Multi-path Speculative Decoding on Llama (test)
14.77Throughput (tokens/s)Traversal
Evaluation Results
| Method | Links | |
|---|---|---|
| TraversalExpansion Mode=Static2026.02 | 14.77 | |
| SpecInfer NDEExpansion Mode=Neural Dynamic Expansion2026.02 | 14.27 | |
| SpecTr NDEExpansion Mode=Neural Dynamic Expansion2026.02 | 14.23 | |
| NaiveTree NDEExpansion Mode=Neural Dynamic Expansion2026.02 | 14.01 | |
| Khisti NDEExpansion Mode=Neural Dynamic Expansion2026.02 | 13.24 | |
| NSS NDEExpansion Mode=Neural Dynamic Expansion2026.02 | 12.45 |