Brevis compresses model checkpoints by synthesizing self-contained DSL programs that reconstruct tensors bit-exactly, reducing storage requirements and achieving high compression and decompression speeds.
Brevis is the first to formulate tensor compression as program synthesis and learn a production prior to guide synthesis.
Before reading this…
Applications
To understand this paper, make sure you know these concepts first:
Model checkpoints are growing in both number and size, which makes archival, transfer, and deployment increasingly costly. General-purpose compressors can reduce storage requirements but ignore tensor structure, whereas existing tensor-specific compressors rely on fixed and format-specific pipelines. We present Brevis, which formulates lossless tensor compression as program synthesis. We design a typed domain-specific language (DSL) that captures recurring tensor structures, such as repeated regions and floating-point fields, through a set of reversible operators. Given a tensor, Brevis synthesizes a self-contained DSL program that reconstructs it bit-exactly. A checkpoint-specific production prior, learned from a small representative sample of tensors, guides a bounded A* search to synthesize compact programs, which can later be executed directly for bit-exact decompression. On 10 public checkpoints spanning language, audio, and image generation models, Brevis reduces 2.13 TB of checkpoint data to 1.41 TB, a 33.93% storage reduction. It produces archives up to 30.87% smaller than those of four general-purpose compressors, including zstd and gzip, and smaller archives than the tensor-specific compressors ZipNN and DFloat11. Under a practical concurrency configuration, Brevis achieves 3.60 GB/s compression and 6.61 GB/s decompression while preserving every source byte.
Robust Low-Tubal-Rank Tensor Completion under Cross-Concentrated Sampling
This paper proposes Robust Iterative t-CUR (R-ItCUR), a tensor-native algorithm…
Toward Alias-Free Channel Extrapolation in Upper Mid-Band Systems: A Spatial-Frequency-Temporal Tens…
This paper proposes a tensor-structured multi-domain channel extrapolation frame…
InstantInfer: Enabling Fast LLM Cold Start with Communicating Finite Automata
The paper proposes the Communicating Finite Automata (CFA) abstraction and a fra…
3D Gaussian Splatting for Scientific Particle Data Compression and Rendering
The paper proposes ParticleGS, a visualization-aware framework for compressing l…
Lossless Address Coding for Quantum Networks
This paper proposes a lossless source coding scheme for quantum network addressi…
PRISM: Evaluating POSIX Storage Systems for AI Research Workflows
The paper introduces PRISM, an evaluation framework for assessing POSIX storage…
Hilbert Operator for Progressive Encoding (HOPE): A Mathematical Framework for Deconstructing Learne…
The paper introduces HOPE, a mathematical framework for network compression that…
From Bit-Position Sensitivity to Unequal Error Protection for DNN Inference Memory
This paper characterizes per-bit-position fault sensitivity in machine learning…