This is a hybrid Mamba-Transformer model based on the Llama 3.1 architecture, distilled from Llama 3.3 70B into a 8B parameter model using Qwerky's proprietary distillation method. The model uses MAMBA layers interleaved with attention layers for efficient sequence modeling. The results are a 8B parameter model comparable in quality to Llama's 3.1 8B but running at speeds as fast or faster than Llama's 3.2 3B model.

Requirements

CUDA-compatible GPU
Python 3.8+
PyTorch 2.0+
transformers, safetensors, mamba-ssm, causal-conv1d, flash-attn

Installation

bash

pip install transformers torch safetensors
pip install flash-attn mamba-ssm causal-conv1d --no-build-isolation

Usage - Transformers

python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("QwerkyAI/QRe-Llama-3-8B-Instruct-QDistill")
model = AutoModelForCausalLM.from_pretrained(
    "QwerkyAI/QRe-Llama-3-8B-Instruct-QDistill",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

inputs = tokenizer("Hello, how are you?", return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0]))

Usage - vLLM

bash

pip install vllm qwerky-vllm-models
vllm serve QwerkyAI/QRe-Llama-3-8B-Instruct-QDistill

Model Files

config.json - Model configuration with auto_map
modeling_qwerky_llama_mamba_hybrid.py - Custom modeling class
configuration_qwerky_llama_mamba_hybrid.py - Custom configuration class
model.safetensors - Model weights

License

See LICENSE file for details.