Graph Transformer Efficiency Benchmarking on tolokers 2
9.78Latency Speedup (Forward Pass)CSR fused kernels
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| CSR fused kernelsHeads (H)=8, Head Dimension (D)=1282026.05 | 9.78 | 2.18 | 69.74 | 44.8 | |
| CSR fused kernelsHeads (H)=2, Head Dimension (D)=642026.05 | 5.59 | 1.89 | 32.59 | 28.55 | |
| Ours (WSB/TC)Number of Heads (H)=4, Head Dimension (D)=1282026.05 | 2.69 | 0.71 | 1.23 | 1.2 | |
| Ours (CSR fused)Number of Heads (H)=4, Head Dimension (D)=1282026.05 | 2 | 1.1 | 1.57 | 1.13 | |
| PyGHeads (H)=2, Head Dimension (D)=642026.05 | 0.45 | 0.7 | 0.61 | 0.71 | |
| PyGHeads (H)=8, Head Dimension (D)=1282026.05 | 0.35 | 0.44 | 0.6 | 0.67 |