ML Model Compression via Hybrid Sparsification and Quantization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing machine learning model compression techniques uniformly apply sparsification and quantization, which negatively impacts model quality, especially for outlier values, leading to suboptimal performance in resource-constrained applications.

Innovation Solution

The method involves apportioning machine learning model values into data structures with defined structured sparse patterns and non-uniform data representations, allowing inlier and outlier values to be stored with different sparsity and precision levels.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If uniform sparsification and quantization are applied to all model weights, then model size and computation requirements are reduced, but model quality significantly deteriorates due to disproportionate impact on outlier values

Engineering Contradiction:
Improvemodel sizeVSAvoidmodel quality
Core Design Contradiction:
Quantity of substanceVSManufacturing precision

Solution Approach 1:

The patent applies different quantization bit widths to different subsets of model weights based on their statistical properties. Inlier weights (following the dominant distribution) are quantized to lower precision (e.g., 4-bit), while outlier weights (following tail distributions) are quantized to higher precision (e.g., 8-bit or 16-bit). This local differentiation preserves model quality by protecting sensitive outlier values while achieving compression on the majority of weights.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent segments model weights into multiple subsets based on their statistical characteristics, specifically separating inlier weights from outlier weights using distribution analysis. This segmentation allows independent quantization strategies to be applied to each subset, optimizing the balance between compression and quality preservation for different weight categories.

Inventive Principle:
Principle #1Segmentation

2Quantity of substance

If uniform quantization is applied to reduce storage requirements, then memory consumption decreases, but performance deteriorates due to loss of precision in critical outlier values

Engineering Contradiction:
Improvememory consumptionVSAvoidmodel performance
Core Design Contradiction:
Quantity of substanceVSReliability

Solution Approach 1:

The patent implements local quality by assigning different quantization precisions to different weight subsets. Weights identified as outliers through statistical analysis receive higher precision allocation (e.g., 8-bit or 16-bit quantization), while inlier weights use lower precision (e.g., 4-bit). This ensures that memory-constrained applications maintain performance by preserving precision where it matters most.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent changes the quantization parameter (bit width) dynamically based on the statistical properties of weight subsets. By analyzing weight distributions and identifying outliers, the system adapts the precision parameter to match the actual needs of different weight groups, optimizing the trade-off between memory efficiency and model performance.

Inventive Principle:
Principle #35Parameter changes

3Productivity

If aggressive sparsification is applied to accelerate model execution, then computation speed increases, but model accuracy deteriorates due to removal of important connections

Engineering Contradiction:
Improvecomputation speedVSAvoidmodel accuracy
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent applies structured sparsification patterns (e.g., block sparsity) that preserve local connectivity structures while achieving overall sparsity. By maintaining contiguous blocks of weights rather than random removal, the method preserves important local relationships in the model while still achieving computation acceleration through sparse matrix operations.

Inventive Principle:
Principle #3Local quality

Data Source

PatentUS20250094864A1Compression of machine learning models via sparsification and quantization
Publication Date: 2025.03.20 NVIDIA CORP
  • US20250094864A1 patent drawing
  • US20250094864A1 patent drawing
  • US20250094864A1 patent drawing

AI summary

Machine learning is a process that learns a model from a given dataset, where the model can then be used to make a prediction about new data. In order to reduce the size, computation, and latency of a machine learning model, a compression technique can be employed which includes model sparsification and quantization. To limit the extent to which the quality of the model is impacted when uniformly applying sparsification and quantization to all values of the model, the present disclosure provides for a hybrid sparsification and quantization of the model.