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
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


impl GGUFTransformer {
    /// Load transformer weights from GGUF model
    ///
    /// # Arguments
    ///
    /// * `model` - Parsed GGUF model
    /// * `file_data` - Original file bytes for tensor extraction
    ///
    /// # Errors
    ///
    /// Returns error if required tensors are missing or malformed
    pub fn from_gguf(model: &GGUFModel, file_data: &[u8]) -> Result<Self> {
        // Phase 2: Validate config at construction boundary.
        let config = ValidatedModelConfig::from_gguf(model)?.into_inner();

        // Load token embedding
        let token_embedding = model.get_tensor_f32("token_embd.weight", file_data)?;
        // GH-278: Position embedding — standard GGUF name + aprender export fallback
        let position_embedding = model
            .get_tensor_f32("token_pos_embd.weight", file_data)
            .or_else(|_| model.get_tensor_f32("model.position_embedding.weight", file_data))
            .ok();

        // Load layers
        let mut layers = Vec::with_capacity(config.num_layers);
        for layer_idx in 0..config.num_layers {
            let layer = Self::load_layer(model, file_data, layer_idx)?;
            layers.push(layer);
        }

        // Load output norm (raw gamma values - no delta transformation needed)
        let output_norm_weight = model.get_tensor_f32("output_norm.weight", file_data)?;
        // GH-278: Output norm bias — standard + aprender fallback
        let output_norm_bias = model
            .get_tensor_f32("output_norm.bias", file_data)
            .or_else(|_| model.get_tensor_f32("model.norm.bias", file_data))
            .ok();

        // Load LM head (output projection)
        // Fall back to token_embd.weight for tied embeddings (Qwen2, some LLaMA variants)
        let lm_head_weight = model
            .get_tensor_f32("output.weight", file_data)
            .or_else(|_| model.get_tensor_f32("token_embd.weight", file_data))?;
        let lm_head_bias = model.get_tensor_f32("output.bias", file_data).ok();

        Ok(Self {
            config,
            token_embedding,
            position_embedding,
            layers,
            output_norm_weight,
            output_norm_bias,
            lm_head_weight,
            lm_head_bias,
        })
    }

    /// Load a single transformer layer
    ///
    /// Supports both tensor naming conventions:
    /// - phi-2 style: combined `attn_qkv.weight`
    /// - llama style: separate `attn_q.weight`, `attn_k.weight`, `attn_v.weight`
    fn load_layer(
        model: &GGUFModel,
        file_data: &[u8],
        layer_idx: usize,
    ) -> Result<GGUFTransformerLayer> {
        let prefix = format!("blk.{}", layer_idx);

        // Attention norm weights
        let attn_norm_weight =
            model.get_tensor_f32(&format!("{}.attn_norm.weight", prefix), file_data)?;
        // GH-278: Attention norm bias — standard GGUF + aprender fallback
        let attn_norm_bias = model
            .get_tensor_f32(&format!("{}.attn_norm.bias", prefix), file_data)
            .or_else(|_| {
                model.get_tensor_f32(&format!("{}.input_layernorm.bias", prefix), file_data)
            })
            .ok();

        // QKV weights - try combined first (phi-2), fall back to separate (llama)
        let (qkv_weight, qkv_bias) = if let Ok(combined) =
            model.get_tensor_f32(&format!("{}.attn_qkv.weight", prefix), file_data)
        {
            // phi-2 style: combined QKV tensor
            let bias = model
                .get_tensor_f32(&format!("{}.attn_qkv.bias", prefix), file_data)
                .ok();
            (combined, bias)
        } else {
            // llama style: separate Q, K, V tensors - concatenate them
            let q_weight = model.get_tensor_f32(&format!("{}.attn_q.weight", prefix), file_data)?;
            let k_weight = model.get_tensor_f32(&format!("{}.attn_k.weight", prefix), file_data)?;
            let v_weight = model.get_tensor_f32(&format!("{}.attn_v.weight", prefix), file_data)?;

            // Concatenate Q, K, V weights
            let mut qkv = Vec::with_capacity(q_weight.len() + k_weight.len() + v_weight.len());
            qkv.extend_from_slice(&q_weight);
            qkv.extend_from_slice(&k_weight);
            qkv.extend_from_slice(&v_weight);

            // Try to get biases (llama usually doesn't have them)
            let q_bias = model
                .get_tensor_f32(&format!("{}.attn_q.bias", prefix), file_data)
                .ok();
            let k_bias = model
                .get_tensor_f32(&format!("{}.attn_k.bias", prefix), file_data)
                .ok();
            let v_bias = model
                .get_tensor_f32(&format!("{}.attn_v.bias", prefix), file_data)
                .ok();

            let bias = match (q_bias, k_bias, v_bias) {
                (Some(q), Some(k), Some(v)) => {
                    let mut combined_bias = Vec::with_capacity(q.len() + k.len() + v.len());
                    combined_bias.extend_from_slice(&q);
                    combined_bias.extend_from_slice(&k);
                    combined_bias.extend_from_slice(&v);
                    Some(combined_bias)
                },
                _ => None,
            };

            (qkv, bias)
        };

        // Attention output
        let attn_output_weight =
            model.get_tensor_f32(&format!("{}.attn_output.weight", prefix), file_data)?;
        let attn_output_bias = model
            .get_tensor_f32(&format!("{}.attn_output.bias", prefix), file_data)
            .ok();

        // FFN gate (SwiGLU models like llama have this)
        let ffn_gate_weight = model
            .get_tensor_f32(&format!("{}.ffn_gate.weight", prefix), file_data)
            .ok();
        let ffn_gate_bias = model
            .get_tensor_f32(&format!("{}.ffn_gate.bias", prefix), file_data)
            .ok();

        // FFN up/down projections
        let ffn_up_weight =
            model.get_tensor_f32(&format!("{}.ffn_up.weight", prefix), file_data)?;
        // GH-278: FFN biases — standard GGUF + aprender fallback
        let ffn_up_bias = model
            .get_tensor_f32(&format!("{}.ffn_up.bias", prefix), file_data)
            .or_else(|_| {
                model.get_tensor_f32(&format!("{}.mlp.up_proj.bias", prefix), file_data)
            })
            .ok();
        let ffn_down_weight =
            model.get_tensor_f32(&format!("{}.ffn_down.weight", prefix), file_data)?;
        let ffn_down_bias = model
            .get_tensor_f32(&format!("{}.ffn_down.bias", prefix), file_data)
            .or_else(|_| {
                model.get_tensor_f32(&format!("{}.mlp.down_proj.bias", prefix), file_data)
            })
            .ok();

        // FFN norm (models with separate FFN normalization)
        let ffn_norm_weight = model
            .get_tensor_f32(&format!("{}.ffn_norm.weight", prefix), file_data)
            .ok();
        // GH-278: FFN norm bias — standard GGUF + aprender fallback
        let ffn_norm_bias = model
            .get_tensor_f32(&format!("{}.ffn_norm.bias", prefix), file_data)
            .or_else(|_| {
                model.get_tensor_f32(
                    &format!("{}.post_attention_layernorm.bias", prefix),
                    file_data,
                )
            })
            .ok();

        // GH-279: QK norm weights (Qwen3 per-head RMSNorm on Q and K)
        let attn_q_norm_weight = model
            .get_tensor_f32(&format!("{prefix}.attn_q_norm.weight"), file_data)
            .ok();
        let attn_k_norm_weight = model
            .get_tensor_f32(&format!("{prefix}.attn_k_norm.weight"), file_data)
            .ok();

        Ok(GGUFTransformerLayer {
            attn_norm_weight,
            attn_norm_bias,
            qkv_weight,
            qkv_bias,
            attn_output_weight,
            attn_output_bias,
            ffn_gate_weight,
            ffn_gate_bias,
            ffn_up_weight,
            ffn_up_bias,
            ffn_down_weight,
            ffn_down_bias,
            ffn_norm_weight,
            ffn_norm_bias,
            attn_q_norm_weight,
            attn_k_norm_weight,
        })
    }
}