consult-architecture

What problem does it solve?

Guide architecture-aware experiment design for Mamba2 state-space models, translating theoretical ideas into feasible training configurations and resource estimates.

Core Features & Use Cases

• Analyze compatibility of techniques with Mamba2 without relying on transformer-specific assumptions.
• Estimate VRAM and compute implications of architectural changes (e.g., d_state, context length, number of layers) within given hardware constraints.
• Connect architectural questions to concrete code references (models/mamba2.py, models/encoding.py, training/trainer.py) to ground recommendations.
• Use cases include evaluating new loss designs, input encodings, and experiment scope decisions to inform design and risk assessment.

Quick Start

Ask it to evaluate whether a proposed architectural change is feasible within available VRAM and how it would impact training stability.