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

//! Compare CPU vs GPU layer 0 outputs
use realizar::gguf::{MappedGGUFModel, OwnedQKVWeights, OwnedQuantizedModel};
use realizar::quantize::{fused_q4k_parallel_matvec, fused_q6k_parallel_matvec};
use realizar::rms_norm;

const GGUF_TYPE_Q4_K: u32 = 12;
const GGUF_TYPE_Q6_K: u32 = 14;

fn fused_matmul(input: &[f32], data: &[u8], qtype: u32, in_dim: usize, out_dim: usize) -> Vec<f32> {
    match qtype {
        GGUF_TYPE_Q4_K => fused_q4k_parallel_matvec(data, input, in_dim, out_dim).expect("q4k"),
        GGUF_TYPE_Q6_K => fused_q6k_parallel_matvec(data, input, in_dim, out_dim).expect("q6k"),
        _ => panic!("Unsupported qtype: {}", qtype),
    }
}

fn main() {
    let path = std::env::args().nth(1).unwrap_or_else(|| {
        "../aprender/models/qwen2.5-coder-1.5b-instruct-q4_k_m.gguf".to_string()
    });
    let mapped = MappedGGUFModel::from_path(&path).expect("Failed to load");
    let model = OwnedQuantizedModel::from_mapped(&mapped).expect("Failed to parse");

    let hidden_dim = model.config().hidden_dim;
    let eps = model.config().eps;

    // Token 791
    let token_id = 791u32;
    let start = token_id as usize * hidden_dim;
    let embedding: Vec<f32> = model.token_embedding()[start..start + hidden_dim].to_vec();

    println!("=== CPU Layer 0 Trace ===");
    println!("Embedding first 5: {:?}", &embedding[..5]);

    let layer = &model.layers()[0];

    // RMSNorm
    let normed = rms_norm(&embedding, &layer.attn_norm_weight, eps);
    println!("RMSNorm first 3: {:?}", &normed[..3]);

    // Q, K, V projections
    let OwnedQKVWeights::Separate {
        q: q_weight,
        k: k_weight,
        v: v_weight,
    } = &layer.qkv_weight
    else {
        panic!("Expected separate QKV")
    };

    let mut q = fused_matmul(
        &normed,
        &q_weight.data,
        q_weight.qtype,
        q_weight.in_dim,
        q_weight.out_dim,
    );
    let mut k = fused_matmul(
        &normed,
        &k_weight.data,
        k_weight.qtype,
        k_weight.in_dim,
        k_weight.out_dim,
    );
    let mut v = fused_matmul(
        &normed,
        &v_weight.data,
        v_weight.qtype,
        v_weight.in_dim,
        v_weight.out_dim,
    );

    // PMAT-COR-001 Fix: Add QKV biases (Qwen2.5 requires them!)
    if let Some(ref bias) = layer.qkv_bias {
        let q_dim = q.len();
        let k_dim = k.len();
        // Bias is concatenated: [Q bias (q_dim), K bias (k_dim), V bias (v_dim)]
        for i in 0..q_dim {
            q[i] += bias[i];
        }
        for i in 0..k_dim {
            k[i] += bias[q_dim + i];
        }
        for i in 0..v.len() {
            v[i] += bias[q_dim + k_dim + i];
        }
        println!("QKV bias applied: len={}", bias.len());
    } else {
        println!("WARNING: No QKV bias found!");
    }

    println!("Q first 3: {:?}", &q[..3]);
    println!("K first 5: {:?}", &k[..5]);
    println!("V first 5: {:?}", &v[..5]);

    // Single-token attention: output = V (softmax(single element) = 1.0)
    let head_dim = hidden_dim / model.config().num_heads;
    let group_size = model.config().num_heads / model.config().num_kv_heads;
    let mut attn_out = Vec::with_capacity(hidden_dim);
    for h in 0..model.config().num_heads {
        let kv_head = h / group_size;
        let kv_start = kv_head * head_dim;
        attn_out.extend_from_slice(&v[kv_start..kv_start + head_dim]);
    }
    println!("Attn out first 3: {:?}", &attn_out[..3]);

    // Output projection
    let out_proj = fused_matmul(
        &attn_out,
        &layer.attn_output_weight.data,
        layer.attn_output_weight.qtype,
        layer.attn_output_weight.in_dim,
        layer.attn_output_weight.out_dim,
    );
    println!("Output proj first 3: {:?}", &out_proj[..3]);

    // Residual 1
    let residual1: Vec<f32> = embedding
        .iter()
        .zip(out_proj.iter())
        .map(|(e, o)| e + o)
        .collect();
    println!("Residual1 first 3: {:?}", &residual1[..3]);

    // FFN norm
    let ffn_normed = rms_norm(
        &residual1,
        layer.ffn_norm_weight.as_ref().expect("ffn"),
        eps,
    );

    // FFN gate + up
    let gate_weight = layer.ffn_gate_weight.as_ref().expect("gate");
    let gate = fused_matmul(
        &ffn_normed,
        &gate_weight.data,
        gate_weight.qtype,
        gate_weight.in_dim,
        gate_weight.out_dim,
    );
    let up = fused_matmul(
        &ffn_normed,
        &layer.ffn_up_weight.data,
        layer.ffn_up_weight.qtype,
        layer.ffn_up_weight.in_dim,
        layer.ffn_up_weight.out_dim,
    );
    println!("FFN gate first 3: {:?}", &gate[..3]);
    println!("FFN up first 3: {:?}", &up[..3]);

    // SwiGLU: silu(gate) * up
    let swiglu: Vec<f32> = gate
        .iter()
        .zip(up.iter())
        .map(|(g, u)| (g / (1.0 + (-*g).exp())) * u)
        .collect();
    println!("SwiGLU first 3: {:?}", &swiglu[..3]);

    // FFN down
    let ffn_down = fused_matmul(
        &swiglu,
        &layer.ffn_down_weight.data,
        layer.ffn_down_weight.qtype,
        layer.ffn_down_weight.in_dim,
        layer.ffn_down_weight.out_dim,
    );
    println!("FFN down first 3: {:?}", &ffn_down[..3]);

    // Layer output
    let output: Vec<f32> = residual1
        .iter()
        .zip(ffn_down.iter())
        .map(|(r, f)| r + f)
        .collect();
    println!("Layer output first 3: {:?}", &output[..3]);

    println!("\n=== Comparison with GPU (from debug_layer_by_layer) ===");
    println!("GPU RMSNorm: [-0.9617413, 0.3485948, 0.40064743]");
    println!("GPU Q: [-0.25516012, -0.6839622, -0.04699441]");
    println!("GPU Attn: [-0.11200829, -0.066736706, 0.2103174]");
    println!("GPU Out proj: [-0.95968825, 0.011617601, -0.059631474]");
    println!("GPU Residual1: [-0.9881397, 0.023871362, -0.047377713]");
    println!("GPU FFN gate: [-0.6621405, -0.15498713, -1.1224853]");
    println!("GPU FFN up: [0.58142, 0.092911005, 0.32663453]");
    println!("GPU SwiGLU: [-0.13099347, -0.0066431654, -0.09002926]");
    println!("GPU FFN down: [-0.029813662, 0.17109162, 0.08764193]");
    println!("GPU Layer out: [-1.0179534, 0.19496298, 0.04026422]");
}