Prism is a training-free method to accelerate long-context LLM pre-filling. It addresses the "blind spot" in standard mean pooling caused by Rotary Positional Embeddings (RoPE) by disentangling attention into high-frequency and low-frequency bands.

Key Features:

• Dual-Band Importance Estimation: Separates semantic (low-freq) and positional (high-freq) signals.
• Energy-Based Calibration: Restores attenuated signals automatically.
• Speed: Up to 5.1× speedup on 128K context with negligible accuracy loss.
• Implementation: purely block-level operations with custom Triton kernels.

• prism/
  • prism.py: Core implementation of Prism.
  • kernels/: Custom Triton kernels for efficient block importance estimation with Prism and block-sparse attention.
• baselines: Baseline implementations (e.g., MInference, FlexPrefill, Xattention).
• eval/: Evaluation harnesses.
• data/: Example data for demonstration.
• scripts/: Shell scripts to reproduce the experiments in the paper.

# For core Prism implementation only
uv pip install -e .

# For lm-eval evals
uv pip install -e "eval/lm-evaluation-harness["hf", "longbench", "ruler"]"

# For lmms-eval evals
uv pip install -e "eval/lmms-eval["qwen", "metrics"]"

# For baselines
# FlashAttention
uv pip install flash_attn --no-build-isolation
# Minference
uv pip install minference
# XAttention
git clone git@github.com:mit-han-lab/Block-Sparse-Attention.git
cd Block-Sparse-Attention && uv pip install -e .

A simple example Prism using Qwen3-0.6B with a RULER example.

python -m prism.prism

Alternatively, patch your model with Prism for broader usage:

from prism import prism_attention_forward
from prism.utils.patch import apply_patch
from transformers import AutoModelForCausalLM

apply_patch(
    forward_fn=prism_attention_forward,
    model_id="Qwen/Qwen3-8B",
)

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B")

To reproduce the evaluation results, please refer to the scripts in the scripts directory.

# Example: Running LongBench Evaluation on Qwen3-8B
bash scripts/longbench.sh

Note: For RULER evaluation, we use Qwen3 with YaRN extrapolation, consistent with the official implementation. Please ensure your MODEL_ID points to a model path containing the modified config.json required for long-context processing.

If you find this work helpful, please consider citing our paper as following:

@misc{wang2026prismspectralawareblocksparseattention,
      title={Prism: Spectral-Aware Block-Sparse Attention}, 
      author={Xinghao Wang and Pengyu Wang and Xiaoran Liu and Fangxu Liu and Jason Chu and Kai Song and Xipeng Qiu},
      year={2026},
      eprint={2602.08426},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2602.08426}, 
}