,
        hidden_size: 8,
        num_hidden_layers: 1,
        num_attention_heads: 2,
        intermediate_size: 16,
        max_position_embeddings: 64,
        layer_norm_eps: 1e-12,
        hidden_act: HiddenAct::Gelu,
        type_vocab_size: 0,
        arch_family: crate::model::encoder::config::ArchFamily::Bert,
        padding_token_id: 0,
        compute_dtype: numr::dtype::DType::F32,
        rope_freq_base: 10000.0,
        causal: false,
        ffn_variant: crate::model::encoder::config::FfnVariant::Standard,
        token_type_embed_size: 0,
        num_kv_heads: 0,
        head_dim_explicit: None,
        rms_eps: 1e-6,
        sliding_window: None,
        embed_scale: false,
        max_tokens_per_forward: None,
    };

    let d = &device;
    let encoder = Encoder::from_weights(config, Pooling::Mean, |name| match name {
        "embeddings.word_embeddings.weight" => Ok(Tensor::from_slice(
            &vec![0.1f32; vocab_size * 8],
            &[vocab_size, 8],
            d,
        )),
        "embeddings.position_embeddings.weight" => {
            Ok(Tensor::from_slice(&vec![0.01f32; 64 * 8], &[64, 8], d))
        }
        "embeddings.layer_norm.weight" => Ok(Tensor::from_slice(&[1.0f32; 8], &[8], d)),
        "embeddings.layer_norm.bias" => Ok(Tensor::from_slice(&[0.0f32; 8], &[8], d)),
        n if n.ends_with("query.weight")
            || n.ends_with("key.weight")
            || n.ends_with("value.weight")
            || n.ends_with("attention.output.dense.weight") =>
        {
            Ok(Tensor::from_slice(&vec![0.02f32; 8 * 8], &[8, 8], d))
        }
        n if n.ends_with("query.bias")
            || n.ends_with("key.bias")
            || n.ends_with("value.bias")
            || n.ends_with("attention.output.dense.bias")
            || n.ends_with("output.dense.bias") =>
        {
            Ok(Tensor::from_slice(&[0.0f32; 8], &[8], d))
        }
        n if n.ends_with("LayerNorm.weight") => Ok(Tensor::from_slice(&[1.0f32; 8], &[8], d)),
        n if n.ends_with("LayerNorm.bias") => Ok(Tensor::from_slice(&[0.0f32; 8], &[8], d)),
        n if n.ends_with("intermediate.dense.weight") => {
            Ok(Tensor::from_slice(&vec![0.02f32; 16 * 8], &[16, 8], d))
        }
        n if n.ends_with("intermediate.dense.bias") => {
            Ok(Tensor::from_slice(&[0.0f32; 16], &[16], d))
        }
        n if n.ends_with("output.dense.weight") => {
            Ok(Tensor::from_slice(&vec![0.02f32; 8 * 16], &[8, 16], d))
        }
        _ => Err(Error::ModelError {
            reason: format!("unknown weight: {name}"),
        }),
    })
    .unwrap();

    let pipeline = EmbeddingPipeline::new(encoder, tokenizer, device);
    (pipeline, client)
}

#[test]
fn test_embed_text_returns_hidden_size() {
    let (pipeline, client) = make_test_pipeline();
    let emb = pipeline.embed_text(&client, "hello").unwrap();
    assert_eq!(emb.len(), 8);
}

#[test]
fn test_embed_texts_batch() {
    let (pipeline, client) = make_test_pipeline();
    let embs = pipeline.embed_texts(&client, &["hello", "world"]).unwrap();
    assert_eq!(embs.len(), 2);
    assert_eq!(embs[0].len(), 8);
    assert_eq!(embs[1].len(), 8);
}

#[test]
fn test_embed_texts_empty() {
    let (pipeline, client) = make_test_pipeline();
    let embs = pipeline.embed_texts(&client, &[]).unwrap();
    assert!(embs.is_empty());
}

/// Build a pipeline whose position embeddings are non-uniform so that
/// unmasked padding produces a detectably different mean-pool output.
fn make_pipeline_with_distinct_positions()
-> (EmbeddingPipeline<CpuRuntime>, numr::runtime::cpu::CpuClient) {
    let (client, device) = cpu_setup();

    let tokenizer = splintr::from_pretrained("cl100k_base").unwrap();
    let vocab_size = tokenizer.vocab_size();

    let config = EncoderConfig {
        vocab_size,
        hidden_size: 8,
        num_hidden_layers: 1,
        num_attention_heads: 2,
        intermediate_size: 16,
        max_position_embeddings: 64,
        layer_norm_eps: 1e-12,
        hidden_act: HiddenAct::Gelu,
        type_vocab_size: 0,
        arch_family: crate::model::encoder::config::ArchFamily::Bert,
        padding_token_id: 0,
        compute_dtype: numr::dtype::DType::F32,
        rope_freq_base: 10000.0,
        causal: false,
        ffn_variant: crate::model::encoder::config::FfnVariant::Standard,
        token_type_embed_size: 0,
        num_kv_heads: 0,
        head_dim_explicit: None,
        rms_eps: 1e-6,
        sliding_window: None,
        embed_scale: false,
        max_tokens_per_forward: None,
    };

    let d = &device;
    // Position embeddings: position i has value (i+1) * 0.1 in all 8 dims,
    // so different positions produce distinctly different hidden states.
    let mut pos_emb_data = vec![0.0f32; 64 * 8];
    for pos in 0..64usize {
        let v = (pos + 1) as f32 * 0.1;
        for dim in 0..8usize {
            pos_emb_data[pos * 8 + dim] = v;
        }
    }

    let encoder = Encoder::from_weights(config, Pooling::Mean, |name| match name {
        "embeddings.word_embeddings.weight" => Ok(Tensor::from_slice(
            &vec![0.1f32; vocab_size * 8],
            &[vocab_size, 8],
            d,
        )),
        "embeddings.position_embeddings.weight" => {
            Ok(Tensor::from_slice(&pos_emb_data, &[64, 8], d))
        }
        "embeddings.layer_norm.weight" => Ok(Tensor::from_slice(&[1.0f32; 8], &[8], d)),
        "embeddings.layer_norm.bias" => Ok(Tensor::from_slice(&[0.0f32; 8], &[8], d)),
        n if n.ends_with("query.weight")
            || n.ends_with("key.weight")
            || n.ends_with("value.weight")
            || n.ends_with("attention.output.dense.weight") =>
        {
            Ok(Tensor::from_slice(&vec![0.02f32; 8 * 8], &[8, 8], d))
        }
        n if n.ends_with("query.bias")
            || n.ends_with("key.bias")
            || n.ends_with("value.bias")
            || n.ends_with("attention.output.dense.bias")
            || n.ends_with("output.dense.bias") =>
        {
            Ok(Tensor::from_slice(&[0.0f32; 8], &[8], d))
        }
        n if n.ends_with("LayerNorm.weight") => Ok(Tensor::from_slice(&[1.0f32; 8], &[8], d)),
        n if n.ends_with("LayerNorm.bias") => Ok(Tensor::from_slice(&[0.0f32; 8], &[8], d)),
        n if n.ends_with("intermediate.dense.weight") => {
            Ok(Tensor::from_slice(&vec![0.02f32; 16 * 8], &[16, 8], d))
        }
        n if n.ends_with("intermediate.dense.bias") => {
            Ok(Tensor::from_slice(&[0.0f32; 16], &[16], d))
        }
        n if n.ends_with("output.dense.weight") => {
            Ok(Tensor::from_slice(&vec![0.02f32; 8 * 16], &[8, 16], d))
        }
        _ => Err(Error::ModelError {
            reason: format!("unknown weight: {name}"),
        }),
    })
    .unwrap();

    let pipeline = EmbeddingPipeline::new(encoder, tokenizer, device);
    (pipeline, client)
}

/// Core correctness test: embedding a short sequence alone (no padding)
/// must produce the same vector as embedding it in a batch alongside a
/// longer sequence (where it is padded on the right).
///
/// Without an attention mask the pad tokens contribute to the mean-pool
/// output, causing V1 != V1'.  With the mask they are excluded and
/// V1 == V1' (within float epsilon).
#[test]
fn embed_texts_with_padding_excludes_pad_contamination() {
    let (pipeline, client) = make_pipeline_with_distinct_positions();

    // Embed "hello" alone — no padding, no mask needed.
    let solo = pipeline.embed_texts(&client, &["hello"]).unwrap();
    let v1 = &solo[0];

    // Embed "hello" together with a longer text — "hello" gets padded.
    let batch = pipeline
        .embed_texts(&client, &["hello", "this is a longer input sequence"])
        .unwrap();
    let v1_prime = &batch[0];

    assert_eq!(v1.len(), v1_prime.len());

    // Both should agree to within a small epsilon; if masking is broken
    // they will differ by the contribution of pad-token hidden states.
    let max_diff = v1
        .iter()
        .zip(v1_prime.iter())
        .map(|(a, b)| (a - b).abs())
        .fold(0.0f32, f32::max);

    assert!(
        max_diff < 1e-5,
        "pad contamination detected: max element-wise diff = {max_diff:.3e}"
    );
}

/// Regression guard: a single-text call (no padding path) must not regress.
#[test]
fn embed_text_single_is_stable() {
    let (pipeline, client) = make_pipeline_with_distinct_positions();
    let e1 = pipeline.embed_text(&client, "hello world").unwrap();
    let e2 = pipeline.embed_text(&client, "hello world").unwrap();
    assert_eq!(e1, e2);
}