DL Architectural DAGs 2026: Computation Graphs from 200+ Deep Learning Models

A research dataset of computation graphs (DAGs) extracted from over 200 deep learning models via torch.export, annotated with hardware-performance-counter (HPC) proxy features per node. It is designed for research on operator scheduling, memory mapping, and performance prediction of DNN workloads on heterogeneous hardware. The dataset was created by user 'nieche' and last updated on May 11, 2026.

Use Cases

• Research on operator scheduling strategies based on annotated hardware-performance-counter features.
• Developing memory mapping and allocation algorithms for DNN workloads based on graph structure.
• Training performance prediction models for deep learning models on heterogeneous hardware.
• Benchmarking compiler and runtime systems for deep learning using a corpus of real model graphs.

Strengths

• Contains computation graphs from over 200 distinct deep learning models.
• Nodes are annotated with hardware-performance-counter (HPC) proxy features.
• Specifically designed for research on heterogeneous hardware performance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the primary 'models' split is approximate (~200), and exact scale is unknown.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

huggingface

Collection Method

Extracted from deep learning models via torch.export.

Freshness

Last updated 2026-05-11 13:30:36; freshness should be verified.