How state-spaces/mamba works
Mamba is a PyTorch state-space sequence-modeling library, exposing Mamba-1, Mamba-2/SSD, and Mamba-3 blocks plus a complete causal language-model example. The principal workflow is token IDs entering generation, passing through the language-model head, embedding/backbone blocks, and a selected SSM implementation, then returning logits or generated tokens. Lower layers provide selective-scan, SSD, and GPU kernels, with inference-state caching for decoding. The graph reflects README capabilities and sampled call sites; some unsampled variant wiring is shown only where documented.
Main entry points: mamba3.py Mamba-3 Block · generation.py Generation API · block.py Sequence Block · mixer_seq_simple.py LM Head
Components 18
Public API
- Generation API
mamba_ssm/utils/generation.py
Model Stack
- LM Head
mamba_ssm/models/mixer_seq_simple.py - Mixer Backbone
mamba_ssm/models/mixer_seq_simple.py - Token Embedding
mamba_ssm/models/mixer_seq_simple.py - Sequence Block
mamba_ssm/modules/block.py
SSM Algorithms
- Mamba-1 Block
mamba_ssm/modules/mamba_simple.py - Mamba-2 Block
mamba_ssm/modules/mamba2.py - Mamba-3 Block
mamba_ssm/modules/mamba3.py - Selective SSM
mamba_ssm/ops/selective_scan_interface.py
GPU Kernels
- SSD Operations
mamba_ssm/ops/triton/ssd_combined.py - CUDA Scan
csrc/selective_scan/selective_scan.cpp - Mamba-3 Kernels
mamba_ssm/ops/triton/mamba3/mamba3_siso_combined.py - Mamba-3 Tile Kernels
mamba_ssm/ops/tilelang/mamba3/mamba3_mimo.py - Mamba-3 Step
mamba_ssm/ops/cute/mamba3/mamba3_step_fn.py
Inference
- Inference State
mamba_ssm/utils/generation.py
Other components
- Model Caller
- Logits
- Generated Tokens
Connections 18
- Model Caller → Generation API: calls
- Generation API → LM Head: invokes model
- LM Head → Mixer Backbone: runs backbone
- Mixer Backbone → Token Embedding: embeds tokens
- Mixer Backbone → Sequence Block: runs blocks
- Sequence Block → Mamba-1 Block: selects mixer
- Sequence Block → Mamba-2 Block: selects mixer
- Sequence Block → Mamba-3 Block: selects mixer
- Mamba-1 Block → Selective SSM: calls scan
- Mamba-2 Block → SSD Operations: runs SSD
- Selective SSM → CUDA Scan: dispatches CUDA
- Mamba-3 Block → Mamba-3 Kernels: dispatches kernels
- Mamba-3 Block → Mamba-3 Tile Kernels: dispatches kernels
- Mamba-3 Block → Mamba-3 Step: runs step
- Mamba-3 Block → Inference State: updates state
- LM Head → Logits: returns logits
- Generation API → Inference State: manages cache
- Generation API → Generated Tokens: samples tokens