NeuralZip is a method for lossless compression of floating‑point model checkpoints that pre‑computes a reusable setup based on exponent statistics. By grouping tensor chunks with similar exponent distributions, sharing Huffman codes, and packing recurring exponent tuples, it achieves 1.81–21.33× faster compression than baselines while preserving exact bit‑to‑bit reconstruction. The setup can be transferred across compatible architectures and reused during training, reducing GPU memory usage by up to 27.5% without affecting model outputs.
Lossless compression can reduce the storage and movement of model weights without changing their floating-point values, but repeated statistical analysis and code construction add computational overhe...
The paper introduces Losslessly Compressed Tensors (LCT), a GPU‑friendly framework that repeatedly encodes and decodes tensors during training. LCT uses Huffman coding with precomputed codebooks to exploit redundancy in floating‑point data, achieving up to a 1.455× compression ratio. Experiments on models such as NanoGPT and Qwen3‑4B show memory savings of 24.6% and 23.8% respectively, enabling larger context lengths with only modest increases in step time.
arXiv:2605. 01708v3 Announce Type: replace-cross Abstract: Contemporary systems serving large language models (LLMs) have adopted prefill-decode disaggregation to load-balance between the compute-bound prefill phase and the memory-bound decode phase.
BitSparse introduces a GPU‑resident, lossless compression format for sparse ReLU and ReLU² activations in feed‑forward networks. By integrating custom Triton kernels into an autograd function, it compresses and reconstructs activation caches without loss, supporting BF16 and FP8 formats and offering optional sign‑bit packing to approach theoretical limits. The method reduces per‑token VRAM usage by about 30% for Modded‑NanoGPT and Nemotron‑H, with negligible overhead and a new preallocated buffer system that eliminates host synchronization during training.
TensorDex is a lossless, tensor‑centric storage system designed to reduce the storage footprint of large language models by performing delta compression at the tensor level. It decomposes models into individual tensors, uses compact fingerprints to predict which tensors can be compressed together, and incrementally clusters tensors to select effective reference bases as new models arrive. In experiments on 2,890 Hugging Face models, TensorDex cuts storage usage by 70.5%, outperforms the previous best system by 37%, and achieves compression and decompression speeds of 22.9 GB/s and 28.4 GB/s, respectively.
arXiv:2501.13558v4 Announce Type: replace Abstract: Splat-based radiance fields have rapidly diversified in primitive type, kernel support, and appearance parameterization. This diversity makes compr...
The paper "LLM Inference in a Flash!" proposes an integer‑only quantization scheme and a dictionary‑based KV cache compression technique to enable large language model inference on compute‑in‑flash (CIF) devices. By eliminating floating‑point operations and reducing KV cache traffic through sparse dictionary coding, the authors achieve minimal accuracy loss while cutting dynamic KV cache traffic by 15× on Llama‑3.1‑8B and Qwen‑2.5‑7B models.
The paper presents a budget‑aware compression pipeline for deploying 70B‑parameter language models on a single NVIDIA GPU. It examines how pruning, quantization, and KV‑cache compression interact, showing that layer‑wise pruning improves weight quantization robustness and that KV‑cache sparsification complements INT8 KV quantization without harming decoding speed. Using these insights, the authors compressed a 70B model to ~33 GB, achieving ~57 tokens/s on 10k‑token prompts on an A40 while maintaining accuracy within 5% on standard benchmarks.
The paper introduces a hardware-software co‑design framework that compresses Mixture‑of‑Experts (MoE) model weights into low‑precision, hardware‑native sparse representations, enabling efficient execution on Sparse Tensor Cores (SpTCs). By relaxing discrete support selection through continuous reparameterization, the method jointly optimizes quantized weights and supports a router‑weighted reconstruction objective, achieving up to 4.35 percentage‑point gains in joint sparse‑quantization accuracy while retaining 96.09% of the original model’s performance. A custom grouped sparse GEMM kernel further boosts inference speed, outperforming NVIDIA’s baseline by up to 1.65× and reducing latency by up to 4.03× on B200 GPUs.
arXiv:2607. 18187v1 Announce Type: cross Abstract: Large-scale scientific simulations generate volumetric data at rates that far outpace advances in storage and network bandwidth, making effective lossy compression increasingly critical.
arXiv:2608. 08081v1 Announce Type: cross Abstract: Large mixture-of-experts (MoE) language models with 26--120 billion parameters exceed the memory capacity of consumer devices through three simultaneous pressures: resident weight matrices, key-value (KV) cache state that grows linearly with context, and dozens of expert sublayers that must be paged on demand.
STAPLE is a method that integrates static sparse memory into pretrained language models by adding token‑indexed embeddings as a residual pathway while preserving the original feed‑forward network. On the OLMo‑2 1B and 7B models, STAPLE outperforms dense training and other static sparse baselines, improving benchmark scores by up to 4.27 points and achieving faster decoding with less GPU memory and energy usage. The approach demonstrates that large pretrained models can be expanded efficiently without costly sparse pretraining from scratch.