Expert-parallelism Load Balancing on EPLB
51BalancednessPACE++
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PACE++Backbone=DS-R1-Qwen3-8B2026.05 | 51 | 0.035 | |
| PACE++Backbone=Qwen3.5-4B2026.05 | 48.9 | 0.053 | |
| GRPOBackbone=DS-R1-Qwen3-8B2026.05 | 47.8 | 0.056 | |
| Max@kBackbone=DS-R1-Qwen3-8B2026.05 | 45.8 | 0.08 | |
| TTT-DBackbone=Qwen3.5-4B2026.05 | 44.6 | 0.054 | |
| GRPOBackbone=Qwen3.5-4B2026.05 | 44.3 | 0.055 | |
| TTT-DBackbone=DS-R1-Qwen3-8B2026.05 | 44.3 | 0.05 | |
| Max@kBackbone=Qwen3.5-4B2026.05 | 40.7 | 0.04 |