Build state-space LLM architectures

Build and deploy Mamba state-space models with O(n) linear complexity

Long sequences (100K+ tokens), streaming inference, or memory-constrained deployments beating Transformer quadratic scaling.

• · model config (d_model, n_layer, d_state)
• · training data or checkpoint
• · inference prompt

• · trained model
• · generated sequence
• · memory/latency benchmarks

• · mamba-ssm
• · torch
• · causal-conv1d
• · transformers

• · Linux/NVIDIA GPU
• · CUDA 11.6+
• · PyTorch 1.12+

• · installation stalls on source build
• · causal-conv1d missing → 5× slower inference

• · 5× faster inference than Transformers
• · no KV cache → O(1) memory per token
• · million-token sequences verified