[NeurIPS 2025 Oral] Official Code for Exploring Diffusion Transformer Designs via Grafting

Run IBM Granite 4.0 locally on Raspberry Pi 5 with Ollama.This is a privacy-first AI. Your data never leaves your device because it runs 100% locally. There are no cloud uploads and no third-party tracking.

mech-interp suite for Granite4 models that use Mamba-2 architecture

Early baby steps towards a long-term vision regarding Mamba-2's state interpretability.

A simple, minimalistic, and explainable code implementation of of Nemotron 3 Nano in JAX

Mechanistic interpretability for State-Space Models: SAEs, feature visualization, and a Hub registry for Mamba/Mamba-2.

Systematic study of LoRA fine-tuning strategies for IBM Granite 4.0-H-Micro (Mamba-2 + Transformer hybrid). Demonstrates the impact of architecture-aware target selection and SSM core parameter co-training, including analysis of PEFT serialization behavior. Reports up to 37% relative improvement over LoRA-only baselines.

A simple, minimalistic, and explainable JAX implementation of Mamba 2 & Mamba 3

Integrated mechanistic interpretability + sparse autoencoder framework for Hybrid SSM-Attention models (Mamba-2, Hymba, RWKV-7). v0.1.2 alpha: real forward-pass intervention + mean-ablation patching shipped, CPU smoke; GPU/real adapters in v0.2.