CoCoDiff: Optimizing Collective Communications for Distributed Diffusion Transformer Inference Under Ulysses Sequence Parallelism

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Ma, Bin, Ding, Xingjian, Bicer, Tekin, Su, Pengfei, Li, Dong
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866915946850942976
author Ma, Bin
Ding, Xingjian
Bicer, Tekin
Su, Pengfei
Li, Dong
author_facet Ma, Bin
Ding, Xingjian
Bicer, Tekin
Su, Pengfei
Li, Dong
contents Diffusion Transformers (DiTs) are increasingly adopted in scientific computing, yet growing model sizes and resolutions make distributed multi-GPU inference essential. Ulysses sequence parallelism scales DiT inference but introduces frequent all-to-all collectives that dominate latency. Overlapping these with computation is difficult due to tight data dependencies, large message volumes, and asymmetric interconnect bandwidths. We introduce CoCoDiff, a distributed DiT inference engine exploiting two observations: (1) V requires only linear projection while Q/K need additional normalization and RoPE, creating opportunities to overlap V's communication with Q/K computation; (2) adjacent denoising steps produce similar tensors, yielding temporal redundancy. CoCoDiff introduces three mechanisms: Tile-Aware Parallel All-to-all (TAPA) decomposes collectives into topology-aligned phases; V-First scheduling hides V's communication behind Q/K computation; and V-Major selective communication transmits only active projections on slow interconnects. On the Aurora supercomputer with four DiT models across 1-8 nodes (up to 96 Intel GPU tiles), CoCoDiff achieves an average speedup of 3.6x, peaking at 8.4x.
format Preprint
id arxiv_https___arxiv_org_abs_2604_14561
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle CoCoDiff: Optimizing Collective Communications for Distributed Diffusion Transformer Inference Under Ulysses Sequence Parallelism
Ma, Bin
Ding, Xingjian
Bicer, Tekin
Su, Pengfei
Li, Dong
Distributed, Parallel, and Cluster Computing
Diffusion Transformers (DiTs) are increasingly adopted in scientific computing, yet growing model sizes and resolutions make distributed multi-GPU inference essential. Ulysses sequence parallelism scales DiT inference but introduces frequent all-to-all collectives that dominate latency. Overlapping these with computation is difficult due to tight data dependencies, large message volumes, and asymmetric interconnect bandwidths. We introduce CoCoDiff, a distributed DiT inference engine exploiting two observations: (1) V requires only linear projection while Q/K need additional normalization and RoPE, creating opportunities to overlap V's communication with Q/K computation; (2) adjacent denoising steps produce similar tensors, yielding temporal redundancy. CoCoDiff introduces three mechanisms: Tile-Aware Parallel All-to-all (TAPA) decomposes collectives into topology-aligned phases; V-First scheduling hides V's communication behind Q/K computation; and V-Major selective communication transmits only active projections on slow interconnects. On the Aurora supercomputer with four DiT models across 1-8 nodes (up to 96 Intel GPU tiles), CoCoDiff achieves an average speedup of 3.6x, peaking at 8.4x.
title CoCoDiff: Optimizing Collective Communications for Distributed Diffusion Transformer Inference Under Ulysses Sequence Parallelism
topic Distributed, Parallel, and Cluster Computing
url https://arxiv.org/abs/2604.14561