Qianchao Zhu

Xinming Wei, Chao Jin, Tuo Dai et al.

Large-scale expert parallelism (EP) is becoming pivotal for training and serving frontier MoE models, but it also amplifies device-level expert load imbalance into compute stragglers, token all-to-all bottlenecks, and activation-memory spikes. Existing balancers redistribute experts periodically based on historical load, which becomes unreliable for production deployments with non-stationary load patterns. We present UltraEP, the first exact-load, real-time balancer for large-EP MoE training and serving prefill on rack-scale nodes (RSNs). Built upon the extended scale-up connectivity of RSNs, UltraEP rebalances every microbatch and layer on critical paths, which requires nontrivial co-design of plan solving and expert replication communication to minimize exposed overhead. To this end, UltraEP eagerly reacts to post-gating load with efficient quota-driven planning, and executes the resulting irregular expert-state transfers with RSN-native persistent tile streaming and relay-based fan-out mitigation. Averaged across MoE models from 106B to 671B parameters in training and prefill, UltraEP achieves 94.3% of the force-balanced ideal throughput, delivering 1.49$\times$ improvement over non-balancing, while reducing the final inter-rank imbalance from 1.30$-$4.01 to 1.01$-$1.04. Additionally, we validate UltraEP's scalability and robustness in production MoE training with 2560 GPUs.

Yuliang Liu, Guohao Wu, Shenglong Zhang et al.

The efficient distributed training of Large Language Models (LLMs) is severely hampered by the extreme variance in context lengths. This data heterogeneity, amplified by conventional packing strategies and asymmetric forward-backward costs, leads to critical inefficiencies such as cascading workload imbalances and severe hardware underutilization. Existing solutions attempt to mitigate these challenges, but often at the expense of memory or communication efficiency. To address these challenges, we introduce SlimPack, a framework that fundamentally rethinks data packing and scheduling by decomposing samples into fine-grained slices. This slice-level decomposition immediately mitigates critical memory and communication bottlenecks by transforming large, volatile workloads into a stream of smaller, manageable units. This flexibility is then harnessed for our core innovation, Asymmetric Partitioning, which assembles balanced scheduling units uniquely optimized for the different demands of the forward and backward passes. Orchestrated by a two-phase solver and a high-fidelity simulator, SlimPack holistically resolves imbalances across all parallel dimensions. Extensive experiments demonstrate that SlimPack achieves up to a $2.8\times$ training throughput improvement over baselines, breaking the conventional trade-off by delivering both superior balance and high resource efficiency.

Siran Liu, Yang Ye, Qianchao Zhu et al.

Autoregressive decoding inherently limits the inference throughput of Large Language Model (LLM) due to its sequential dependency. Speculative decoding mitigates this by verifying multiple predicted tokens in parallel, but its efficiency remains constrained by what we identify as verification heterogeneity -- the uneven difficulty of verifying different speculative candidates. In practice, a small subset of high-confidence predictions accounts for most successful verifications, yet existing methods treat all candidates uniformly, leading to redundant computation. We present HeteroSpec, a heterogeneity-adaptive speculative decoding framework that allocates verification effort in proportion to candidate uncertainty. HeteroSpec estimates verification complexity using a lightweight entropy-based quantifier, partitions candidates via a data-driven stratification policy, and dynamically tunes speculative depth and pruning thresholds through coordinated optimization. Across five benchmarks and four LLMs, HeteroSpec delivers an average 4.24$\times$ decoding speedup over state-of-the-art methods such as EAGLE-3, while preserving exact output distributions. Crucially, HeteroSpec requires no model retraining and remains compatible with other inference optimizations, making it a practical direction for improving speculative decoding efficiency.

Qianchao Zhu, Jiangfei Duan, Chang Chen et al.

Large language models (LLMs) now support extremely long context windows, but the quadratic complexity of vanilla attention results in significantly long Time-to-First-Token (TTFT) latency. Existing approaches to address this complexity require additional pretraining or finetuning, and often sacrifice model accuracy. In this paper, we first provide both theoretical and empirical foundations for near-lossless sparse attention. We find dynamically capturing head-specific sparse patterns at runtime with low overhead is crucial. To address this, we propose SampleAttention, an adaptive structured and near-lossless sparse attention. Leveraging observed significant sparse patterns, SampleAttention attends to a fixed percentage of adjacent tokens to capture local window patterns, and employs a two-stage query-guided key-value filtering approach, which adaptively select a minimum set of key-values with low overhead, to capture column stripe patterns. Comprehensive evaluations show that SampleAttention can seamlessly replace vanilla attention in off-the-shelf LLMs with nearly no accuracy loss, and reduces TTFT by up to $2.42\times$ compared with FlashAttention.