vllm Op benchmark

Hi, have you open-sourced the vLLM MoE benchmarks?

In the paper, you mention:

> The full MoE performance is also shown in Figure 9. On average, TileLink achieves a 1.14× speedup over vLLM. The maximal speedup over cuBLAS+NCCL is 20.76×. Note that existing libraries such as FLUX and Async-TP.PyTorch do not support overlapping MoE layers. TileLink supports MoE thanks to its flexible primitives and dynamic mappings.

I have a question regarding this. Do you have numbers comparing vLLM end-to-end (E2E) performance with TileLink MoE E2E performance?

As far as I know, vLLM executes in the following sequence:
>qkv_projection + attention + o_proj + All-Reduce + LayerNorm + MoE Block + All-Reduce + LayerNorm

Do you have numbers comparing TileLink with vLLM in this setting?
If so, how have you implemented the comparison? Have you also overlapped the non-moe rs and ag?

The issue names no files, tests, or entry points. Start by locating the vLLM and TileLink MoE benchmark implementations and the measurements behind Figure 9; done would mean publishing or documenting an end-to-end comparison, including whether non-MoE communication is overlapped.

Written by the indexing model from the issue text.