ternlang-core 0.3.3

Compiler and VM for Ternlang — balanced ternary language with affirm/tend/reject trit semantics, @sparseskip codegen, and BET bytecode execution.
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

// Module:  stdlib/nn/multihead_attn.tern
// Purpose: Ternary Multi-Head Attention
// Author:  RFI-IRFOS
// Ref:     https://ternlang.com

// Multi-head attention using trit tensors. Naturally sparse, meaning 
// attention maps bypass irrelevant tokens via '@sparseskip'.

struct MultiHeadAttn {
    heads: int,
    q_proj: trittensor<4 x 4>,
    k_proj: trittensor<4 x 4>,
    v_proj: trittensor<4 x 4>
}

fn split_heads(tensor: trittensor<4 x 4>) -> trittensor<4 x 4> {
    // Dummy implementation for shape manipulation
    return tensor;
}

fn concat_heads(tensor: trittensor<4 x 4>) -> trittensor<4 x 4> {
    // Dummy implementation
    return tensor;
}

fn scaled_dot_trit(q: trittensor<4 x 4>, k: trittensor<4 x 4>) -> trittensor<4 x 4> {
    // In ternary, 'scaling' by sqrt(d_k) isn't necessary because states 
    // are bounded to (-1, 0, 1). We just use sparse matmul.
    @sparseskip
    let score: trittensor<4 x 4> = q * k;
    return score;
}