Sequence Length Imbalance Causing Distributed Training Stragglers

Sequence length imbalance across micro-batches within the same global batch causes certain pipeline stages to compute significantly longer than others. ByteDance's production trace analysis found that 21.4% of straggler-affected jobs showed measurable throughput degradation due to variable-length sequences creating uneven compute load across data-parallel ranks.

Sequence Length Imbalance Causing Distributed Training Stragglers is a Reliability failure seen during ML training runs. Sequence length imbalance across micro-batches within the same global batch causes certain pipeline stages to compute significantly longer than others. ByteDance's production trace analysis found that 21.4% of straggler-affected jobs showed measurable throughput degradation due to variable-length sequences creating uneven compute load across data-parallel ranks. Common tags: Bytedance, Straggler, Sequence Length, Flash Attention.

Attention computation scales quadratically with sequence length (O(n^2) for dense attention). Standard random batching distributes long sequences unevenly across micro-batches. No cross-rank sequence length synchronization before micro-batch formation. Taken together, these mechanisms explain why the failure is reproducible, why it tends to surface on specific workloads or scales, and why generic mitigation attempts often fall short without addressing the underlying cause.

• Long sequences in one micro-batch delay entire pipeline iteration
• Throughput drops even when padding is removed via variable-length batching
• Straggler pattern shifts across iterations as sequences are resampled

• Data-parallel + pipeline-parallel training with variable-length sequence datasets
• LLM training with packed sequences and attention masking
• Jobs using dynamic batching without sequence length sorting

Use this checklist to test the hypothesis against a small reproduction. No single line proves the root cause, so preserve the preceding events and compare one variable at a time.

• ✓Reproduce the failure from a clean checkpoint/seed: the symptom must appear without warm-up state from a previous run.
• ✓Verified signal present: Iteration time correlates with max sequence length in global batch rather than mean
• ✓Verified signal present: Attention entropy (computation) peaks on micro-batches containing longest documents
• ✓Verified signal present: Per-rank FLOP utilization varies and tracks max sequence length per data-parallel shard
• ✓A targeted fix from the "How to fix it" section eliminates or substantially reduces the symptom within one validation pass.

Select the traceback or open the failed terminal, then run Denpex locally to see the initiating rank, collateral failures, exact fix, and verification command without uploading the log.

• Attention computation scales quadratically with sequence length (O(n^2) for dense attention)
• Standard random batching distributes long sequences unevenly across micro-batches
• No cross-rank sequence length synchronization before micro-batch formation