nieche/DL-Architectural-DAGs-2026

DL Architectural DAGs — HPC Features

A research dataset of computation graphs (DAGs) extracted from 200+ deep learning models via torch.export, annotated with hardware-performance-counter (HPC) proxy features per node.

Designed for research on operator scheduling, memory mapping, and performance prediction of DNN workloads on heterogeneous hardware (CPUs, GPUs, accelerators).

MEAN PARALLELISM = 2.19

---

Splits

Split	Rows	Description
models	~200	One row per model — metadata + aggregate stats
nodes	~300 000	One row per graph node — per-operation HPC features

---

Schema

models split

Column	Type	Description
model_id	string	Unique key: {source}__{model_name}
source	string	timm / torchvision / hf
model_name	string	Original model identifier
export_mode	string	strict or non-strict (torch.export mode used)
node_count	int32	Total number of graph nodes
total_flops	float64	Sum of FLOPs across all nodes
total_transfer_bytes	float64	Sum of output-tensor bytes across all nodes

nodes split

Column	Type	Description
model_id	string	Foreign key → models.model_id
source	string	timm / torchvision / hf
model_name	string	Model this node belongs to
node_index	int32	Topological order index in the graph
node_name	string	ATen IR node name (e.g. addmm, convolution)
flops	float64	Estimated FLOPs (MACs × 2)
mem_byte	float64	Input-tensor bytes read by this node (weight-load proxy)
transfer_byte	float64	Output-tensor bytes produced (data-volume / communication proxy)
parents	list[string]	Predecessor node names — encodes DAG edges

---

Models Covered

Library	Count	Architectures
timm	~100	ResNet, EfficientNet, ViT, DeiT, Swin, ConvNeXt, MobileNet, DenseNet, RegNet, MLP-Mixer, ResMLP, CaiT, LeViT, NFNet, Twins, PVT, HRNet, CrossViT, EfficientFormer, VGG, …
torchvision	~51	AlexNet, VGG, SqueezeNet, ResNet, DenseNet, Inception, GoogLeNet, MobileNet, ShuffleNet, EfficientNet, RegNet, ViT-B, Swin-T/S/B, ConvNeXt, …
HuggingFace Transformers	~50	BERT family, DistilBERT, RoBERTa, ALBERT, ELECTRA, GPT-2, OPT, GPT-Neo, DeBERTa, MobileBERT, FNet, MPNet, TinyBERT, domain-BERTs, …

---

FLOPs Estimation Method

Operation	Formula
mm / addmm (Linear)	2 × M × N × K
bmm (batched matmul)	2 × B × M × N × K
convolution (any-D, grouped)	2 × output_elements × (C_in/g × kernel_dims)
scaled_dot_product_attention	2 × B × H × L² × D
einsum (DeBERTa disentangled)	2 × output_elements
element-wise (ReLU, Add, LayerNorm, …)	output_elements

---

Usage

from datasets import load_dataset

ds = load_dataset("nieche/DL-Architectural-DAGs-2026")

# All nodes of ResNet-50 from timm
resnet_nodes = ds["nodes"].filter(lambda x: x["model_id"] == "timm__resnet50")

# Top-10 most FLOPs-intensive nodes across the whole dataset
df = ds["nodes"].to_pandas()
print(df.nlargest(10, "flops")[["model_id", "node_name", "flops"]])

---

Citation

@dataset{dl_architectural_dags_2026,
  title  = {DL Architectural DAGs — HPC Features},
  year   = {2026},
  url    = {https://huggingface.co/datasets/nieche/DL-Architectural-DAGs-2026}
}