privacy-filter-rs 0.1.0

OpenAI Privacy Filter — PII detection inference in pure Rust with Burn ML
Documentation 
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

/// Load pretrained weights from a safetensors file.
///
/// Weight key reference (HuggingFace format):
///
///   model.embed_tokens.weight                         [200064, 640]  bf16
///   model.layers.{i}.input_layernorm.weight            [640]         bf16
///   model.layers.{i}.self_attn.q_proj.weight           [896, 640]    bf16
///   model.layers.{i}.self_attn.q_proj.bias             [896]         bf16
///   model.layers.{i}.self_attn.k_proj.weight           [128, 640]    bf16
///   model.layers.{i}.self_attn.k_proj.bias             [128]         bf16
///   model.layers.{i}.self_attn.v_proj.weight           [128, 640]    bf16
///   model.layers.{i}.self_attn.v_proj.bias             [128]         bf16
///   model.layers.{i}.self_attn.o_proj.weight           [640, 896]    bf16
///   model.layers.{i}.self_attn.o_proj.bias             [640]         bf16
///   model.layers.{i}.self_attn.sinks                   [14]          f32
///   model.layers.{i}.post_attention_layernorm.weight   [640]         bf16
///   model.layers.{i}.mlp.router.weight                 [128, 640]    bf16
///   model.layers.{i}.mlp.router.bias                   [128]         bf16
///   model.layers.{i}.mlp.experts.gate_up_proj          [128, 640, 1280] bf16
///   model.layers.{i}.mlp.experts.gate_up_proj_bias     [128, 1280]   bf16
///   model.layers.{i}.mlp.experts.down_proj             [128, 640, 640]  bf16
///   model.layers.{i}.mlp.experts.down_proj_bias        [128, 640]    bf16
///   model.norm.weight                                  [640]         bf16
///   score.weight                                       [33, 640]     bf16
///   score.bias                                         [33]          bf16

use std::collections::HashMap;
use burn::prelude::*;
use burn::module::{Param, ParamId};
use half::bf16;
use safetensors::SafeTensors;

use crate::config::ModelConfig;
use crate::model::privacy_filter::PrivacyFilterModel;

// ── Raw tensor map ────────────────────────────────────────────────────────────

pub struct WeightMap {
    tensors: HashMap<String, (Vec<f32>, Vec<usize>)>,
}

impl WeightMap {
    pub fn from_file(path: &str) -> anyhow::Result<Self> {
        let bytes = std::fs::read(path)?;
        let st = SafeTensors::deserialize(&bytes)?;
        let mut tensors = HashMap::with_capacity(st.len());

        for (raw_key, view) in st.tensors() {
            let key = raw_key.to_string();
            let shape: Vec<usize> = view.shape().to_vec();
            let data = view.data();

            let f32s: Vec<f32> = match view.dtype() {
                safetensors::Dtype::BF16 => data
                    .chunks_exact(2)
                    .map(|b| bf16::from_le_bytes([b[0], b[1]]).to_f32())
                    .collect(),
                safetensors::Dtype::F32 => data
                    .chunks_exact(4)
                    .map(|b| f32::from_le_bytes([b[0], b[1], b[2], b[3]]))
                    .collect(),
                other => anyhow::bail!("unsupported dtype {:?} for key {key}", other),
            };

            tensors.insert(key, (f32s, shape));
        }

        Ok(Self { tensors })
    }

    /// Take a tensor by key, removing it from the map.
    pub fn take<B: Backend, const N: usize>(
        &mut self,
        key: &str,
        device: &B::Device,
    ) -> anyhow::Result<Tensor<B, N>> {
        let (data, shape) = self.tensors.remove(key)
            .ok_or_else(|| anyhow::anyhow!("weight key not found: {key}"))?;

        if shape.len() != N {
            anyhow::bail!("rank mismatch for {key}: expected {N}, got {}", shape.len());
        }

        Ok(Tensor::<B, N>::from_data(
            TensorData::new(data, shape),
            device,
        ))
    }

    pub fn print_keys(&self) {
        let mut keys: Vec<&str> = self.tensors.keys().map(String::as_str).collect();
        keys.sort();
        for k in keys {
            let (_, s) = &self.tensors[k];
            println!("  {k:70}  {s:?}");
        }
    }
}

// ── Weight assignment helpers ────────────────────────────────────────────────

/// Assign weight (transpose [out, in] → [in, out]) + bias into nn::Linear.
fn set_linear_wb<B: Backend>(
    linear: &mut burn::nn::Linear<B>,
    w: Tensor<B, 2>,
    b: Tensor<B, 1>,
) {
    linear.weight = Param::initialized(ParamId::new(), w.transpose());
    linear.bias = Some(Param::initialized(ParamId::new(), b));
}

/// Assign weight only (transpose [out, in] → [in, out]) into nn::Linear.
#[allow(dead_code)]
fn set_linear_w<B: Backend>(
    linear: &mut burn::nn::Linear<B>,
    w: Tensor<B, 2>,
) {
    linear.weight = Param::initialized(ParamId::new(), w.transpose());
}

// ── Model loader ─────────────────────────────────────────────────────────────

pub fn load_model<B: Backend>(
    config: &ModelConfig,
    weights_path: &str,
    device: &B::Device,
) -> anyhow::Result<PrivacyFilterModel<B>> {
    eprintln!("Loading weights from {weights_path} ...");
    let mut wm = WeightMap::from_file(weights_path)?;
    eprintln!("Loaded {} tensors from safetensors file.", wm.tensors.len());

    let mut model = PrivacyFilterModel::new(config, device);

    // 1. Token embedding
    let emb_w: Tensor<B, 2> = wm.take("model.embed_tokens.weight", device)?;
    model.embed_tokens = Param::initialized(ParamId::new(), emb_w);

    // 2. Transformer layers
    for (i, layer) in model.layers.iter_mut().enumerate() {
        let p = format!("model.layers.{i}");

        // Input layernorm
        let ln_w: Tensor<B, 1> = wm.take(&format!("{p}.input_layernorm.weight"), device)?;
        layer.input_layernorm.weight = Param::initialized(ParamId::new(), ln_w);

        // Self-attention projections
        set_linear_wb(
            &mut layer.self_attn.q_proj,
            wm.take(&format!("{p}.self_attn.q_proj.weight"), device)?,
            wm.take(&format!("{p}.self_attn.q_proj.bias"), device)?,
        );
        set_linear_wb(
            &mut layer.self_attn.k_proj,
            wm.take(&format!("{p}.self_attn.k_proj.weight"), device)?,
            wm.take(&format!("{p}.self_attn.k_proj.bias"), device)?,
        );
        set_linear_wb(
            &mut layer.self_attn.v_proj,
            wm.take(&format!("{p}.self_attn.v_proj.weight"), device)?,
            wm.take(&format!("{p}.self_attn.v_proj.bias"), device)?,
        );
        set_linear_wb(
            &mut layer.self_attn.o_proj,
            wm.take(&format!("{p}.self_attn.o_proj.weight"), device)?,
            wm.take(&format!("{p}.self_attn.o_proj.bias"), device)?,
        );

        // Attention sinks
        let sinks: Tensor<B, 1> = wm.take(&format!("{p}.self_attn.sinks"), device)?;
        layer.self_attn.sinks = Param::initialized(ParamId::new(), sinks);

        // Post-attention layernorm
        let pln_w: Tensor<B, 1> = wm.take(&format!("{p}.post_attention_layernorm.weight"), device)?;
        layer.post_attention_layernorm.weight = Param::initialized(ParamId::new(), pln_w);

        // MoE router
        // Router weight in safetensors is [num_experts, hidden_size] (F.linear format)
        // For our manual matmul (hidden @ weight), we need [hidden_size, num_experts]
        let router_w: Tensor<B, 2> = wm.take(&format!("{p}.mlp.router.weight"), device)?;
        layer.mlp.router_weight = Param::initialized(ParamId::new(), router_w.transpose());
        let router_b: Tensor<B, 1> = wm.take(&format!("{p}.mlp.router.bias"), device)?;
        layer.mlp.router_bias = Param::initialized(ParamId::new(), router_b);

        // Expert weights (stored as [num_experts, in_dim, out_dim] for direct matmul)
        let gu: Tensor<B, 3> = wm.take(&format!("{p}.mlp.experts.gate_up_proj"), device)?;
        layer.mlp.gate_up_proj = Param::initialized(ParamId::new(), gu);
        let gu_b: Tensor<B, 2> = wm.take(&format!("{p}.mlp.experts.gate_up_proj_bias"), device)?;
        layer.mlp.gate_up_proj_bias = Param::initialized(ParamId::new(), gu_b);
        let dp: Tensor<B, 3> = wm.take(&format!("{p}.mlp.experts.down_proj"), device)?;
        layer.mlp.down_proj = Param::initialized(ParamId::new(), dp);
        let dp_b: Tensor<B, 2> = wm.take(&format!("{p}.mlp.experts.down_proj_bias"), device)?;
        layer.mlp.down_proj_bias = Param::initialized(ParamId::new(), dp_b);

        // Cache expert weights as CPU vecs for fast per-expert slicing
        layer.mlp.cache_weights();
    }

    // 3. Final norm
    let norm_w: Tensor<B, 1> = wm.take("model.norm.weight", device)?;
    model.norm.weight = Param::initialized(ParamId::new(), norm_w);

    // 4. Classification head
    // score.weight: [num_labels, hidden_size] -> transpose to [hidden_size, num_labels]
    let score_w: Tensor<B, 2> = wm.take("score.weight", device)?;
    model.score_weight = Param::initialized(ParamId::new(), score_w.transpose());
    let score_b: Tensor<B, 1> = wm.take("score.bias", device)?;
    model.score_bias = Param::initialized(ParamId::new(), score_b);

    // Check for remaining keys (should be empty if we loaded everything)
    if !wm.tensors.is_empty() {
        eprintln!("Warning: {} unused weight keys remain:", wm.tensors.len());
        wm.print_keys();
    }

    eprintln!("Model loaded successfully.");
    Ok(model)
}