tract-core 0.23.5

Tiny, no-nonsense, self contained, TensorFlow and ONNX inference
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

use tract_data::TooEarly;

use crate::internal::*;
use crate::ops::array::Slice;
use crate::ops::dummy::Dummy;
use crate::ops::konst::Const;
use crate::ops::source::TypedSource;
use crate::optim::{CONST_FOLD_MEM_BUDGET, OptimizerSession};

/// Replaces stateless nodes whose inputs are all constant with the constant they
/// evaluate to, walking forward through single-successor chains so one patch can
/// collapse a whole run.
///
/// [`CONST_FOLD_MEM_BUDGET`] is weighed only against an input that has several
/// successors, so a large weight is not duplicated once per consumer; a
/// sole-consumer input, or any input of a `Slice`, is exempt whatever its size.
/// Every fold, including each step of the walk, must additionally keep its output
/// within the allowance.
#[derive(Clone, Debug, Default)]
pub struct PropConst(usize);

impl super::TypedPass for PropConst {
    fn reset(&mut self) -> TractResult<()> {
        self.0 = 0;
        Ok(())
    }
    fn next(
        &mut self,
        _session: &mut OptimizerSession,
        model: &TypedModel,
    ) -> TractResult<Option<TypedModelPatch>> {
        for node in &model.nodes[self.0..] {
            if node.op_is::<Const>() && node.outputs[0].fact.konst.is_none() {
                self.0 = node.id;
                let mut patch = TypedModelPatch::default();
                let wire =
                    patch.add_const(&node.name, node.op_as::<Const>().unwrap().val().clone())?;
                patch.shunt_outside(model, node.id.into(), wire)?;
                return Ok(Some(patch));
            }
            let inputs = model.node_input_facts(node.id)?;
            if !node.op_is::<Const>()
                && !node.op_is::<Dummy>()
                && !node.op_is::<TypedSource>()
                && node.op.is_stateless()
                && inputs.iter().zip(&node.inputs).all(|(fact, outlet)| {
                    fact.konst.is_some()
                        && (model.node(outlet.node).outputs[outlet.slot].successors.len() == 1
                            || node.op_is::<Slice>()
                            || (fact.datum_type.is_number()
                                && fact
                                    .mem_size()
                                    .as_i64()
                                    .is_some_and(|m| m as u64 <= CONST_FOLD_MEM_BUDGET)))
                })
            {
                let inputs =
                    inputs.iter().map(|f| f.konst.clone().unwrap().into_tvalue()).collect();
                let input_mem: u64 = model
                    .node_input_facts(node.id)?
                    .iter()
                    .map(|f| f.mem_size().as_i64().unwrap_or(i64::MAX) as u64)
                    .sum();
                match node.op.eval_with_session(node.id, &TurnState::default(), inputs) {
                    Ok(mut res) => {
                        self.0 = node.id;
                        let output_mem: u64 = res
                            .iter()
                            .map(|t| (t.datum_type().size_of() * t.volume()) as u64)
                            .sum();
                        if output_mem > input_mem.max(CONST_FOLD_MEM_BUDGET) {
                            continue;
                        }
                        let mut node = node;
                        loop {
                            let Some(succ) = model.single_succ(node.id)? else {
                                break;
                            };
                            if succ.inputs.len() > 1 || !succ.op.is_stateless() {
                                break;
                            }
                            let Ok(succ_res) = succ.op.eval_with_session(
                                node.id,
                                &TurnState::default(),
                                res.clone(),
                            ) else {
                                break;
                            };
                            let succ_mem: u64 = succ_res
                                .iter()
                                .map(|t| (t.datum_type().size_of() * t.volume()) as u64)
                                .sum();
                            if succ_mem > input_mem.max(CONST_FOLD_MEM_BUDGET) {
                                break;
                            }
                            res = succ_res;
                            node = succ;
                        }
                        let mut patch = TypedModelPatch::default();
                        for (ix, output) in res.into_iter().enumerate() {
                            let exotic_fact =
                                model.outlet_fact(OutletId::new(node.id, ix))?.exotic_fact.clone();

                            let name = if ix > 0 {
                                format!("{}.{ix}", node.name)
                            } else {
                                node.name.clone()
                            };
                            let wire = patch.wire_node(
                                name,
                                Const::new_with_opt_exotic_fact(
                                    output.into_arc_tensor(),
                                    exotic_fact,
                                )?,
                                &[],
                            )?[0];
                            patch.shunt_outside(model, (node.id, ix).into(), wire)?;
                        }
                        self.0 = node.id;
                        return Ok(Some(patch));
                    }
                    Err(e) => {
                        if !e.root_cause().is::<TooEarly>() {
                            Err(e).with_context(|| {
                                format!("Eager eval {node} during optimisation")
                            })?;
                        }
                    }
                }
            }
        }
        Ok(None)
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::ops::math;

    fn const_i64(model: &mut TypedModel, name: &str, shape: [usize; 2]) -> TractResult<OutletId> {
        let len = shape[0] * shape[1];
        let t = Tensor::from_shape(&shape, &(0..len as i64).collect::<Vec<_>>())?;
        model.add_const(name, t)
    }

    /// Sides are derived from the budget rather than fixed, so changing
    /// [`CONST_FOLD_MEM_BUDGET`] cannot silently move a case to the other side of
    /// the guard.
    fn side_for(bytes: u64) -> usize {
        (bytes / std::mem::size_of::<i64>() as u64).isqrt() as usize
    }

    /// Two consumers of one constant, both foldable. The constant is well over the
    /// element count the guard used to allow but inside the memory budget, so the
    /// graph must collapse entirely: once both consumers are constant the shared
    /// input is dead and nothing has been duplicated.
    #[test]
    fn fold_through_shared_const_inside_mem_budget() -> TractResult<()> {
        let mut model = TypedModel::default();
        let side = side_for(CONST_FOLD_MEM_BUDGET / 16);
        let shared = const_i64(&mut model, "shared", [side, side])?;
        let one = model.add_const("one", tensor2(&[[1i64]]))?;
        let two = model.add_const("two", tensor2(&[[2i64]]))?;
        let a = model.wire_node("a", math::mul(), &[shared, two])?[0];
        let b = model.wire_node("b", math::add(), &[shared, one])?[0];
        let sum = model.wire_node("sum", math::add(), &[a, b])?[0];
        model.select_output_outlets(&[sum])?;

        let decluttered = model.into_decluttered()?;
        let live: Vec<&str> = decluttered
            .nodes
            .iter()
            .filter(|n| !n.op_is::<Const>() && !n.op_is::<Dummy>())
            .map(|n| n.name.as_str())
            .collect();
        assert!(live.is_empty(), "expected a fully folded graph, got {live:?}");
        Ok(())
    }

    /// The same shape of graph with a constant over the budget keeps its consumers,
    /// so a large weight is not turned into one copy per consumer.
    #[test]
    fn keep_shared_const_over_mem_budget() -> TractResult<()> {
        let mut model = TypedModel::default();
        let side = side_for(CONST_FOLD_MEM_BUDGET * 2);
        let shared = const_i64(&mut model, "shared", [side, side])?;
        let one = model.add_const("one", tensor2(&[[1i64]]))?;
        let two = model.add_const("two", tensor2(&[[2i64]]))?;
        let a = model.wire_node("a", math::mul(), &[shared, two])?[0];
        let b = model.wire_node("b", math::add(), &[shared, one])?[0];
        model.select_output_outlets(&[a, b])?;

        let decluttered = model.into_decluttered()?;
        assert!(
            decluttered.nodes.iter().any(|n| !n.op_is::<Const>() && !n.op_is::<Dummy>()),
            "a shared constant over the memory budget must not be folded per consumer"
        );
        Ok(())
    }
}