aprender-serve 0.29.0

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

//! Compare final hidden states between single and multi-token inference
use realizar::gguf::{MappedGGUFModel, OwnedQuantizedModel};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let path = "/home/noah/.cache/huggingface/hub/models--Qwen--Qwen2-0.5B-Instruct-GGUF/snapshots/198f08841147e5196a6a69bd0053690fb1fd3857/qwen2-0_5b-instruct-q4_0.gguf";
    let mapped = MappedGGUFModel::from_path(path)?;
    let model = OwnedQuantizedModel::from_mapped(&mapped)?;
    let vocab = mapped.model.vocabulary().expect("vocab");

    println!("=== Final Hidden State Analysis ===\n");

    let hidden_dim = model.config().hidden_dim;
    let num_layers = model.layers().len();
    let _eps = model.config().eps;

    println!("Model: {} layers, {} hidden_dim", num_layers, hidden_dim);

    // We need to manually trace through the model to get intermediate states
    // Let's use a simpler approach: use forward() and compare logits structure

    // Actually, let's compare the dot product of final hidden with specific token embeddings
    // to understand why "!" gets such high logits

    // First, let's verify that the LM head is correctly computing logits
    // logit[i] = final_hidden · lm_head_weight[i]

    // Get logits for both cases
    let single_logits = model.forward(&[17u32])?; // "2"
    let multi_logits = model.forward(&[17u32, 10, 17, 28])?; // "2+2="

    // Now let's check the LM head weight for token 0 ("!")
    // For tied embeddings, lm_head_weight[i] = token_embedding[i]
    // So logit[0] = hidden · embedding[0]

    // The question is: what's different about the hidden state that makes it
    // have high dot product with token 0 embedding in multi-token case?

    println!("\n=== LM Head Analysis ===");
    println!("LM head weight type: {}", model.lm_head_weight().qtype);
    println!("LM head in_dim: {}", model.lm_head_weight().in_dim);
    println!("LM head out_dim: {}", model.lm_head_weight().out_dim);

    // Check token 0 vs token 19 ("4") embeddings
    let emb_0 = &model.token_embedding()[0..hidden_dim];
    let emb_19 = &model.token_embedding()[19 * hidden_dim..20 * hidden_dim];

    let emb_0_norm: f32 = emb_0.iter().map(|x| x * x).sum::<f32>().sqrt();
    let emb_19_norm: f32 = emb_19.iter().map(|x| x * x).sum::<f32>().sqrt();

    println!("\nEmbedding norms:");
    println!("  Token 0 ('!'): {:.4}", emb_0_norm);
    println!("  Token 19 ('4'): {:.4}", emb_19_norm);

    // Check if there's a systematic pattern in token 0's embedding
    let emb_0_mean: f32 = emb_0.iter().sum::<f32>() / hidden_dim as f32;
    let emb_0_sum: f32 = emb_0.iter().sum();
    let emb_0_abs_sum: f32 = emb_0.iter().map(|x| x.abs()).sum();

    let emb_19_mean: f32 = emb_19.iter().sum::<f32>() / hidden_dim as f32;
    let emb_19_sum: f32 = emb_19.iter().sum();
    let emb_19_abs_sum: f32 = emb_19.iter().map(|x| x.abs()).sum();

    println!("\nEmbedding statistics:");
    println!(
        "  Token 0 ('!'): sum={:.4}, abs_sum={:.4}, mean={:.6}",
        emb_0_sum, emb_0_abs_sum, emb_0_mean
    );
    println!(
        "  Token 19 ('4'): sum={:.4}, abs_sum={:.4}, mean={:.6}",
        emb_19_sum, emb_19_abs_sum, emb_19_mean
    );

    // Hypothesis: If emb_0 has a large positive sum and the final hidden also has a large positive sum,
    // this would explain the high dot product

    // The logit formula is: logit[i] = hidden · lm_head_weight[i]
    // If lm_head_weight is Q8_0 quantized, it's dequantized then dot product computed

    // Let me check if there's a bias in the LM head
    if let Some(bias) = model.lm_head_bias() {
        println!("\nLM head has bias!");
        println!("  Bias[0] ('!'): {:.4}", bias[0]);
        println!("  Bias[19] ('4'): {:.4}", bias[19]);

        // Check if bias[0] is unusually high
        let bias_mean: f32 = bias.iter().sum::<f32>() / bias.len() as f32;
        let bias_max = bias.iter().copied().fold(f32::NEG_INFINITY, f32::max);
        let bias_min = bias.iter().copied().fold(f32::INFINITY, f32::min);

        println!(
            "  Bias stats: mean={:.4}, range=[{:.4}, {:.4}]",
            bias_mean, bias_min, bias_max
        );

        // Find rank of token 0's bias
        let mut bias_indexed: Vec<_> = bias.iter().enumerate().collect();
        bias_indexed.sort_by(|a, b| b.1.partial_cmp(a.1).unwrap());
        let bias_0_rank = bias_indexed.iter().position(|(i, _)| *i == 0).unwrap_or(0) + 1;
        println!("  Token 0 bias rank: {} of {}", bias_0_rank, bias.len());
    } else {
        println!("\nNo LM head bias");
    }

    // Now let's see the difference in logits more carefully
    println!("\n=== Logit Analysis ===");

    // Key tokens
    let tokens_to_check = vec![(0, "!"), (19, "4"), (17, "2"), (11, ",")];

    println!(
        "\n{:>15} {:>12} {:>12} {:>12}",
        "Token", "Single", "Multi", "Diff"
    );
    for (tok, name) in &tokens_to_check {
        let single = single_logits[*tok];
        let multi = multi_logits[*tok];
        let diff = multi - single;
        println!(
            "{:>15} {:>12.4} {:>12.4} {:>12.4}",
            format!("{} ('{}')", tok, name),
            single,
            multi,
            diff
        );
    }

    // Calculate mean shift in logits
    let mean_shift: f32 = multi_logits
        .iter()
        .zip(single_logits.iter())
        .map(|(m, s)| m - s)
        .sum::<f32>()
        / multi_logits.len() as f32;

    let token_0_shift = multi_logits[0] - single_logits[0];

    println!("\nMean logit shift (multi - single): {:.4}", mean_shift);
    println!("Token 0 shift: {:.4}", token_0_shift);
    println!(
        "Token 0 relative shift: {:.4} (vs mean)",
        token_0_shift - mean_shift
    );

    // Check how many tokens got boosted more than token 0
    let shifts: Vec<f32> = multi_logits
        .iter()
        .zip(single_logits.iter())
        .map(|(m, s)| m - s)
        .collect();
    let tokens_boosted_more = shifts.iter().filter(|&&s| s > token_0_shift).count();

    println!("Tokens with larger boost than '!': {}", tokens_boosted_more);

    // Top tokens by boost
    let mut boost_indexed: Vec<_> = shifts.iter().enumerate().collect();
    boost_indexed.sort_by(|a, b| b.1.partial_cmp(a.1).unwrap());

    println!("\nTop 10 tokens by boost (multi - single):");
    for (tok, boost) in boost_indexed.iter().take(10) {
        let s = vocab.get(*tok).map_or("?", |s| s.as_str());
        println!("  Token {} ({:?}): +{:.4}", tok, s, boost);
    }

    Ok(())
}