BitMoE

1 bit Mixture of Experts utilizing BitNet ++ Mixture of Experts. Also will add distribution amongst GPUs.

install

$ pip3 install bitmoe

usage

import torch
from bitmoe.main import BitMoE

# Set the parameters
dim = 10  # Dimension of the input
hidden_dim = 20  # Dimension of the hidden layer
output_dim = 30  # Dimension of the output
num_experts = 5  # Number of experts in the BitMoE model

# Create the model
model = BitMoE(dim, hidden_dim, output_dim, num_experts)

# Create random inputs
batch_size = 32  # Number of samples in a batch
sequence_length = 100  # Length of the input sequence
x = torch.randn(batch_size, sequence_length, dim)  # Random input tensor

# Forward pass
output = model(x)  # Perform forward pass using the model

# Print the output shape
print(output)  # Print the output tensor
print(output.shape)  # Print the shape of the output tensor

License

MIT

Todo

• Implement better gating mechanisms
• Implement better routing algorithm
• Implement better BitFeedForward
• Implement