Awesome

MoE Mamba

Implementation of MoE Mamba from the paper: "MoE-Mamba: Efficient Selective State Space Models with Mixture of Experts" in Pytorch and Zeta. The SwitchMoE architecture is from the Switch Transformer paper. And, I still need help with it. If you want to help please join the Agora discord and server and help in the MoE Mamba channel.

PAPER LINK

Install

pip install moe-mamba

Usage

MoEMambaBlock

import torch 
from moe_mamba import MoEMambaBlock

x = torch.randn(1, 10, 512)
model = MoEMambaBlock(
    dim=512,
    depth=6,
    d_state=128,
    expand=4,
    num_experts=4,
)
out = model(x)
print(out)

MoEMamba

import torch 
from moe_mamba.model import MoEMamba 


# Create a tensor of shape (1, 1024, 512)
x = torch.randint(0, 10000, (1, 512))

# Create a MoEMamba model
model = MoEMamba(
    num_tokens=10000,
    dim=512,
    depth=1,
    d_state=512,
    causal=True,
    shared_qk=True,
    exact_window_size=True,
    dim_head=64,
    m_expand=4,
    num_experts=4,
)

# Forward pass
out = model(x)

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

Code Quality 🧹

• make style to format the code
• make check_code_quality to check code quality (PEP8 basically)
• black .
• ruff . --fix

Citation

@misc{pióro2024moemamba,
    title={MoE-Mamba: Efficient Selective State Space Models with Mixture of Experts}, 
    author={Maciej Pióro and Kamil Ciebiera and Krystian Król and Jan Ludziejewski and Sebastian Jaszczur},
    year={2024},
    eprint={2401.04081},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

License

MIT