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

// Module:  stdlib/models/diffusion_analog.tern
// Purpose: Discrete Ternary Diffusion Model
// Author:  RFI-IRFOS
// Ref:     https://ternlang.com

// Forward diffusion gradually corrupts 'affirm'/'reject' states into 'tend'.
// Reverse diffusion denoises 'tend' back into structure.

struct TritDiffusion {
    steps: int,
    w: trittensor<4 x 4>
}

fn noise_schedule_trit(step: int, max_steps: int) -> float {
    // Returns probability of corruption
    return 0.1;
}

fn denoise_step_trit(model: TritDiffusion, noisy: trittensor<4 x 1>, step: int) -> trittensor<4 x 1> {
    @sparseskip
    let less_noisy: trittensor<4 x 1> = model.w * noisy;
    return less_noisy;
}

fn sample_trit(model: TritDiffusion) -> trittensor<4 x 1> {
    // Start with pure noise (all 'tend')
    let current: trittensor<4 x 1> = { [tend], [tend], [tend], [tend] };
    
    // Iterative denoise ...
    return current;
}