Total: 1
Training trillion-parameter models relies on pipeline parallelism to scale across thousands of accelerators. However, architectures exemplified by the Qwen family are evolving from uniform Transformer blocks into heterogeneous compositions that combine sparse Mixture-of-Experts (MoE) with different attention variants. This heterogeneity breaks the uniformity assumption behind existing pipeline systems: partitions chosen from serial layer costs become imbalanced after communication is overlapped with computation, because different layer combinations hide different fractions of their communication. Furthermore, at the scale of 10K+ GPUs, MoE routing variation adds a runtime source of stalls, creating transient idle slots that static plans cannot anticipate. We present Tessera, a framework that co-optimizes partitioning and overlap scheduling for heterogeneous pipeline training, and adapts to runtime dynamics. Tessera introduces: (1) an overlap scheduler that synthesizes fine-grained interleavings for each layer combination; (2) an overlap-aware partitioner that uses profiled post-overlap cost to select partitions balanced for parallel execution; and (3) a dynamic bubble optimizer that fills routing-induced idle slots with movable tasks at runtime. Deployed on production clusters for pre-training Qwen3 and Qwen3-Next, Tessera improves throughput by 20%–33% over our production baseline across five workloads at scales from 4,096 to 12,288 GPUs, reaching 39% MFU on a trillion-parameter model. In controlled experiments against Megatron-Core MoE with public recipes, Tessera achieves up to 1.24× higher MFU.