aprender-serve 0.65.1

Pure Rust ML inference engine built from scratch - model serving for GGUF and safetensors
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


#[test]
fn test_from_gguf_separate_qkv() {
    use crate::gguf::test_helpers::create_q4k_test_data;
    use crate::gguf::{GGUFConfig, OwnedQKVWeights, OwnedQuantizedLayer, OwnedQuantizedModel};

    let hidden_dim = 64;
    let num_heads = 4;
    let num_kv_heads = 2;
    let head_dim = hidden_dim / num_heads;
    let kv_dim = num_kv_heads * head_dim;

    let config = GGUFConfig {
        architecture: "llama".to_string(),
        constraints: crate::gguf::ArchConstraints::from_architecture("llama"),
        hidden_dim,
        intermediate_dim: 128,
        num_heads,
        num_kv_heads,
        num_layers: 1,
        vocab_size: 100,
        context_length: 256,
        rope_theta: 10000.0,
        eps: 1e-5,
        rope_type: 0,
            explicit_head_dim: None,
            query_pre_attn_scalar: None,
        bos_token_id: None,
            eos_token_id: None,
    };

    let q_weight = create_q4k_test_data(hidden_dim, hidden_dim);
    let k_weight = create_q4k_test_data(hidden_dim, kv_dim);
    let v_weight = create_q4k_test_data(hidden_dim, kv_dim);

    let layer = OwnedQuantizedLayer {
        attn_norm_weight: vec![1.0f32; hidden_dim],
        attn_norm_bias: None,
        qkv_weight: OwnedQKVWeights::Separate {
            q: q_weight,
            k: k_weight,
            v: v_weight,
        },
        qkv_bias: None,
        attn_output_weight: create_q4k_test_data(hidden_dim, hidden_dim),
        attn_output_bias: None,
        ffn_up_weight: create_q4k_test_data(hidden_dim, 128),
        ffn_up_bias: None,
        ffn_down_weight: create_q4k_test_data(128, hidden_dim),
        ffn_down_bias: None,
        ffn_gate_weight: Some(create_q4k_test_data(hidden_dim, 128)),
        ffn_gate_bias: None,
        ffn_norm_weight: Some(vec![1.0f32; hidden_dim]),
        ffn_norm_bias: None,
        attn_q_norm_weight: None,
        attn_k_norm_weight: None,
        post_attn_norm_weight: None,
        post_ffw_norm_weight: None,
    };

    let model = OwnedQuantizedModel {
        config: config.clone(),
        token_embedding: vec![0.1f32; 100 * hidden_dim],
        position_embedding: None,
        layers: vec![layer],
        encoder_layers: vec![],
        encoder_output_norm_weight: None,
        encoder_output_norm_bias: None,
        output_norm_weight: vec![1.0f32; hidden_dim],
        output_norm_bias: None,
        lm_head_weight: create_q4k_test_data(hidden_dim, 100),
        lm_head_bias: None,
        #[cfg(feature = "cuda")]
        cuda_executor: None,
        #[cfg(feature = "cuda")]
        cuda_kernel_count: std::sync::atomic::AtomicU64::new(0),
        #[cfg(feature = "cuda")]
        cached_weight_names: std::sync::Mutex::new(std::collections::HashSet::new()),
    };

    let q4_model = QuantizedAprTransformerQ4::from_gguf(&model);

    // Separate QKV should be concatenated into fused format
    assert_eq!(q4_model.layers.len(), 1);
    let layer = &q4_model.layers[0];
    // Concatenated QKV: data should be q_data + k_data + v_data
    let q_data_len = create_q4k_test_data(hidden_dim, hidden_dim).data.len();
    let k_data_len = create_q4k_test_data(hidden_dim, kv_dim).data.len();
    let v_data_len = create_q4k_test_data(hidden_dim, kv_dim).data.len();
    assert_eq!(
        layer.qkv_weight.data.len(),
        q_data_len + k_data_len + v_data_len
    );
    assert_eq!(layer.qkv_weight.in_dim, hidden_dim);
    assert_eq!(layer.qkv_weight.out_dim, hidden_dim + kv_dim + kv_dim);
}

#[test]
fn test_from_gguf_gqa_config() {
    use crate::gguf::test_helpers::create_test_model_with_config;
    use crate::gguf::GGUFConfig;

    let config = GGUFConfig {
        architecture: "llama".to_string(),
        constraints: crate::gguf::ArchConstraints::from_architecture("llama"),
        hidden_dim: 128,
        intermediate_dim: 256,
        num_heads: 8,
        num_kv_heads: 2,
        num_layers: 2,
        vocab_size: 200,
        context_length: 512,
        rope_theta: 10000.0,
        eps: 1e-5,
        rope_type: 0,
            explicit_head_dim: None,
            query_pre_attn_scalar: None,
        bos_token_id: None,
            eos_token_id: None,
    };

    let gguf_model = create_test_model_with_config(&config);
    let q4_model = QuantizedAprTransformerQ4::from_gguf(&gguf_model);

    assert_eq!(q4_model.config.num_heads, 8);
    assert_eq!(q4_model.config.num_kv_heads, 2);
    // create_test_model_with_config creates 1 layer regardless of config.num_layers
    assert_eq!(q4_model.layers.len(), 1);
    assert_eq!(q4_model.config.context_length, 512);
    assert_eq!(q4_model.config.rope_theta, 10000.0);
}

#[test]
fn test_from_gguf_with_ffn_gate() {
    use crate::gguf::test_helpers::create_q4k_test_data;
    use crate::gguf::{GGUFConfig, OwnedQKVWeights, OwnedQuantizedLayer, OwnedQuantizedModel};

    let hidden_dim = 64;
    let config = GGUFConfig {
        architecture: "llama".to_string(),
        constraints: crate::gguf::ArchConstraints::from_architecture("llama"),
        hidden_dim,
        intermediate_dim: 128,
        num_heads: 4,
        num_kv_heads: 4,
        num_layers: 1,
        vocab_size: 100,
        context_length: 256,
        rope_theta: 10000.0,
        eps: 1e-5,
        rope_type: 0,
            explicit_head_dim: None,
            query_pre_attn_scalar: None,
        bos_token_id: None,
            eos_token_id: None,
    };

    let kv_dim = 4 * (hidden_dim / 4);
    let qkv_out = hidden_dim + 2 * kv_dim;

    let layer = OwnedQuantizedLayer {
        attn_norm_weight: vec![1.0f32; hidden_dim],
        attn_norm_bias: None,
        qkv_weight: OwnedQKVWeights::Fused(create_q4k_test_data(hidden_dim, qkv_out)),
        qkv_bias: None,
        attn_output_weight: create_q4k_test_data(hidden_dim, hidden_dim),
        attn_output_bias: None,
        ffn_up_weight: create_q4k_test_data(hidden_dim, 128),
        ffn_up_bias: None,
        ffn_down_weight: create_q4k_test_data(128, hidden_dim),
        ffn_down_bias: None,
        ffn_gate_weight: Some(create_q4k_test_data(hidden_dim, 128)),
        ffn_gate_bias: None,
        ffn_norm_weight: Some(vec![1.0f32; hidden_dim]),
        ffn_norm_bias: None,
        attn_q_norm_weight: None,
        attn_k_norm_weight: None,
        post_attn_norm_weight: None,
        post_ffw_norm_weight: None,
    };

    let model = OwnedQuantizedModel {
        config: config.clone(),
        token_embedding: vec![0.1f32; 100 * hidden_dim],
        position_embedding: None,
        layers: vec![layer],
        encoder_layers: vec![],
        encoder_output_norm_weight: None,
        encoder_output_norm_bias: None,
        output_norm_weight: vec![1.0f32; hidden_dim],
        output_norm_bias: None,
        lm_head_weight: create_q4k_test_data(hidden_dim, 100),
        lm_head_bias: None,
        #[cfg(feature = "cuda")]
        cuda_executor: None,
        #[cfg(feature = "cuda")]
        cuda_kernel_count: std::sync::atomic::AtomicU64::new(0),
        #[cfg(feature = "cuda")]
        cached_weight_names: std::sync::Mutex::new(std::collections::HashSet::new()),
    };

    let q4_model = QuantizedAprTransformerQ4::from_gguf(&model);
    let layer = &q4_model.layers[0];

    // Gate weight should be present
    assert!(layer.ffn_gate_weight.is_some());
    let gate = layer.ffn_gate_weight.as_ref().expect("test value should be present");
    assert_eq!(gate.in_dim, hidden_dim);
    assert_eq!(gate.out_dim, 128);

    // Other weights should be present
    assert_eq!(layer.attn_output_weight.in_dim, hidden_dim);
    assert_eq!(layer.attn_output_weight.out_dim, hidden_dim);
    assert_eq!(layer.ffn_up_weight.in_dim, hidden_dim);
    assert_eq!(layer.ffn_up_weight.out_dim, 128);
    assert_eq!(layer.ffn_down_weight.in_dim, 128);
    assert_eq!(layer.ffn_down_weight.out_dim, hidden_dim);
}