PyTorch Conference Europe 2026

Pipeline parallelism is vital for large models, but advanced schedules for SOTA LLMs are difficult to express in current PyTorch. MoE communication dominates the critical path, making latency hiding essential. Leading systems use fw-bw overlapping; fw-fw and bw-bw overlapping further boost throughput.

Schedules like ZeroBubbleV and DualPipeV rely on dI-dW backward splitting for fine-grained overlap. However, eager-mode implementations require a patchwork of fragile integrations (multi-threading, custom autograd functions, activation checkpointing, etc.) that rely on implicit behavior and hand-written logic with poor torch.compile compatibility and upstream composability.

We present Graph-Based PP: stages are compiled to reusable FX graphs executed via an explicit schedule language. Users write standard PyTorch code while specifying schedules at varying granularity; all manipulations run as graph passes, abstracting complexity away from user code and into the compiler/runtime, allowing for greater composability.

We have integrated Graph-PP into TorchTitan and AutoParallel on real MoE workloads, targeting upstream inclusion in torch.distributed.