ZipLLM: Efficient LLM Storage via Model-Aware Synergistic Data Deduplication and
summary
Hugging Face model storage grows ~275× from 2023 to 2025 to ~77 PB and fine-tuned variants dominate, so ZipLLM targets cross-model redundancy with tensor-level deduplication (metadata shrink ~1000× and parallel ingest) plus model clustering by bit distance. ZipLLM applies lossless XOR deltas with BitX (split-field XOR then ZSTD) to compress sparse mantissa changes; evaluation on 1,700 repos yields about 54% storage reduction and multi-GB/s ingest throughput, ~20% better than prior methods.
AI-written summary · about
chapters (6)
- 0:00ZipLLM and storage crisis
- 1:42LLM storage study observations
- 5:29Problems with deduplication
- 7:16Tensor deduplication and BitX compression
- 10:49Evaluation results and comparison
- 12:41Conclusion and future work
related talks
11mWhen Less Is More: A Technical Overview of LLMs and the Strength of Smaller Models
38mOptimizing LLM Inference for the Rest of Us
30mLLMs: Beyond the Hype - A Practical Journey to Scale
21mScaling LLM Inference with vLLM
30mAccelerated LLM Inference with Anyscale
38mOptimizing vLLM Performance through Quantization
24mLLM Inference Optimization From Simple To Sophisticated
15mvLLM: Easy, Fast, and Cheap LLM Serving for Everyone