Superlinear Multi-Step Attention — a subquadratic attention mechanism that preserves random context access (structural non-exclusion) for extremely long sequences.

This is an experimental release demonstrating the Superlinear attention architecture integrated into a modified nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 hybrid model.

WARNING (Security): This model requires trust_remote_code=True, which executes Python code from this repository. Review the code before running in sensitive environments.

Model Description

Superlinear attention reformulates causal self-attention as a multi-step search problem:

Accumulation — Efficiently processes the sequence and produces per-position representatives (via Mamba-2 layers in the hybrid architecture).
Span Search — Scores a sublinear number of candidate spans using learned routing, then selects top-k spans per query.
Span Attention — Computes standard token-level attention within the selected contiguous spans.
Combination — Produces outputs using softmax-weighted gating over span attention outputs.

In the baseline N=2 implementation, both span-search and span-attention scale as O(L^(3/2)), enabling practical inference at multi-million-token context lengths where dense attention becomes prohibitive.

Key property: Random context access (structural non-exclusion) — any eligible token position can be selected by the content-dependent routing mechanism; no fixed sparsity pattern permanently excludes positions.

Quickstart

python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained(
    "concavity-ai/superlinear-exp-v0.1",
    trust_remote_code=True
)

model = AutoModelForCausalLM.from_pretrained(
    "concavity-ai/superlinear-exp-v0.1",
    torch_dtype=torch.float16,
    device_map="cuda",
    trust_remote_code=True,
)

messages = [{"role": "user", "content": "Explain the Transformer architecture."}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to("cuda")

output = model.generate(inputs, max_new_tokens=1000, do_sample=True, temperature=0.1, top_p=0.99)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Dependencies

This model uses custom Python code (trust_remote_code=True) and CUDA extensions.

Recommended: follow the Superlinear repo install

The simplest supported path is to use the installation flow from the Superlinear repo (it pins a known-good CUDA toolchain and builds mamba-ssm[causal-conv1d] from source to avoid wheel/ABI mismatches):

https://github.com/concavity-ai/superlinear#installation

Copy/paste one-liner (from the Superlinear repo root):

bash

conda env create -f environment.yml \
  && conda run -n superlinear pip install torch --index-url https://download.pytorch.org/whl/cu128 \
  && conda run -n superlinear pip install -e ".[server,model]" \
  && conda run -n superlinear bash -lc 'CUDA_HOME="$CONDA_PREFIX" pip install "mamba-ssm[causal-conv1d]" --no-build-isolation --no-cache-dir --no-binary :all:'

Optional: pip-only (if `mamba-ssm` already works in your env)

If python -c "import mamba_ssm, causal_conv1d" already succeeds in the environment you’ll run inference in, you already have a working PyTorch/CUDA pairing for the extension in that environment — you should not need to reinstall PyTorch.

Install the remaining Python deps + Superlinear:

bash

pip install -U "transformers<5" accelerate safetensors
pip install -U vllm triton

# Superlinear kernels (span-attention)
pip install -U git+https://github.com/concavity-ai/superlinear.git

Building `mamba-ssm` from source (only if needed)

If you must build mamba-ssm[causal-conv1d] yourself, you need a CUDA toolkit with nvcc and CUDA_HOME pointing at it (example: /usr/local/cuda):

bash

CUDA_HOME=/usr/local/cuda \
  pip install -U "mamba-ssm[causal-conv1d]" \
  --no-build-isolation --no-cache-dir --no-binary :all:

Recommended Inference Settings

For long-context inference with the Superlinear attention mechanism, use the following configuration:

python

model = AutoModelForCausalLM.from_pretrained(
    "concavity-ai/superlinear-exp-v0.1",
    # Attention implementation
    _attn_implementation='block-span-gqa',
    decode_kernel='staged-gqa',
    
    # Performance optimizations
    enable_cuda_graph=True,
    enable_shared_fused_moe=True,
    
    # Superlinear attention hyperparameters
    span_attention_sw_index=65,           # Local window boundary index
    span_attention_num_spans=3,           # Top-k spans per query
    span_attention_backward_factor=3,     # Backward span extent multiplier
    span_attention_forward_factor=1,      # Forward span extent multiplier
    span_attention_search_power=0.55,     # Search exponent (controls anchor budget)
    span_attention_span_power=0.55,       # Span exponent (controls span scale)
    
    torch_dtype=torch.float16,
    device_map="cuda",
    trust_remote_code=True,
)

Hyperparameter Notes

Parameter	Description	Typical Value
`span_attention_num_spans`	Number of routed spans selected per query (top-k)	2 or 3
`span_attention_backward_factor`	Backward extent of each span relative to base scale	2–4
`span_attention_forward_factor`	Forward extent of each span relative to base scale	0–2
`span_attention_search_power`	Exponent controlling the number of candidate anchors	0.5–0.667
`span_attention_span_power`	Exponent controlling span length scaling	0.5–0.667

Sliding window length from span_attention_sw_index: Internally, the kernels compute the sliding-window length as:

window_len = floor((sw_index + 1) ** (1 / search_power)) - 1

We parameterize the local sliding window using sw_index (a stride/stripe index) rather than specifying window_len directly. This keeps the sliding-window boundary aligned with the same index space used by span search, so span-search begins immediately after the sliding-window region and avoids gaps between local attention and routed spans.

Example: with search_power=0.55 and sw_index=65,

window_len = floor(66 ** (1 / 0.55)) - 1 = 2032

Hardware Requirements

GPU: NVIDIA GPU with sufficient VRAM (tested on B200 180GB)
KV Cache: ~6GB per million tokens (model-dependent)
Precision: FP16 recommended

Measured Throughput (Single B200, Batch=1)

Context Length	Prefill (tok/s)	Decode (tok/s)
1M tokens	~20,202	~109
10M tokens	~5,576	~76

Your results may vary depending on hardware, batch size, and configuration.

Intended Use

This is an architecture-and-systems feasibility study release. It demonstrates that:

The Superlinear attention mechanism is structurally random-context-access-preserving
The architecture achieves asymptotically subquadratic attention complexity
The resulting irregular span pattern can be implemented with practical performance at very long context lengths

Limitations

Not a comprehensive quality study: We do not present extensive ablations or claim state-of-the-art accuracy on benchmarks.
Limited evaluation: Initial validation focused on NIAH (Needle In A Haystack) retrieval task and throughput measurements.
Experimental: This release is intended for research and experimentation, not production use.
Memory: Full KV cache must be retained for random context access; memory usage scales with context length.

What's in This Repository

├── config.json                     # Model configuration
├── generation_config.json          # Default generation settings
├── tokenizer.json                  # Tokenizer
├── tokenizer_config.json
├── special_tokens_map.json
├── chat_template.jinja             # Chat template
├── configuration_superlinear_exp.py  # Custom config class
├── modeling_superlinear_exp.py     # Custom model implementation
├── moe.py                          # MoE components
├── model-*.safetensors             # Model weights (16 shards)
├── model.safetensors.index.json    # Weight index
├── LICENSE                         # NVIDIA Open Model License
├── NOTICE                          # Required attribution
└── README.md                       # This file

License

Model Weights

This model is a derivative of nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 and is distributed under the NVIDIA Open Model License Agreement.

See LICENSE for the full license text.

Use of this model must be consistent with NVIDIA's Trustworthy AI terms.

Code

The modeling code in this repository is provided for loading and running the model. For the broader Superlinear project codebase, see github.com/concavity-ai/superlinear (Apache-2.0).

Attribution

Upstream Model:

NVIDIA Nemotron-3-Nano-30B-A3B (nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16)

Paper:

bibtex

@article{huang2026superlinear,
  title={Superlinear Multi-Step Attention},
  author={Huang, Yufeng},
  journal={arXiv preprint arXiv:2601.18401},
  year={2026}
}

Patent Notice

Patent applications have been filed related to aspects of the methods described in this work.

Contact

Author: Yufeng Huang
Email: yufeng@concavity.ai
Organization: Concavity AI

concavity-ai/superlinear-exp-v0.1

Superlinear-Exp-v0.1

Model Description

Quickstart

Dependencies

Recommended: follow the Superlinear repo install

Optional: pip-only (if mamba-ssm already works in your env)

Building mamba-ssm from source (only if needed)

Recommended Inference Settings

Hyperparameter Notes

Hardware Requirements

Measured Throughput (Single B200, Batch=1)

Intended Use

Limitations

What's in This Repository

License

Model Weights

Code

Attribution

Patent Notice

Contact

concavity-ai/
superlinear-exp-v0.1

Optional: pip-only (if `mamba-ssm` already works in your env)

Building `mamba-ssm` from source (only if needed)