realizar 0.8.4

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


// ============================================================================
// from_apr_bytes: Successful deserialization with layers
// ============================================================================

#[test]
fn test_from_apr_bytes_preserves_layer_data() {
    let transformer = AprTransformer {
        config: AprTransformerConfig {
            architecture: "llama".to_string(),
            hidden_dim: 8,
            num_layers: 1,
            num_heads: 2,
            num_kv_heads: 2,
            vocab_size: 10,
            intermediate_dim: 16,
            context_length: 128,
            rope_theta: 10000.0,
            eps: 1e-5,
            eos_token_id: None,
        ..Default::default()
        },
        token_embedding: vec![0.5; 80],
        layers: vec![AprTransformerLayer {
            attn_norm_weight: vec![1.0; 8],
            attn_norm_bias: None,
            qkv_weight: vec![0.1; 8 * 24],
            qkv_bias: None,
            attn_output_weight: vec![0.2; 64],
            attn_output_bias: None,
            ffn_gate_weight: Some(vec![0.3; 128]),
            ffn_gate_bias: None,
            ffn_up_weight: vec![0.4; 128],
            ffn_up_bias: None,
            ffn_down_weight: vec![0.5; 128],
            ffn_down_bias: None,
            ffn_norm_weight: Some(vec![1.0; 8]),
            ffn_norm_bias: None,
            attn_q_norm_weight: None,
            attn_k_norm_weight: None,
            linear_attn_z_weight: None,
            linear_attn_b_weight: None,
            linear_attn_a_weight: None,
            linear_attn_conv1d_weight: None,
            linear_attn_a_log: None,
            linear_attn_dt_bias: None,
            linear_attn_norm_weight: None,
            moe_gate_weight: None,
            moe_expert_gate_up: None,
            moe_expert_down: None,
            moe_shared_gate: None,
            moe_shared_up: None,
            moe_shared_down: None,
            moe_shared_expert_gate_weight: None,
        }],
        output_norm_weight: vec![1.0; 8],
        output_norm_bias: None,
        lm_head_weight: vec![0.6; 80],
        lm_head_bias: None,
        q4k_layers: None,
        lm_head_weight_q6k: None,
        lm_head_weight_q4k: None,
    };

    let bytes = GgufToAprConverter::to_apr_bytes(&transformer).expect("should serialize");
    let restored = GgufToAprConverter::from_apr_bytes(&bytes).expect("should deserialize");

    // Verify layer count
    assert_eq!(restored.layers.len(), 1);

    // Verify token embedding values
    assert!((restored.token_embedding[0] - 0.5).abs() < 1e-6);

    // Verify layer weight values
    assert!((restored.layers[0].attn_norm_weight[0] - 1.0).abs() < 1e-6);
    assert!((restored.layers[0].qkv_weight[0] - 0.1).abs() < 1e-6);
    assert!((restored.layers[0].attn_output_weight[0] - 0.2).abs() < 1e-6);

    // Verify lm_head values
    assert!((restored.lm_head_weight[0] - 0.6).abs() < 1e-6);
}

// ============================================================================
// from_apr_bytes: Error path - truncated tensor index
// ============================================================================

#[test]
fn test_from_apr_bytes_truncated_at_tensor_index() {
    let transformer = AprTransformer {
        config: AprTransformerConfig {
            architecture: "test".to_string(),
            hidden_dim: 8,
            num_layers: 0,
            num_heads: 2,
            num_kv_heads: 2,
            vocab_size: 10,
            intermediate_dim: 16,
            context_length: 128,
            rope_theta: 10000.0,
            eps: 1e-5,
            eos_token_id: None,
        ..Default::default()
        },
        token_embedding: vec![0.0; 80],
        layers: vec![],
        output_norm_weight: vec![1.0; 8],
        output_norm_bias: None,
        lm_head_weight: vec![0.0; 80],
        lm_head_bias: None,
        q4k_layers: None,
        lm_head_weight_q6k: None,
        lm_head_weight_q4k: None,
    };

    let bytes = GgufToAprConverter::to_apr_bytes(&transformer).expect("should serialize");

    // Truncate right after the tensor index starts (at tensor_index_offset + 5 bytes)
    let tensor_index_offset = u64::from_le_bytes(bytes[24..32].try_into().expect("slice")) as usize;
    let truncated = &bytes[..tensor_index_offset + 5];

    let result = GgufToAprConverter::from_apr_bytes(truncated);
    assert!(
        result.is_err(),
        "Should fail when tensor index is truncated"
    );
}

// ============================================================================
// GgufToAprConverter::stats with non-trivial model
// ============================================================================

#[test]
fn test_stats_with_converted_model() {
    let gguf_data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let transformer = GgufToAprConverter::convert(&gguf_data).expect("should convert");
    let stats = GgufToAprConverter::stats(&transformer);

    assert_eq!(stats.architecture, "llama");
    assert_eq!(stats.hidden_dim, 64);
    assert_eq!(stats.vocab_size, 32);
    assert_eq!(stats.num_layers, 1);
    assert!(stats.total_parameters > 0);
    assert!(stats.memory_bytes_f32 > 0);
    assert!(stats.memory_mb() > 0.0);
    assert!(stats.parameters_m() > 0.0);
}