Create b158.py
Browse files
b158.py
ADDED
|
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import torch.nn as nn
|
| 3 |
+
import torchao
|
| 4 |
+
|
| 5 |
+
def bitnet_b158_quantize(tensor):
|
| 6 |
+
"""
|
| 7 |
+
Konvertiert ein Gewicht-Tensor in das ternäre 1.58-Bit Format {-1, 0, +1}
|
| 8 |
+
inklusive der notwendigen Per-Channel Skalierung.
|
| 9 |
+
"""
|
| 10 |
+
# 1. Berechne den durchschnittlichen Absolutwert pro Kanal (Zeile)
|
| 11 |
+
scale = tensor.abs().mean(dim=-1, keepdim=True).clamp(min=1e-5)
|
| 12 |
+
|
| 13 |
+
# 2. Skaliere den Tensor und runde auf die nächste ganze Zahl
|
| 14 |
+
quantized = torch.round(tensor / scale)
|
| 15 |
+
|
| 16 |
+
# 3. Zwinge die Werte strikt in den Bereich von -1 bis +1
|
| 17 |
+
quantized = torch.clamp(quantized, min=-1.0, max=1.0)
|
| 18 |
+
|
| 19 |
+
return quantized, scale
|
| 20 |
+
|
| 21 |
+
class BitLinear158(nn.Module):
|
| 22 |
+
"""
|
| 23 |
+
Ein Ersatz für nn.Linear, der die 1.58-Bit Ternary-Inferenz ausführt.
|
| 24 |
+
"""
|
| 25 |
+
def __init__(self, in_features, out_features, bias=False):
|
| 26 |
+
super().__init__()
|
| 27 |
+
self.in_features = in_features
|
| 28 |
+
self.out_features = out_features
|
| 29 |
+
self.register_buffer("weight_158", torch.zeros((out_features, in_features), dtype=torch.int8))
|
| 30 |
+
self.register_buffer("scale", torch.zeros((out_features, 1), dtype=torch.bfloat16))
|
| 31 |
+
|
| 32 |
+
@torch.no_grad()
|
| 33 |
+
def from_float(self, float_layer):
|
| 34 |
+
# Transformiere die originalen Gewichte
|
| 35 |
+
q_w, scale = bitnet_b158_quantize(float_layer.weight.data)
|
| 36 |
+
self.weight_158.copy_(q_w.to(torch.int8))
|
| 37 |
+
self.scale.copy_(scale.to(torch.bfloat16))
|
| 38 |
+
return self
|
| 39 |
+
|
| 40 |
+
def forward(self, x):
|
| 41 |
+
# Die Magie: x wird mit den Integer-Gewichten (-1, 0, 1) verarbeitet
|
| 42 |
+
# Auf der Hardwarebene entspricht dies reinen Additionen/Subtraktionen
|
| 43 |
+
out = nn.functional.linear(x, self.weight_158.to(x.dtype))
|
| 44 |
+
return out * self.scale
|