trustformers-models 0.1.1

Model implementations for TrustformeRS
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

use crate::llama3::{
    config::LLaMA3Config,
    model::{
        LLaMA3Attention, LLaMA3ForCausalLM, LLaMA3MLP, LLaMA3Model, LLaMA3RmsNorm,
        LLaMA3RotaryEmbedding,
    },
    tasks::{format_llama3_chat, LLaMA3ChatModel},
};
use scirs2_core::ndarray::{ArrayD, IxDyn};
use trustformers_core::{tensor::Tensor, traits::Config, traits::Layer};

// ─────────────────────────────────────────────────────────────────────────
// Config
// ─────────────────────────────────────────────────────────────────────────

#[test]
fn test_llama3_config_small_test_valid() {
    let config = LLaMA3Config::small_test();
    assert!(config.validate().is_ok(), "small_test config must be valid");
}

#[test]
fn test_llama3_config_8b_preset() {
    let config = LLaMA3Config::llama3_8b();
    assert_eq!(config.hidden_size, 4096);
    assert_eq!(config.num_attention_heads, 32);
    assert_eq!(config.num_key_value_heads, 8);
    assert_eq!(config.vocab_size, 128256);
    assert_eq!(config.max_position_embeddings, 8192);
    assert!((config.rope_theta - 500000.0).abs() < 1.0);
    assert!(config.uses_gqa(), "8B must use GQA");
    assert!(config.validate().is_ok());
}

#[test]
fn test_llama3_config_70b_preset() {
    let config = LLaMA3Config::llama3_70b();
    assert_eq!(config.hidden_size, 8192);
    assert_eq!(config.num_attention_heads, 64);
    assert_eq!(config.num_key_value_heads, 8);
    assert_eq!(config.num_hidden_layers, 80);
    assert_eq!(config.num_query_groups(), 8);
    assert_eq!(config.head_dim(), 128); // 8192 / 64
    assert!(config.validate().is_ok());
}

#[test]
fn test_llama3_config_invalid_head_division() {
    let config = LLaMA3Config {
        hidden_size: 65, // not divisible by 4
        ..LLaMA3Config::small_test()
    };
    assert!(
        config.validate().is_err(),
        "hidden_size not divisible by num_attention_heads must fail"
    );
}

#[test]
fn test_llama3_config_architecture_label() {
    assert_eq!(LLaMA3Config::small_test().architecture(), "LLaMA-3");
}

// ─────────────────────────────────────────────────────────────────────────
// RMSNorm
// ─────────────────────────────────────────────────────────────────────────

#[test]
fn test_llama3_rmsnorm_construction() {
    let norm = LLaMA3RmsNorm::new(64, 1e-5);
    assert!(norm.is_ok(), "RMSNorm construction must succeed");
    let norm = norm.expect("checked");
    assert_eq!(norm.parameter_count(), 64);
}

#[test]
fn test_llama3_rmsnorm_forward_shape() {
    let norm = LLaMA3RmsNorm::new(8, 1e-5).expect("construction failed");
    let input = Tensor::ones(&[8]).expect("tensor");
    let output = norm.forward(input.clone()).expect("forward failed");
    assert_eq!(output.shape(), input.shape(), "RMSNorm must preserve shape");
}

// ─────────────────────────────────────────────────────────────────────────
// RoPE
// ─────────────────────────────────────────────────────────────────────────

#[test]
fn test_llama3_rope_half_dim() {
    let config = LLaMA3Config::small_test();
    let rope = LLaMA3RotaryEmbedding::new(
        config.head_dim(),
        config.max_position_embeddings,
        config.rope_theta,
    );
    assert_eq!(rope.half_dim(), config.head_dim() / 2);
}

#[test]
fn test_llama3_rope_apply_preserves_shape() {
    let rope = LLaMA3RotaryEmbedding::new(16, 64, 500000.0);
    let q = Tensor::ones(&[4, 16]).expect("tensor");
    let k = Tensor::ones(&[4, 16]).expect("tensor");
    let positions: Vec<usize> = (0..4).collect();
    let (q_out, k_out) = rope.apply_rotary_emb(&q, &k, &positions).expect("RoPE failed");
    assert_eq!(q_out.shape(), q.shape(), "q shape must be preserved");
    assert_eq!(k_out.shape(), k.shape(), "k shape must be preserved");
}

// ─────────────────────────────────────────────────────────────────────────
// GQA repeat_kv
// ─────────────────────────────────────────────────────────────────────────

#[test]
fn test_llama3_repeat_kv_no_op_when_no_gqa() {
    // Build a config with num_kv == num_q → no expansion
    let config = LLaMA3Config {
        num_attention_heads: 4,
        num_key_value_heads: 4,
        ..LLaMA3Config::small_test()
    };
    let attn = LLaMA3Attention::new(&config).expect("construction failed");
    let kv = Tensor::ones(&[8]).expect("tensor");
    let expanded = attn.repeat_kv(&kv).expect("repeat_kv failed");
    assert_eq!(
        expanded.shape(),
        kv.shape(),
        "no expansion when groups == 1"
    );
}

#[test]
fn test_llama3_repeat_kv_expands_correctly() {
    let config = LLaMA3Config::small_test(); // 4 Q, 2 KV → 2× sharing
    let attn = LLaMA3Attention::new(&config).expect("construction failed");
    let head_dim = config.head_dim(); // 64/4 = 16
    let kv = Tensor::ones(&[config.num_key_value_heads * head_dim]).expect("tensor");

    let expanded = attn.repeat_kv(&kv).expect("repeat_kv failed");
    let expanded_len: usize = expanded.shape().iter().product();
    let original_len: usize = kv.shape().iter().product();
    assert_eq!(
        expanded_len,
        original_len * config.num_query_groups(),
        "expanded KV must be num_query_groups × larger"
    );
}

// ─────────────────────────────────────────────────────────────────────────
// Attention forward
// ─────────────────────────────────────────────────────────────────────────

#[test]
fn test_llama3_attention_forward_shape() {
    let config = LLaMA3Config::small_test(); // hidden=64
    let attn = LLaMA3Attention::new(&config).expect("construction failed");
    let seq_len = 3_usize;
    let arr = ArrayD::from_shape_vec(
        IxDyn(&[seq_len, config.hidden_size]),
        vec![0.01f32; seq_len * config.hidden_size],
    )
    .expect("reshape");
    let input = Tensor::F32(arr);
    let output = attn.forward(input).expect("attention forward failed");
    let out_len: usize = output.shape().iter().product();
    assert_eq!(out_len, seq_len * config.hidden_size);
}

// ─────────────────────────────────────────────────────────────────────────
// SwiGLU MLP
// ─────────────────────────────────────────────────────────────────────────

#[test]
fn test_llama3_mlp_construction() {
    let config = LLaMA3Config::small_test();
    let mlp = LLaMA3MLP::new(&config);
    assert!(mlp.is_ok(), "MLP construction must succeed");
    assert!(mlp.expect("checked").parameter_count() > 0);
}

#[test]
fn test_llama3_mlp_swiglu_output_shape() {
    let config = LLaMA3Config::small_test(); // hidden=64, intermediate=128
    let mlp = LLaMA3MLP::new(&config).expect("construction failed");
    // Linear requires at least 2D: shape [seq_len=1, hidden=64]
    let arr = ArrayD::from_shape_vec(IxDyn(&[1, 64]), vec![0.01f32; 64]).expect("reshape");
    let input = Tensor::F32(arr);
    let output = mlp.forward(input).expect("MLP forward failed");
    // Output shape: [1, hidden_size] → total = 64
    let total: usize = output.shape().iter().product();
    assert_eq!(total, 64, "MLP output must equal hidden_size");
}

// ─────────────────────────────────────────────────────────────────────────
// Full model
// ─────────────────────────────────────────────────────────────────────────

#[test]
fn test_llama3_model_construction() {
    let model = LLaMA3Model::new(LLaMA3Config::small_test());
    assert!(model.is_ok(), "LLaMA3Model construction must succeed");
    let model = model.expect("checked");
    assert!(model.parameter_count() > 0);
}

#[test]
fn test_llama3_for_causal_lm_forward() {
    let config = LLaMA3Config::small_test();
    let model = LLaMA3ForCausalLM::new(config.clone()).expect("construction failed");
    let input_ids = vec![1u32, 2, 3];
    let logits = model.forward(input_ids).expect("forward failed");
    let total: usize = logits.shape().iter().product();
    assert_eq!(
        total,
        3 * config.vocab_size,
        "logits = seq_len × vocab_size"
    );
}

// ─────────────────────────────────────────────────────────────────────────
// Chat template
// ─────────────────────────────────────────────────────────────────────────

#[test]
fn test_llama3_chat_format_contains_tokens() {
    let messages = vec![
        ("user".to_string(), "Hello!".to_string()),
        ("assistant".to_string(), "Hi there!".to_string()),
    ];
    let prompt = format_llama3_chat("You are helpful.", &messages);
    assert!(
        prompt.contains("<|begin_of_text|>"),
        "must start with begin_of_text"
    );
    assert!(prompt.contains("<|start_header_id|>system<|end_header_id|>"));
    assert!(prompt.contains("<|start_header_id|>user<|end_header_id|>"));
    assert!(prompt.contains("<|start_header_id|>assistant<|end_header_id|>"));
    assert!(prompt.contains("<|eot_id|>"));
    assert!(prompt.contains("You are helpful."));
    assert!(prompt.contains("Hello!"));
    assert!(prompt.contains("Hi there!"));
}

#[test]
fn test_llama3_chat_model_build_prompt() {
    let model = LLaMA3ChatModel::new(LLaMA3Config::small_test()).expect("construction failed");
    let msgs = vec![("user".to_string(), "Write a poem.".to_string())];
    let prompt = model.build_prompt("You are a poet.", &msgs);
    assert!(prompt.contains("Write a poem."));
    assert!(prompt.contains("You are a poet."));
}