Baseline: Bucket Attention

[sahiljoshi515]

🧱 Baseline Addition PR

📘 Description

Baseline: Compressed Retrieval Attention (Do you think this is a better name?)

✅ Baseline Checklist

(Follow the docs/baslines/developer_guide.md)

Item	Status	Notes (Note comments, remove placehoders)
🧩 1. baseline code	[ Y]	This is a new baseline.
⚙️ 4. Performance Profiled	[ Y]	See below
📖 5. Documentation Created	[Y ]	docs/baselines/Bucket_Attn.md

📊 Results Summary

quality

results in different setups

Our Results:

	niah_multikey_3	qa_2
Our Code (10x sparsty,)	100	56

research performance (H200)

Hyper parameters: K=4, L = 30, top_t = 4

📈 Masker Overhead Analysis:

• Sparse Attention: 4.400 ms
• Baseline (no maskers): 0.299 ms
• Masker overhead: 4.101 ms (1373.8%)
  ❌ High overhead - significant masker cost

Hyper parameters: K=4, L = 60, top_t = 4

📈 Masker Overhead Analysis:

• Sparse Attention: 7.318 ms
• Baseline (no maskers): 0.299 ms
• Masker overhead: 7.019 ms (2350.5%)
  ❌ High overhead - significant masker cost

---