Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
PipeFusion: Patch-level Pipeline Parallelism for Diffusion Transformers Inference
Authors: Jiarui Fang, Jinzhe Pan, Aoyu Li, Xibo Sun, WANG Jiannan
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that Pipe Fusion achieves state-of-the-art performance on 8 L40 PCIe GPUs for Pixart, Stable-Diffusion 3, and Flux.1 models. |
| Researcher Affiliation | Collaboration | Jiarui Fang Byte Dance EMAIL Jinzhe Pan HUST EMAIL Aoyu Li Byte Dance EMAIL Xibo Sun Tencent EMAIL Jiannan Wang The University of Hong Kong EMAIL |
| Pseudocode | No | The paper describes methods through textual explanation and visual diagrams (e.g., Figure 1: Workflow of Di Ts inference, Figure 3: Workflow of the Pipe Fusion as patch-level pipelined parallelism) but does not include formal pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our Source code is available at https://github.com/xdit-project/x Di T. |
| Open Datasets | Yes | We use the COCO Captions 2014 [37] dataset to evaluate the FID scores. |
| Dataset Splits | Yes | During the evaluation, a subset comprising 30,000 images is sampled from the validation set and resized to 256px to serve as the reference dataset. Concurrently, each experiment generates 30,000 images of 256px, each paired with a caption derived from the COCO Captions 2014 dataset, as the sample dataset. |
| Hardware Specification | Yes | In our experimental setup, we deployed our trials on an 8 L40-48GB (PCIe Gen4x16) cluster to evaluate three prominent Di T models |
| Software Dependencies | Yes | The software stack utilized includes Py Torch 2.4.1, CUDA Runtime 12.1.105, and diffusers 0.30.3. |
| Experiment Setup | Yes | We employ a default warmup step of 1 for both Distri Fusion and Pipe Fusion. The software stack utilized includes Py Torch 2.4.1, CUDA Runtime 12.1.105, and diffusers 0.30.3. For Pipe Fusion, we select the best latency performance by searching the patch number M from 2, 4, 8, 16, 32. |