Method Drift›Long-context / context-window extension
SnapKV
SnapKV: LLM Knows What You are Looking for Before GenerationLong-context / context-window extension · first seen Apr 22, 2024
superseded — cited as a baseline and beaten by newer methods
2 papers critique it · 3 beat it on benchmarks
What papers say
Verbatim critique sentences, each from a paper that cites SnapKV as a baseline.
“Although these methods generally have low additional overhead, they often lead to noticeable performance degradation.”
— A$^2$ATS: Retrieval-Based KV Cache Reduction via Windowed Rotary Position Embedding and Query-Aware Vector Quantization“While these methods differ in selecting tokens for KV cache retention, they generally apply a uniform budget size across layers, even though the optimal budget size may vary.”
— ZigZagkv: Dynamic KV Cache Compression for Long-context Modeling based on Layer Uncertainty
Beaten on benchmarks
Head-to-head results where a newer method reports beating SnapKV. Values are copied from the source paper's tables — verify against the cited paper.
- A$^2$ATS: Retrieval-Based KV Cache Reduction via Windowed Rotary Position Embedding and Query-Aware Vector Quantization
A2ATS beats SnapKV · Accuracy [Llama-3.1-8B-Instruct, Sparsity ~0.060]
86.6 vs 72.7
- A$^2$ATS: Retrieval-Based KV Cache Reduction via Windowed Rotary Position Embedding and Query-Aware Vector Quantization
A2ATS beats SnapKV · Accuracy [MegaBeam-Mistral-7B-512K, Sparsity ~0.062]
86.3 vs 67.6
- ZigZagkv: Dynamic KV Cache Compression for Long-context Modeling based on Layer Uncertainty
ZigZagKV beats SnapKV · Avg. [KV Size = 128]
43.30 vs 42.90
- ZigZagkv: Dynamic KV Cache Compression for Long-context Modeling based on Layer Uncertainty
ZigZagKV beats SnapKV · Attention loss [Mistral Budget 128]
2.447 vs 2.706
- ZigZagkv: Dynamic KV Cache Compression for Long-context Modeling based on Layer Uncertainty
ZigZagKV beats SnapKV · Attention loss [Mistral Budget 256]
1.249 vs 1.543
- ZigZagkv: Dynamic KV Cache Compression for Long-context Modeling based on Layer Uncertainty
ZigZagKV beats SnapKV · Attention loss [Mistral Budget 512]
0.611 vs 0.893
- ZigZagkv: Dynamic KV Cache Compression for Long-context Modeling based on Layer Uncertainty
ZigZagKV beats SnapKV · Attention loss [LLaMA Budget 128]
1.504 vs 1.763
- ZigZagkv: Dynamic KV Cache Compression for Long-context Modeling based on Layer Uncertainty
ZigZagKV beats SnapKV · Attention loss [LLaMA Budget 256]
0.637 vs 0.901
- ZigZagkv: Dynamic KV Cache Compression for Long-context Modeling based on Layer Uncertainty
ZigZagKV beats SnapKV · Attention loss [LLaMA Budget 512]
0.226 vs 0.463
- ZigZagkv: Dynamic KV Cache Compression for Long-context Modeling based on Layer Uncertainty
ZigZagKV beats SnapKV · Hidden state loss [Mistral Budget 128]
2.544 vs 2.550
- ZigZagkv: Dynamic KV Cache Compression for Long-context Modeling based on Layer Uncertainty
ZigZagKV beats SnapKV · Hidden state loss [Mistral Budget 256]
1.495 vs 1.509
- ZigZagkv: Dynamic KV Cache Compression for Long-context Modeling based on Layer Uncertainty
ZigZagKV beats SnapKV · Hidden state loss [Mistral Budget 512]
0.830 vs 0.851
Newer alternatives
Recent methods in the same sub-problem, not yet superseded in the knowledge base.