For DS4 Flash, with 2x Sparks, I am getting 35-85 TPS in single stream, fresh context after quite a bit of RoCe config and the DSpark MTP, on vLLM with Ray and tensor parallel = 2. For multi-stream, it tops out all stream at well north of 100-120. This all degrades with context, but I rarely fill context that much, and if I do it's coding where it's non-real-time.

For something like GLM, it's larger, has a larger number of active experts, and doesn't support tensor parallel. This means performance doesn't really scale with more Sparks. You can layer split, but then you are still seeing each layer in series and so if anything performance gets slightly worse. I would not expect more than 10-20 TPS on GLM with 2-4 Sparks.