entrenar 0.7.8

Training & Optimization library with autograd, LoRA, quantization, and model merging
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
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268

//! Sparse model export with sparsity metadata
//!
//! Exports pruned model weights along with a `sparsity_metadata.json` sidecar
//! containing per-tensor sparsity statistics.

use crate::prune::pipeline::metrics::PruningMetrics;
use serde::{Deserialize, Serialize};
use std::collections::HashMap;
use std::path::{Path, PathBuf};

/// Result of sparse model export
#[derive(Debug, Clone)]
pub struct SparseExportResult {
    /// Path to the exported weight file
    pub weights_path: PathBuf,
    /// Path to the sparsity metadata sidecar
    pub metadata_path: PathBuf,
    /// Global sparsity ratio
    pub global_sparsity: f32,
    /// Number of tensors exported
    pub num_tensors: usize,
}

/// Sparsity metadata sidecar (serialized to sparsity_metadata.json)
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct SparsityMetadata {
    /// Format version
    pub version: String,
    /// Global sparsity (fraction of zero parameters)
    pub global_sparsity: f32,
    /// Total parameters
    pub total_parameters: usize,
    /// Parameters pruned (zero)
    pub parameters_pruned: usize,
    /// Per-tensor sparsity information
    pub tensors: Vec<TensorSparsityInfo>,
}

/// Per-tensor sparsity statistics
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct TensorSparsityInfo {
    /// Tensor name
    pub name: String,
    /// Sparsity ratio for this tensor
    pub sparsity: f32,
    /// Number of zero elements
    pub zero_count: usize,
    /// Total elements
    pub total_count: usize,
}

/// Export a sparse (pruned) model with sparsity metadata sidecar
///
/// Produces:
/// - Weight file (SafeTensors format via bytemuck)
/// - `sparsity_metadata.json` with per-tensor sparsity stats
pub fn export_sparse_model(
    weights: &HashMap<String, Vec<f32>>,
    shapes: &HashMap<String, Vec<usize>>,
    metrics: &PruningMetrics,
    output_dir: impl AsRef<Path>,
    filename: &str,
) -> Result<SparseExportResult, std::io::Error> {
    let output_dir = output_dir.as_ref();
    std::fs::create_dir_all(output_dir)?;

    // Compute per-tensor sparsity
    let mut tensor_infos = Vec::new();
    let mut total_zeros = 0usize;
    let mut total_elements = 0usize;

    // Sort keys for deterministic output
    let mut names: Vec<&String> = weights.keys().collect();
    names.sort();

    for name in &names {
        let data = &weights[*name];
        let zero_count = data.iter().filter(|&&v| v == 0.0).count();
        let total = data.len();

        tensor_infos.push(TensorSparsityInfo {
            name: (*name).clone(),
            sparsity: if total > 0 { zero_count as f32 / total as f32 } else { 0.0 },
            zero_count,
            total_count: total,
        });

        total_zeros += zero_count;
        total_elements += total;
    }

    let global_sparsity =
        if total_elements > 0 { total_zeros as f32 / total_elements as f32 } else { 0.0 };

    // Build metadata
    let metadata = SparsityMetadata {
        version: "1.0".to_string(),
        global_sparsity,
        total_parameters: metrics.total_parameters,
        parameters_pruned: metrics.parameters_pruned,
        tensors: tensor_infos,
    };

    // Write weight file as SafeTensors
    let weights_path = output_dir.join(filename);
    {
        use safetensors::tensor::{Dtype, TensorView};

        let tensor_data: Vec<(String, Vec<u8>, Vec<usize>)> = names
            .iter()
            .map(|name| {
                let data = &weights[*name];
                let bytes: Vec<u8> = bytemuck::cast_slice(data).to_vec();
                let shape = shapes.get(*name).cloned().unwrap_or_else(|| vec![data.len()]);
                ((*name).clone(), bytes, shape)
            })
            .collect();

        let views: Vec<(&str, TensorView<'_>)> = tensor_data
            .iter()
            .map(|(name, bytes, shape)| {
                let view = TensorView::new(Dtype::F32, shape.clone(), bytes)
                    .expect("TensorView construction must not fail for valid F32 data");
                (name.as_str(), view)
            })
            .collect();

        let safetensor_bytes = safetensors::serialize(views, None)
            .map_err(|e| std::io::Error::other(e.to_string()))?;

        std::fs::write(&weights_path, safetensor_bytes)?;
    }

    // Write sparsity metadata
    let metadata_path = output_dir.join("sparsity_metadata.json");
    let metadata_json = serde_json::to_string_pretty(&metadata)
        .map_err(|e| std::io::Error::other(e.to_string()))?;
    std::fs::write(&metadata_path, metadata_json)?;

    Ok(SparseExportResult {
        weights_path,
        metadata_path,
        global_sparsity,
        num_tensors: names.len(),
    })
}

#[cfg(test)]
mod tests {
    use super::*;
    use tempfile::TempDir;

    fn make_test_data() -> (HashMap<String, Vec<f32>>, HashMap<String, Vec<usize>>) {
        let mut weights = HashMap::new();
        let mut shapes = HashMap::new();

        // 50% sparse tensor
        let data = vec![1.0, 0.0, 0.0, 2.0, 0.0, 3.0, 0.0, 0.0];
        weights.insert("layer.0.weight".to_string(), data);
        shapes.insert("layer.0.weight".to_string(), vec![2, 4]);

        // 0% sparse tensor
        weights.insert("layer.0.bias".to_string(), vec![0.1, 0.2]);
        shapes.insert("layer.0.bias".to_string(), vec![2]);

        (weights, shapes)
    }

    #[test]
    fn test_export_sparse_creates_files() {
        let (weights, shapes) = make_test_data();
        let metrics = PruningMetrics::new(0.5);
        let tmp = TempDir::new().expect("temp file creation should succeed");

        let result =
            export_sparse_model(&weights, &shapes, &metrics, tmp.path(), "sparse.safetensors")
                .expect("operation should succeed");

        assert!(result.weights_path.exists());
        assert!(result.metadata_path.exists());
        assert_eq!(result.num_tensors, 2);
    }

    #[test]
    fn test_export_sparse_metadata_content() {
        let (weights, shapes) = make_test_data();
        let mut metrics = PruningMetrics::new(0.5);
        metrics.update_sparsity(5, 10);
        let tmp = TempDir::new().expect("temp file creation should succeed");

        export_sparse_model(&weights, &shapes, &metrics, tmp.path(), "sparse.safetensors")
            .expect("parsing should succeed");

        let json = std::fs::read_to_string(tmp.path().join("sparsity_metadata.json"))
            .expect("file read should succeed");
        let meta: SparsityMetadata =
            serde_json::from_str(&json).expect("JSON deserialization should succeed");

        assert_eq!(meta.version, "1.0");
        assert_eq!(meta.total_parameters, 10);
        assert_eq!(meta.parameters_pruned, 5);
        assert_eq!(meta.tensors.len(), 2);
    }

    #[test]
    fn test_per_tensor_sparsity() {
        let (weights, shapes) = make_test_data();
        let metrics = PruningMetrics::new(0.5);
        let tmp = TempDir::new().expect("temp file creation should succeed");

        export_sparse_model(&weights, &shapes, &metrics, tmp.path(), "sparse.safetensors")
            .expect("parsing should succeed");

        let json = std::fs::read_to_string(tmp.path().join("sparsity_metadata.json"))
            .expect("file read should succeed");
        let meta: SparsityMetadata =
            serde_json::from_str(&json).expect("JSON deserialization should succeed");

        // layer.0.bias should have 0% sparsity
        let bias_info = meta
            .tensors
            .iter()
            .find(|t| t.name == "layer.0.bias")
            .expect("operation should succeed");
        assert_eq!(bias_info.sparsity, 0.0);
        assert_eq!(bias_info.zero_count, 0);

        // layer.0.weight should have 5/8 = 62.5% sparsity
        let weight_info = meta
            .tensors
            .iter()
            .find(|t| t.name == "layer.0.weight")
            .expect("operation should succeed");
        assert!(weight_info.sparsity > 0.5);
        assert_eq!(weight_info.zero_count, 5);
    }

    #[test]
    fn test_export_sparse_safetensors_valid() {
        let (weights, shapes) = make_test_data();
        let metrics = PruningMetrics::new(0.5);
        let tmp = TempDir::new().expect("temp file creation should succeed");

        let result =
            export_sparse_model(&weights, &shapes, &metrics, tmp.path(), "sparse.safetensors")
                .expect("operation should succeed");

        // Verify safetensors is valid
        let data = std::fs::read(&result.weights_path).expect("file read should succeed");
        let loaded = safetensors::SafeTensors::deserialize(&data).expect("load should succeed");
        assert_eq!(loaded.len(), 2);
    }

    #[test]
    fn test_export_empty_weights() {
        let weights = HashMap::new();
        let shapes = HashMap::new();
        let metrics = PruningMetrics::new(0.0);
        let tmp = TempDir::new().expect("temp file creation should succeed");

        let result =
            export_sparse_model(&weights, &shapes, &metrics, tmp.path(), "empty.safetensors")
                .expect("operation should succeed");

        assert_eq!(result.num_tensors, 0);
        assert_eq!(result.global_sparsity, 0.0);
    }
}