| # MoE Architecture |
|
|
| The MoE model is the main scaling direction for code completion. The strongest |
| current structure is sparse FFN experts with shared attention and optional |
| multi-token prediction heads. |
|
|
| ## Recommended Structure |
|
|
| - Decoder-only Transformer. |
| - RMSNorm before attention and FFN. |
| - RoPE positional encoding. |
| - Grouped-query attention. |
| - SwiGLU experts. |
| - Top-2 routing during training. |
| - Shared expert for common syntax and indentation. |
| - Router z-loss and auxiliary load-balancing loss. |
| - QK norm for long-context attention stability. |
| - Expert-choice telemetry from the first step, not after loss diverges. |
| - Optional multi-token prediction heads with low loss weight. |
|
|
| ## Best Current Shape |
|
|
| Use shared attention and sparse FFN. Keep early layers mostly dense/shared so |
| the model learns lexical syntax before specializing. |
|
|
| - Early blocks: dense FFN or shared expert heavy. |
| - Middle blocks: sparse MoE every other block. |
| - Late blocks: sparse MoE with stricter router stability checks. |
| - Expert count: start with 8 experts for smoke, then 16-32 for H100 ablation. |
| - Routing: top-2 with capacity factor 1.25 during training. |
| - Inference ablation: top-1 only after top-2 training is stable. |
|
|
| For code completion, the highest-value specialization is not one expert per |
| language. Better targets are syntax/indentation, FIM reconstruction, |
| repository-context patterns, generated-code continuation, and real-code APIs. |
|
|
| ## Routing Metrics |
|
|
| Log: |
|
|
| - Per-layer expert load. |
| - Per-expert token fraction. |
| - Dropped-token count. |
| - Router entropy. |
| - Auxiliary loss. |
| - Domain-to-expert correlation. |
|
|
| ## Failure Modes |
|
|
| - Expert collapse. |
| - Router churn. |
| - Synthetic Python overfitting. |
| - MTP loss overpowering next-token loss. |
| - Shared expert absorbing too much load. |
| - Dropped tokens above 1% for sustained windows. |
|
|
