neura-decoder 0.7.3

Assembling the decoder only models a gguf weighs onto the graph
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

use neura_graph::{Graph, Shape, Value};

pub fn rms(graph: &Graph, value: Value, weight: Value, floor: Value) -> Value {
    let squares = graph.mul(value, value);
    let mean = graph.mean_axis(squares, 3);
    let scale = graph.recip(graph.sqrt(graph.add(mean, floor)));
    graph.mul(graph.mul(value, scale), weight)
}

pub fn l2(graph: &Graph, value: Value, floor: Value) -> Value {
    let squares = graph.mul(value, value);
    let total = graph.sum_axis(squares, 3);
    let scale = graph.recip(graph.sqrt(graph.add(total, floor)));
    graph.mul(value, scale)
}

pub fn softplus(graph: &Graph, value: Value) -> Value {
    let threshold = graph.fill(Shape::scalar(), 20.0);
    let raised = graph.log(graph.add(graph.fill(Shape::scalar(), 1.0), graph.exp(value)));
    graph.select(graph.greater(value, threshold), value, raised)
}

pub fn flatten(graph: &Graph, value: Value) -> Value {
    let elements = graph.shape(value).elements();
    graph.reshape(value, Shape::of([1, 1, 1, elements]))
}