mamba2-8b-3t-4k-hf

This repository contains a Hugging Face Transformers-compatible conversion of nvidia/mamba2-8b-3t-4k.

Notes

• Source checkpoint format: Megatron-LM
• Target format: Hugging Face Transformers
• Loaded via Mamba2ForCausalLM
• Original SentencePiece tokenizer file is preserved in this repo
• Tokenizer is a practical-compatibility T5Tokenizer wrapper rather than a byte-for-byte Megatron GPTSentencePiece clone

Loading

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "ib-ssm/mamba2-8b-3t-4k-hf"

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    torch_dtype="auto",
    device_map="auto",
)