Code Generation on HumanEval (TPS, Speedup, Performance Score)
351.07TPS (Tokens/s)EVICT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| EVICTTemperature=0, Backbone=Ling-flash-2.0, Batch Size=1, TP (Tensor Parallelism)=4, steps=4, topk=8, draft_tokens=322026.05 | 351.07 | — | — | |
| EVICTTemperature=1, Backbone=Ling-flash-2.0, Batch Size=1, TP (Tensor Parallelism)=4, steps=4, topk=8, draft_tokens=322026.05 | 335.78 | — | — | |
| ES-dLLMBase Model=Dream-7B-Base, Few-shot=02026.03 | 305.37 | 6.9 | 38.41 | |
| DualCacheBase Model=Dream-7B-Base, Few-shot=02026.03 | 257.42 | 5.8 | 40.24 | |
| EAGLE-3Temperature=0, Backbone=Ling-flash-2.0, Batch Size=1, TP (Tensor Parallelism)=4, steps=4, topk=8, draft_tokens=322026.05 | 254.55 | — | — | |
| DDDTemperature=0, Backbone=Ling-flash-2.0, Batch Size=1, TP (Tensor Parallelism)=4, steps=4, topk=8, draft_tokens=322026.05 | 253.06 | — | — | |
| EAGLE-3Temperature=1, Backbone=Ling-flash-2.0, Batch Size=1, TP (Tensor Parallelism)=4, steps=4, topk=8, draft_tokens=322026.05 | 249.55 | — | — | |
| DDDTemperature=1, Backbone=Ling-flash-2.0, Batch Size=1, TP (Tensor Parallelism)=4, steps=4, topk=8, draft_tokens=322026.05 | 244.41 | — | — | |
| VanillaTemperature=0, Backbone=Ling-flash-2.0, Batch Size=1, TP (Tensor Parallelism)=42026.05 | 174.51 | — | — | |
| VanillaTemperature=1, Backbone=Ling-flash-2.0, Batch Size=1, TP (Tensor Parallelism)=42026.05 | 169.86 | — | — | |
| LookaheadTemperature=0, Backbone=Ling-flash-2.0, Batch Size=1, TP (Tensor Parallelism)=42026.05 | 130.89 | — | — | |
| LookaheadTemperature=1, Backbone=Ling-flash-2.0, Batch Size=1, TP (Tensor Parallelism)=42026.05 | 128.99 | — | — | |
| STDecBase Model=Dream-7B-Instruct, Decoding Strategy=STDec2026.04 | 64.5 | 5.46 | 60.37 | |
| LocalLeapBase Model=Dream-7B-Instruct, Decoding Strategy=LocalLeap2026.04 | 53.8 | 4.55 | 58.54 | |
| STDecBackbone=LLaDA-8B-Instruct2026.04 | 52.92 | 4.89 | 48.78 | |
| LLaDA + LocalLeapBackbone=LLaDA-8B-Instruct2026.04 | 50.33 | 4.65 | 46.34 | |
| Fast-dLLM (Parallel)Base Model=Dream-7B-Instruct, Decoding Strategy=Fast-dLLM (Parallel)2026.04 | 46.87 | 3.97 | 62.2 | |
| DreamBase Model=Dream-7B-Base, Few-shot=02026.03 | 44.26 | 1 | 42.07 | |
| LLaDA + Fast-dLLM (Parallel)Backbone=LLaDA-8B-Instruct2026.04 | 38.49 | 3.56 | 48.78 | |
| Half-StepBase Model=Dream-7B-Instruct, Decoding Strategy=Half-Step2026.04 | 23.69 | 2 | 35.37 | |
| LLaDA + Half-StepBackbone=LLaDA-8B-Instruct2026.04 | 21.67 | 2 | 35.37 | |
| dKV-CacheBase Model=Dream-7B-Instruct, Decoding Strategy=dKV-Cache2026.04 | 15.21 | 1.29 | 56.1 | |
| LLaDA + dKV-CacheBackbone=LLaDA-8B-Instruct2026.04 | 14.23 | 1.32 | 46.95 | |
| Vanilla DreamBase Model=Dream-7B-Instruct, Decoding Strategy=Vanilla2026.04 | 11.82 | 1 | 59.15 | |
| Vanilla LLaDABackbone=LLaDA-8B-Instruct2026.04 | 10.82 | 1 | 48.17 |