Training Throughput on Open-PerfectBlend
0.52Step Time (s)SpecForge
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SpecForgeTarget Model=Qwen3-30B-A3B, Target Model Parallelism=TP=4, Draft Model Parallelism=ZeRO 2, Max Batch Size=16, Seq Length=40962026.03 | 0.52 | 126,030.8 | 4.31 | |
| EAGLETarget Model=Llama3.1-8B, Target Model Parallelism=ZeRO 2, Draft Model Parallelism=ZeRO 2, Max Batch Size=16, Seq Length=40962026.03 | 1.04 | 63,015.4 | 1 | |
| EAGLETarget Model=Qwen3-30B-A3B, Target Model Parallelism=ZeRO 2, Draft Model Parallelism=ZeRO 2, Max Batch Size=8, Seq Length=40962026.03 | 1.12 | 29,257.1 | 1 | |
| SpecForgeTarget Model=Qwen3-235B-A22B, Target Model Parallelism=TP=8, Draft Model Parallelism=ZeRO 2, Max Batch Size=8, Seq Length=40962026.03 | 1.62 | 20,227.2 | 9.99 | |
| SpecForgeTarget Model=Llama3.1-8B, Target Model Parallelism=TP=1, Draft Model Parallelism=ZeRO 2, Max Batch Size=64, Seq Length=40962026.03 | 2.07 | 126,639.6 | 2.01 | |
| EAGLETarget Model=Llama3.3-70B, Target Model Parallelism=ZeRO 3, Draft Model Parallelism=ZeRO 3, Max Batch Size=8, Seq Length=40962026.03 | 2.21 | 14,827.1 | 1 | |
| SpecForgeTarget Model=Llama3.3-70B, Target Model Parallelism=TP=4, Draft Model Parallelism=ZeRO 2, Max Batch Size=16, Seq Length=40962026.03 | 3.18 | 20,608.8 | 1.39 | |
| EAGLETarget Model=Qwen3-30B-A3B, Target Model Parallelism=ZeRO 3, Draft Model Parallelism=ZeRO 3, Max Batch Size=8, Seq Length=40962026.03 | 5.07 | 6,463.1 | 0.2 | |
| EAGLETarget Model=Qwen3-235B-A22B, Target Model Parallelism=ZeRO 3, Draft Model Parallelism=ZeRO 3, Max Batch Size=8, Seq Length=40962026.03 | 11.2 | 2,025.7 | 1 |