BigMac uses nested pipelines to speed up multimodal LLM training by up to 1.9×

机器之心 · wechat · 2026-07-24

BigMac proposes a nested pipeline schedule to break the compute/memory trade-off in multimodal training

Machine Heart covers BigMac, a new training paradigm for multimodal LLMs that keeps the LLM pipeline as the backbone and safely nests encoder and generator work into dependency-valid slots.

The goal is to avoid the usual Pareto trade-off: either compute-efficient pipelines that keep bubbles low but retain lots of activation memory, or memory-efficient pipelines that reduce activation lifetime but introduce cross-module stalls.
BigMac decouples global scheduling from runtime execution, so LLM pipeline execution can stay optimized while modality modules are inserted without breaking the original order.
The system also adds schedule-aware tooling: a visible global operator plan, a transparent PP interface for model developers, plus profiler / simulator / visualization tools for diagnosis and tuning.
In experiments on two multimodal workloads, BigMac reports 1.08×–1.9× speedups over baselines and stable memory use as batch size grows.
The paper says BigMac is already used as a core component in production training for dots multimodal models.

Original post →

BigMac uses nested pipelines to speed up multimodal LLM training by up to 1.9×

More from Infra