Install in seconds
Install this skill
Copy the command and run it in your terminal. You can review the source before installing.
terminal
git clone https://github.com/Orchestra-Research/AI-Research-SKILLs

Works with Git. The repository opens in your current directory.

🐍
AI EngineeringTeX

Mamba Architecture

by Orchestra-Research

Implements Mamba state-space models achieving O(n) complexity for sequence modeling. Enables 5× faster inference, million-token sequences, and no KV cache. Ideal for streaming and long-context applications.

11.2K stars821 forksAdded 2026/07/19
aiai-researchclaudeclaude-codeclaude-skillscodexgeminigpt-5grpohuggingfacemachine-leanringmegatronskillsvllm

Documentation

README

Mamba - Selective State Space Models

Quick start

Mamba is a state-space model architecture achieving O(n) linear complexity for sequence modeling.

Installation:

# Install causal-conv1d (optional, for efficiency)
pip install causal-conv1d>=1.4.0

# Install Mamba
pip install mamba-ssm
# Or both together
pip install mamba-ssm[causal-conv1d]

Prerequisites: Linux, NVIDIA GPU, PyTorch 1.12+, CUDA 11.6+

Basic usage (Mamba block):

import torch
from mamba_ssm import Mamba

batch, length, dim = 2, 64, 16
x = torch.randn(batch, length, dim).to("cuda")

model = Mamba(
    d_model=dim,      # Model dimension
    d_state=16,       # SSM state dimension
    d_conv=4,         # Conv1d kernel size
    expand=2          # Expansion factor
).to("cuda")

y = model(x)  # O(n) complexity!
assert y.shape == x.shape

Common workflows

Workflow 1: Language model with Mamba-2

Complete LM with generation:

from mamba_ssm.models.mixer_seq_simple import MambaLMHeadModel
from mamba_ssm.models.config_mamba import MambaConfig
import torch

# Configure Mamba-2 LM
config = MambaConfig(
    d_model=1024,           # Hidden dimension
    n_layer=24,             # Number of layers
    vocab_size=50277,       # Vocabulary size
    ssm_cfg=dict(
        layer="Mamba2",     # Use Mamba-2
        d_state=128,        # Larger state for Mamba-2
        headdim=64,         # Head dimension
        ngroups=1           # Number of groups
    )
)

model = MambaLMHeadModel(config, device="cuda", dtype=torch.float16)

# Generate text
input_ids = torch.randint(0, 1000, (1, 20), device="cuda", dtype=torch.long)
output = model.generate(
    input_ids=input_ids,
    max_length=100,
    temperature=0.7,
    top_p=0.9
)

Workflow 2: Use pretrained Mamba models

Load from HuggingFace:

from transformers import AutoTokenizer
from mamba_ssm.models.mixer_seq_simple import MambaLMHeadModel

# Load pretrained model
model_name = "state-spaces/mamba-2.8b"
tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neox-20b")  # Use compatible tokenizer
model = MambaLMHeadModel.from_pretrained(model_name, device="cuda", dtype=torch.float16)

# Generate
prompt = "The future of AI is"
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to("cuda")
output_ids = model.generate(
    input_ids=input_ids,
    max_length=200,
    temperature=0.7,
    top_p=0.9,
    repetition_penalty=1.2
)
generated_text = tokenizer.decode(output_ids[0])
print(generated_text)

Available models:

  • state-spaces/mamba-130m
  • state-spaces/mamba-370m
  • state-spaces/mamba-790m
  • state-spaces/mamba-1.4b
  • state-spaces/mamba-2.8b

Workflow 3: Mamba-1 vs Mamba-2

Mamba-1 (smaller state):

from mamba_ssm import Mamba

model = Mamba(
    d_model=256,
    d_state=16,      # Smaller state dimension
    d_conv=4,
    expand=2
).to("cuda")

Mamba-2 (multi-head, larger state):

from mamba_ssm import Mamba2

model = Mamba2(
    d_model=256,
    d_state=128,     # Larger state dimension
    d_conv=4,
    expand=2,
    headdim=64,      # Head dimension for multi-head
    ngroups=1        # Parallel groups
).to("cuda")

Key differences:

  • State size: Mamba-1 (d_state=16) vs Mamba-2 (d_state=128)
  • Architecture: Mamba-2 has multi-head structure
  • Normalization: Mamba-2 uses RMSNorm
  • Distributed: Mamba-2 supports tensor parallelism

Workflow 4: Benchmark vs Transformers

Generation speed comparison:

# Benchmark Mamba
python benchmarks/benchmark_generation_mamba_simple.py \
  --model-name "state-spaces/mamba-2.8b" \
  --prompt "The future of machine learning is" \
  --topp 0.9 --temperature 0.7 --repetition-penalty 1.2

# Benchmark Transformer
python benchmarks/benchmark_generation_mamba_simple.py \
  --model-name "EleutherAI/pythia-2.8b" \
  --prompt "The future of machine learning is" \
  --topp 0.9 --temperature 0.7 --repetition-penalty 1.2

Expected results:

  • Mamba: 5× faster inference
  • Memory: No KV cache needed
  • Scaling: Linear with sequence length

When to use vs alternatives

Use Mamba when:

  • Need long sequences (100K+ tokens)
  • Want faster inference than Transformers
  • Memory-constrained (no KV cache)
  • Building streaming applications
  • Linear scaling important

Advantages:

  • O(n) complexity: Linear vs quadratic
  • 5× faster inference: No attention overhead
  • No KV cache: Lower memory usage
  • Million-token sequences: Hardware-efficient
  • Streaming: Constant memory per token

Use alternatives instead:

  • Transformers: Need best-in-class performance, have compute
  • RWKV: Want RNN+Transformer hybrid
  • RetNet: Need retention-based architecture
  • Hyena: Want convolution-based approach

Common issues

Issue: CUDA out of memory

Reduce batch size or use gradient checkpointing:

model = MambaLMHeadModel(config, device="cuda", dtype=torch.float16)
model.gradient_checkpointing_enable()  # Enable checkpointing

Issue: Slow installation

Install binary wheels (not source):

pip install mamba-ssm --no-build-isolation

Issue: Missing causal-conv1d

Install separately:

pip install causal-conv1d>=1.4.0

Issue: Model not loading from HuggingFace

Use MambaLMHeadModel.from_pretrained (not AutoModel):

from mamba_ssm.models.mixer_seq_simple import MambaLMHeadModel
model = MambaLMHeadModel.from_pretrained("state-spaces/mamba-2.8b")

Advanced topics

Selective SSM: See references/selective-ssm.md for mathematical formulation, state-space equations, and how selectivity enables O(n) complexity.

Mamba-2 architecture: See references/mamba2-details.md for multi-head structure, tensor parallelism, and distributed training setup.

Performance optimization: See references/performance.md for hardware-aware design, CUDA kernels, and memory efficiency techniques.

Hardware requirements

  • GPU: NVIDIA with CUDA 11.6+
  • VRAM:
    • 130M model: 2GB
    • 370M model: 4GB
    • 790M model: 8GB
    • 1.4B model: 14GB
    • 2.8B model: 28GB (FP16)
  • Inference: 5× faster than Transformers
  • Memory: No KV cache (lower than Transformers)

Performance (vs Transformers):

  • Speed: 5× faster inference
  • Memory: 50% less (no KV cache)
  • Scaling: Linear vs quadratic

Resources

More from Orchestra-Research

Other Claude Code skills by this author in the directory.

1w ago

LitGPT LLM Implementation

Implements and trains LLMs using LitGPT with 20+ architectures like Llama, Gemma, Phi. Use for clean model implementations, understanding architectures, or production fine-tuning with LoRA/QLoRA. Single-file, no abstraction layers.
AI Engineering
+1%11.2K821
🧠
1w ago

nanoGPT

A minimalist, educational GPT implementation in ~300 lines of code. Ideal for learning transformer architectures from scratch; supports training on Shakespeare (CPU) or reproducing GPT-2 on OpenWebText (multi-GPU).
AI Engineering
+1%11.2K821
🧠
1w ago

RWKV Architecture

A skill for using RWKV, the RNN-Transformer hybrid model that offers linear-time inference, infinite context, and constant memory. Covers installation, text generation, long context processing, fine-tuning, and comparison with Transformers.
AI Engineering
+1%11.2K821
🔥
1w ago

TorchTitan LLM Pretraining

Provides PyTorch-native distributed LLM pretraining using torchtitan with 4D parallelism (FSDP2, TP, PP, CP). Use when pretraining Llama 3.1, DeepSeek V3, or custom models at scale from 8 to 512+ GPUs with Float8, torch.compile, and distributed checkpointing.
AI Engineering
+1%11.2K821
1w ago

HuggingFace Tokenizers

Rust-based tokenization library for NLP that processes 1GB of text in under 20 seconds. Supports BPE, WordPiece, and Unigram algorithms, custom vocabulary training, and seamless transformers integration for research and production.
AI Engineering
+1%11.2K821
🧩
1w ago

SentencePiece

SentencePiece is a language-independent tokenizer that works on raw Unicode text without pre-tokenization. It supports BPE and Unigram algorithms for fast, deterministic vocabulary training, ideal for multilingual NLP models and CJK languages.
AI Engineering
+1%11.2K821