ternlang-ml 1.5.0

Ternary ML inference kernels — TritFloat (confidence-native ternary float), TritFloatTensor, sparse_matmul (@sparseskip), TritMatrix, deliberation engine, action gate, and MLP for the Ternlang ecosystem.
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

import torch
import ternlang_ml_C

class TernaryTensor(torch.Tensor):
    """
    Proprietary TernaryTensor utilizing torch_dispatch subclassing mechanism.
    Wraps standard PyTorch tensors and routes matrix multiplication calls 
    through highly optimized BET VM logic.
    """
    
    @classmethod
    def __torch_dispatch__(cls, func, types, args=(), kwargs=None):
        if kwargs is None:
            kwargs = {}
            
        # Intercept matrix multiplication
        if func.__name__ == 'matmul':
            # Route to BET VM logic (C++ extension)
            # This bypasses multiplication when a {0} weight is encountered,
            # delivering massive throughput yields to standard Python developers.
            return ternlang_ml_C.ternary_mm(args[0], args[1])
            
        # Fallback to standard PyTorch implementation
        return super().__torch_dispatch__(func, types, args, kwargs)