dropped tokens

**Dropped Tokens** are **tokens that are discarded in sparse Mixture of Experts models when their selected expert has exceeded its processing capacity buffer — causing information loss, training instability, and inconsistent outputs** — the most visible failure mode of discrete top-k routing in MoE architectures, driving the development of alternative routing strategies (expert choice, soft MoE, capacity-factor tuning) that eliminate or minimize this pathological behavior. **What Are Dropped Tokens?** - **Definition**: In top-k MoE routing, each token selects its preferred experts, but if an expert receives more tokens than its capacity buffer allows (capacity = batch_size / num_experts × capacity_factor), excess tokens are "dropped" — their representation passes through only the residual connection, bypassing the expert FFN entirely. - **Capacity Factor**: The buffer multiplier (typically 1.0–1.5) controlling how many tokens each expert can accept. A capacity factor of 1.0 means each expert can handle exactly (batch_size / num_experts) tokens — any imbalance causes drops. - **Information Loss**: Dropped tokens receive no expert processing — in tasks where every token matters (translation, code generation), dropped tokens introduce systematic errors. - **Non-Deterministic Behavior**: The same input processed in different batch compositions may have different tokens dropped (because drop decisions depend on the batch's routing distribution) — causing inconsistent outputs for identical inputs. **Why Dropped Tokens Are a Problem** - **Quality Degradation**: Token drop rates of 5–15% are common in poorly tuned MoE training — this means 5–15% of tokens in every forward pass receive reduced processing, systematically degrading model quality. - **Training-Inference Mismatch**: Drop rates during training differ from inference (different batch sizes) — the model learns to compensate for drops that don't occur at inference, or encounters drops at inference it never saw during training. - **Gradient Noise**: Tokens dropped in the forward pass still generate gradients through the residual — but these gradients don't reflect the expert processing, introducing noise into the router's gradient signal. - **Unpredictable Quality**: Drop rates vary with input distribution — batches with unusual token distributions experience higher drops, creating unpredictable quality variation in production. - **Fairness Concerns**: Common tokens (that match popular expert specializations) are rarely dropped, while rare or out-of-distribution tokens are frequently dropped — systematically under-serving uncommon inputs. **Mitigation Strategies** **Capacity Factor Tuning**: - Increase capacity factor from 1.0 to 1.5 or 2.0 — allows each expert to accept more tokens. - Trade-off: higher capacity factors increase memory usage and reduce efficiency benefits of sparsity. - Monitoring: track actual drop rate during training and increase capacity until drops are <1%. **Load Balancing Loss**: - Auxiliary loss encouraging uniform expert utilization reduces the routing imbalance that causes drops. - Effective but doesn't guarantee zero drops — extreme batches can still overflow popular experts. **Expert Choice Routing**: - Invert routing direction — experts select tokens instead of tokens selecting experts. - Each expert processes exactly k tokens — drops are eliminated by construction. - Trade-off: variable number of experts per token. **Soft MoE**: - Replace discrete routing with continuous soft weights — every token contributes to every expert. - No discrete assignment means no capacity limits and no drops. - Trade-off: loses inference sparsity benefit. **Dropped Token Impact Analysis** | Drop Rate | Quality Impact | Cause | Action | |-----------|---------------|-------|--------| | **<1%** | Negligible | Normal routing variance | Acceptable | | **1–5%** | Measurable degradation | Moderate imbalance | Increase capacity factor | | **5–15%** | Significant quality loss | Poor load balance | Add/tune balance loss | | **>15%** | Training failure | Router collapse | Switch routing strategy | Dropped Tokens are **the canary in the MoE coal mine** — the most visible symptom of routing pathology that signals expert underutilization, load imbalance, and wasted model capacity, driving the evolution from naive top-k routing toward more sophisticated routing mechanisms that achieve sparse computation without sacrificing tokens.

Go deeper with CFSGPT

Get AI-powered deep-dives, save terms, and run advanced simulations — free account.

Create Free Account