oxionnx 0.1.2

Pure Rust ONNX inference engine — zero C/C++ dependencies
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
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308

//! LayerNorm pattern fusion pass.
//!
//! Matches: ReduceMean → Sub → Pow(2) → ReduceMean → Add(eps) → Sqrt → Div
//! Optionally followed by Mul(scale) → Add(bias).
//! Replaces with a single LayerNorm node.

use crate::graph::{Attributes, Node, OpKind};
use crate::tensor::Tensor;
use std::collections::{HashMap, HashSet};

/// LayerNorm fusion: match the canonical pattern of 7+ nodes:
///   ReduceMean -> Sub -> Pow(2) -> ReduceMean -> Add(eps) -> Sqrt -> Div
/// Optionally followed by Mul(scale) -> Add(bias).
/// Replace with a single LayerNorm node.
pub fn fuse_layer_norm(nodes: Vec<Node>, weights: &HashMap<String, Tensor>) -> Vec<Node> {
    if nodes.len() < 7 {
        return nodes;
    }

    let mut producer: HashMap<String, usize> = HashMap::new();
    for (i, node) in nodes.iter().enumerate() {
        for out in &node.outputs {
            producer.insert(out.clone(), i);
        }
    }

    let mut consumer_count: HashMap<String, usize> = HashMap::new();
    for node in &nodes {
        for inp in &node.inputs {
            if !inp.is_empty() {
                *consumer_count.entry(inp.clone()).or_insert(0) += 1;
            }
        }
    }

    let mut skip: HashSet<usize> = HashSet::new();
    let mut replacements: HashMap<usize, Node> = HashMap::new();

    let single_consumer =
        |name: &str| -> bool { consumer_count.get(name).copied().unwrap_or(0) == 1 };

    let get_producer = |name: &str| -> Option<usize> { producer.get(name).copied() };

    for (i, node) in nodes.iter().enumerate() {
        if skip.contains(&i) {
            continue;
        }

        if !matches!(node.op, OpKind::Div) {
            continue;
        }
        if node.inputs.len() < 2 {
            continue;
        }

        let div_input0 = &node.inputs[0];
        let div_input1 = &node.inputs[1];

        // Step 7: div_input1 should come from Sqrt
        let sqrt_idx = match get_producer(div_input1) {
            Some(idx) if !skip.contains(&idx) => idx,
            _ => continue,
        };
        if !matches!(nodes[sqrt_idx].op, OpKind::Sqrt) {
            continue;
        }
        if !single_consumer(&nodes[sqrt_idx].outputs[0]) {
            continue;
        }

        // Step 6: Sqrt input should come from Add(var, eps)
        if nodes[sqrt_idx].inputs.is_empty() {
            continue;
        }
        let add_eps_idx = match get_producer(&nodes[sqrt_idx].inputs[0]) {
            Some(idx) if !skip.contains(&idx) => idx,
            _ => continue,
        };
        if !matches!(nodes[add_eps_idx].op, OpKind::Add) {
            continue;
        }
        if !single_consumer(&nodes[add_eps_idx].outputs[0]) {
            continue;
        }

        // Step 5: Add(var, eps) - one input should be a small constant (epsilon)
        if nodes[add_eps_idx].inputs.len() < 2 {
            continue;
        }
        let (var_tensor, epsilon) = {
            let inp0 = &nodes[add_eps_idx].inputs[0];
            let inp1 = &nodes[add_eps_idx].inputs[1];
            if let Some(eps_t) = weights.get(inp1) {
                if eps_t.numel() == 1 && eps_t.data[0] < 0.01 {
                    (inp0.clone(), eps_t.data[0])
                } else if let Some(eps_t2) = weights.get(inp0) {
                    if eps_t2.numel() == 1 && eps_t2.data[0] < 0.01 {
                        (inp1.clone(), eps_t2.data[0])
                    } else {
                        continue;
                    }
                } else {
                    continue;
                }
            } else if let Some(eps_t) = weights.get(inp0) {
                if eps_t.numel() == 1 && eps_t.data[0] < 0.01 {
                    (inp1.clone(), eps_t.data[0])
                } else {
                    continue;
                }
            } else {
                continue;
            }
        };

        // Step 4: var should come from ReduceMean(sq, axes)
        let var_reduce_idx = match get_producer(&var_tensor) {
            Some(idx) if !skip.contains(&idx) => idx,
            _ => continue,
        };
        if !matches!(nodes[var_reduce_idx].op, OpKind::ReduceMean) {
            continue;
        }
        if !single_consumer(&nodes[var_reduce_idx].outputs[0]) {
            continue;
        }

        // Step 3: sq should come from Pow(diff, 2)
        if nodes[var_reduce_idx].inputs.is_empty() {
            continue;
        }
        let pow_idx = match get_producer(&nodes[var_reduce_idx].inputs[0]) {
            Some(idx) if !skip.contains(&idx) => idx,
            _ => continue,
        };
        if !matches!(nodes[pow_idx].op, OpKind::Pow) {
            continue;
        }
        if !single_consumer(&nodes[pow_idx].outputs[0]) {
            continue;
        }
        if nodes[pow_idx].inputs.len() < 2 {
            continue;
        }
        let pow_exp_name = &nodes[pow_idx].inputs[1];
        let is_pow2 = if let Some(exp_t) = weights.get(pow_exp_name) {
            exp_t.numel() == 1 && (exp_t.data[0] - 2.0).abs() < 1e-6
        } else {
            false
        };
        if !is_pow2 {
            continue;
        }

        // Step 2: Pow input[0] should come from Sub(X, mean) = diff
        let pow_diff_name = &nodes[pow_idx].inputs[0];
        if pow_diff_name != div_input0 {
            continue;
        }
        let sub_idx = match get_producer(pow_diff_name) {
            Some(idx) if !skip.contains(&idx) => idx,
            _ => continue,
        };
        if !matches!(nodes[sub_idx].op, OpKind::Sub) {
            continue;
        }
        if nodes[sub_idx].inputs.len() < 2 {
            continue;
        }

        // Step 1: Sub input[1] should come from ReduceMean(X, axes) = mean
        let x_name = &nodes[sub_idx].inputs[0];
        let mean_name = &nodes[sub_idx].inputs[1];
        let mean_reduce_idx = match get_producer(mean_name) {
            Some(idx) if !skip.contains(&idx) => idx,
            _ => continue,
        };
        if !matches!(nodes[mean_reduce_idx].op, OpKind::ReduceMean) {
            continue;
        }
        if !single_consumer(&nodes[mean_reduce_idx].outputs[0]) {
            continue;
        }

        if nodes[mean_reduce_idx].inputs.is_empty() {
            continue;
        }
        if &nodes[mean_reduce_idx].inputs[0] != x_name {
            continue;
        }

        let axes = nodes[mean_reduce_idx].attrs.ints("axes");
        let axis = if axes.is_empty() { -1i64 } else { axes[0] };

        let var_axes = nodes[var_reduce_idx].attrs.ints("axes");
        if !var_axes.is_empty() && !axes.is_empty() && var_axes != axes {
            continue;
        }

        // Now check for optional Mul(scale) and Add(bias) after the Div
        let mut final_output = node.outputs[0].clone();
        let mut scale_name: Option<String> = None;
        let mut bias_name: Option<String> = None;
        let mut extra_skip = Vec::new();

        if single_consumer(&node.outputs[0]) {
            for (j, next_node) in nodes.iter().enumerate() {
                if skip.contains(&j) || j == i {
                    continue;
                }
                if !matches!(next_node.op, OpKind::Mul) {
                    continue;
                }
                if next_node.inputs.len() < 2 {
                    continue;
                }
                let (is_match, s_name) = if next_node.inputs[0] == node.outputs[0]
                    && weights.contains_key(&next_node.inputs[1])
                {
                    (true, next_node.inputs[1].clone())
                } else if next_node.inputs[1] == node.outputs[0]
                    && weights.contains_key(&next_node.inputs[0])
                {
                    (true, next_node.inputs[0].clone())
                } else {
                    (false, String::new())
                };
                if is_match {
                    scale_name = Some(s_name);
                    final_output = next_node.outputs[0].clone();
                    extra_skip.push(j);

                    if single_consumer(&next_node.outputs[0]) {
                        for (k, add_node) in nodes.iter().enumerate() {
                            if skip.contains(&k) || k == j || k == i {
                                continue;
                            }
                            if !matches!(add_node.op, OpKind::Add) {
                                continue;
                            }
                            if add_node.inputs.len() < 2 {
                                continue;
                            }
                            let (is_add_match, b_name) = if add_node.inputs[0]
                                == next_node.outputs[0]
                                && weights.contains_key(&add_node.inputs[1])
                            {
                                (true, add_node.inputs[1].clone())
                            } else if add_node.inputs[1] == next_node.outputs[0]
                                && weights.contains_key(&add_node.inputs[0])
                            {
                                (true, add_node.inputs[0].clone())
                            } else {
                                (false, String::new())
                            };
                            if is_add_match {
                                bias_name = Some(b_name);
                                final_output = add_node.outputs[0].clone();
                                extra_skip.push(k);
                                break;
                            }
                        }
                    }
                    break;
                }
            }
        }

        let mut inputs = vec![x_name.clone()];
        if let Some(ref s) = scale_name {
            inputs.push(s.clone());
        }
        if let Some(ref b) = bias_name {
            inputs.push(b.clone());
        }

        let mut attrs = Attributes::default();
        attrs.floats.insert("epsilon".to_string(), epsilon);
        attrs.ints.insert("axis".to_string(), axis);

        let fused = Node {
            op: OpKind::LayerNorm,
            name: format!("{}_fused_layernorm", nodes[mean_reduce_idx].name),
            inputs,
            outputs: vec![final_output],
            attrs,
        };

        skip.insert(sub_idx);
        skip.insert(pow_idx);
        skip.insert(var_reduce_idx);
        skip.insert(add_eps_idx);
        skip.insert(sqrt_idx);
        skip.insert(i);
        for idx in &extra_skip {
            skip.insert(*idx);
        }

        replacements.insert(mean_reduce_idx, fused);
    }

    nodes
        .into_iter()
        .enumerate()
        .filter(|(i, _)| !skip.contains(i))
        .map(|(i, n)| replacements.remove(&i).unwrap_or(n))
        .collect()
}