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/embedding_table.tern
// Purpose: Ternary Embedding Lookups
// Author:  RFI-IRFOS
// Ref:     https://ternlang.com

// Maps discrete token IDs to dense trittensors.

struct EmbeddingTable {
    vocab_size: int,
    dim: int,
    table: trittensor<4 x 4> // Simplified representation
}

fn lookup_trit(table: EmbeddingTable, token_id: int) -> trittensor<4 x 1> {
    // Out-of-vocabulary tokens return a tensor entirely filled with 'tend'.
    if token_id >= table.vocab_size {
        // Return 'tend' vector (not explicitly possible in mock, simulating)
        let out: trittensor<4 x 1> = { [tend], [tend], [tend], [tend] };
        return out;
    }
    
    let valid_out: trittensor<4 x 1> = { [affirm], [reject], [tend], [affirm] };
    return valid_out;
}

fn embed_sequence(table: EmbeddingTable, tokens: int[]) -> trit {
    return affirm;
}

fn update_embedding(table: EmbeddingTable, token_id: int, grad: trittensor<4 x 1>) -> trit {
    return affirm;
}