Massive-Activations-HLA

August 13, 2026 · View on GitHub

Official code and reproducibility toolkit for the paper:

Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateaus

This project studies massive activations (MAs) in hybrid linear attention (HLA) language models. We find that sparse full-attention layers induce architecture-aligned activation events: sharp pre-attention spikes (PAS) and, when full attention becomes denser, persistent inter-spike plateaus (ISP).

Paper · Models on Hugging Face · Quickstart · Reproduction guide · Installation

A synchronized Hugging Face model-card source is maintained at docs/HUGGINGFACE_MODEL_CARD.md so that model compatibility claims stay aligned with this README.

What this repository provides

This repository contains clean, paper-facing code for reproducing the main analysis artifacts:

Artifact	Script	What it reproduces
MA morphology	`scripts/run_morphology.py`	Layerwise token activation curves across models, domains, and hybrid ratios
PAS/ISP metrics	`scripts/run_pas_isp_metrics.py`	Sink-spike alignment and inter-spike retention
Lifecycle atlas	`scripts/run_lifecycle_atlas.py`	Residual / attention / MLP / block-output decomposition of systematic outliers

Released controlled-pretraining checkpoints are hosted separately on Hugging Face:

startlux-models/Massive-Activations-HLA

The repository does not include model weights, datasets, generated figures, or private paths.

Key findings

The paper asks how massive activations behave once full attention is no longer present at every layer. The main findings are:

MAs are architecture-aligned in HLA LLMs. Their largest spikes occur immediately before full-attention layers rather than uniformly across depth.
PAS are robust across models and inputs. The pre-attention spike pattern appears in controlled GDN models, M-A-P checkpoints, and several modern hybrid model families.
ISP reveal a denser-attention transition. As full attention becomes more frequent, isolated PAS are connected by sustained activation plateaus.
Lifecycle atlases expose where outliers are written and canceled. The provided atlas visualizations decompose each layer into residual input, sequence-mixing update, MLP update, and block output.

Overview

Hybrid linear attention models are often introduced as efficient alternatives to full-attention Transformers, but their activation dynamics are not simply Transformer-like behavior with cheaper sequence mixing. In HLA LLMs, full attention layers act as structural anchor points around which large hidden-state outliers emerge and evolve.

The paper focuses on two characteristic MA morphologies:

Pre-attention spikes (PAS): sharp activation spikes immediately before full-attention layers. PAS show that the largest activation events are aligned with the placement of full attention rather than appearing at arbitrary depth.
Inter-spike plateaus (ISP): sustained activation plateaus connecting adjacent PAS. ISP appear as full attention becomes denser, revealing a gradual transition from isolated spikes to persistent, full-attention-like MA dynamics.

The toolkit is designed for three use cases:

Reproduce the released-checkpoint analysis pipeline for the public prompts and checkpoints described below.
Inspect new HLA models by tracing first-token or arbitrary-token MA trajectories.
Compare architectures using the same PAS/ISP metrics and visualization style across controlled GDN models, M-A-P checkpoints, and optional modern hybrid models.

Installation

Create the tested released-GDN environment:

conda create -n ma-hla python=3.12 -y
conda activate ma-hla
python -m pip install --upgrade pip
bash scripts/install_released_gdn_cu126.sh

For HuggingFace datasets:

pip install -r requirements/datasets.txt

The installation script uses the exact tested PyTorch wheel:

torch==2.7.1+cu126

It installs the official causal-conv1d and flash-attn PyTorch 2.7/CUDA 12 wheels, verifies their checksums, and then installs FLA.

The released baseline and no-output-gate GDN checkpoints are tested with flash-linear-attention==0.5.2 (upstream tag v0.5.2, commit 9c8e42e762fce087c27b673af4922795d9edb85e). For the exact A800/CUDA 12.6 environment, follow INSTALL.md.

Full-attention-gate ablation scope

The two gdn-gatedfa-* checkpoints use a post-SDPA, head-specific sigmoid output gate inspired by the G1 design in Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free (official code). Our Gated DeltaNet integration of that gate is not distributed in this repository. Those two checkpoints are therefore released as weights-only research artifacts and are not part of the public, from-scratch quickstart. The loader fails with an explicit compatibility message instead of silently loading them incorrectly.

All baseline and gdn-nooutgate-* checkpoints are covered by the public FLA environment and are reproducible from the published code and weights.

Quickstart: plot a PAS morphology curve

The following command loads the released 340M GDN checkpoint from HuggingFace and plots the first-token MA trajectory for the running example Summer is warm. Winter is cold.

PYTHONPATH=src python scripts/run_morphology.py \
  --models configs/models/released_gdn_models.yaml \
  --model-names gdn_340m_pas_layer12 \
  --prompts configs/datasets/prompts.yaml \
  --token-index 0 \
  --max-length 32 \
  --output outputs/quickstart/morphology

Outputs are written to:

outputs/quickstart/morphology/gdn_340m_pas_layer12/summer/

For a smoke test that does not download datasets, use:

--datasets configs/datasets/five_domains_tiny.yaml --limit 1

This is an offline dataset-input test, not a fully offline model test: the checkpoint is still downloaded from Hugging Face unless it is already cached or you configure a local model path.

Compute PAS/ISP metrics

First capture an ISP model:

PYTHONPATH=src python scripts/run_morphology.py \
  --models configs/models/released_gdn_models.yaml \
  --model-names gdn_340m_isp_3to1 \
  --prompts configs/datasets/prompts.yaml \
  --token-index 0 \
  --max-length 32 \
  --output outputs/quickstart/isp_morphology

Then compute the metrics:

PYTHONPATH=src python scripts/run_pas_isp_metrics.py \
  --models configs/models/released_gdn_models.yaml \
  --model-names gdn_340m_isp_3to1 \
  --trace-root outputs/quickstart/isp_morphology \
  --prompt-names summer \
  --trace-token-index 0 \
  --sink-token 0 \
  --output outputs/quickstart/metrics.csv

The metrics script reports:

sink_spike_alignment: fraction of full-attention events whose sink token reaches its maximum activation at the corresponding pre-attention layer;
inter_spike_retention: fraction of the adjacent PAS reference level retained across intervening linear-attention layers.

When full-attention maps are available, pass --attention-root to extract sink tokens directly from attention:

<attention-root>/<model>/<prompt>/layer_<L>.npy

where L is the 1-based full-attention layer index and each array has shape [heads, query, source] or [query, source].

Plot a lifecycle atlas

PYTHONPATH=src python scripts/run_lifecycle_atlas.py \
  --models configs/models/released_gdn_models.yaml \
  --model-names gdn_340m_pas_layer12 \
  --prompts configs/datasets/prompts.yaml \
  --token-index 0 \
  --max-length 32 \
  --output outputs/quickstart/lifecycle_atlas

The atlas captures and visualizes the signed contribution of:

residual input;
attention / linear-attention update;
MLP update;
block output.

Supported model registries

Registry	Purpose
`configs/models/released_gdn_models.yaml`	Released controlled GDN checkpoints from HuggingFace
`configs/models/map_models.yaml`	Public M-A-P / controlled architecture checkpoints
`configs/models/modern_hybrid_models.yaml`	Optional modern pretrained hybrid models

The released GDN registry points to:

startlux-models/Massive-Activations-HLA

with one checkpoint per subfolder. If you use a local mirror, copy the config and replace hf_id / subfolder with local path entries.

Dataset inputs

The paper evaluates inputs from five regimes: general prose, scientific text, math reasoning, code, and multilingual text. This repository provides:

configs/datasets/prompts.yaml: the running Summer example;
configs/datasets/five_domains.yaml: HuggingFace dataset loaders;
configs/datasets/five_domains_tiny.yaml: offline smoke-test prompts.

The repository does not publish the exact sampled JSONL files used for every paper panel. five_domains.yaml reproduces the public loading protocol, while exact input-level reproduction requires the corresponding sampled JSONL files.

Repository layout

configs/
  datasets/     # prompt and dataset configs
  local/        # ignored local path overrides
  models/       # model registries
docs/
  reproduction.md
examples/
  quickstart_released_gdn.sh
  quickstart_summer_map.sh
scripts/
  run_lifecycle_atlas.py
  run_morphology.py
  run_pas_isp_metrics.py
src/massive_activations_hla/
  analysis/     # metrics and trace loading
  capture/      # model loading and forward hooks
  models/       # registry and released-GDN adapter
  plotting/     # paper-style visualizations
tests/

Notes and limitations

Large modern hybrid models may require substantial GPU memory and model-specific environments.
Baseline and no-output-gate GDN checkpoints use public FLA v0.5.2. The full-attention-gate ablation is weights-only as documented above.
five_domains_tiny.yaml is only a smoke-test input set; it is not a substitute for the full paper evaluation.
This repository is organized for analysis and visualization, not for training new models from scratch.

Licenses

The analysis code in this repository is MIT licensed. Released model weights are Apache-2.0 licensed in the Hugging Face repository. Flash Linear Attention and all datasets retain their own licenses and terms.

Citation

If you use this repository, please cite the paper:

@article{su2026massive,
  title={Massive Activations in Hybrid Linear Attention Large Language Models:
         Pre-Attention Spikes and Inter-Spike Plateaus},
  author={Su, Zunhai and Sun, Bohan and Zhuang, Xialie and Zhang, Shuibai and
          Xiao, He and Xiong, Jing and Zhang, Hengyuan and Zhou, Zhongzhu and
          Zhang, Tiantian and Wong, Ngai and Kuo, Chuan-Wei},
  journal={arXiv preprint arXiv:2608.12149},
  year={2026}
}