Alex (@stel4th) header
Alex (@stel4th) profile picture

i like to code πŸ™ƒ @ClemsonUniv grad | Vroom vroom AI @baseten

United States
Joined July 2017
Photo by Alex (@stel4th)
Alex retweeted
going to start calling it the australian pareto
We just released the #1-ranked decision model, again. bison-1 and groundhog-1 are built on an American backbone (Gemma 4). They deliver improved quality and similar speed to mammoth-1 and elephant-1, with a 128k context window and multimodal support (images + text). baseten.co/blog/meet-bison-1…
4
5
77
4,903
How do you get up to 2.9x prefill throughput on the same GPUs and kernels? We introduce WavePP, a new runtime reaching state of the art for pipeline parallelism (PP) prefill. The question was: β€œWhat if each PP stage ran as an independent, lightweight model worker?” To answer this, we had to rethink everything, from request admission, prefix reuse agreements, and how to divide work. In experiments on TensorRT-LLM PP4, WavePP delivers up to 2.91x throughput on GLM-5.2 and 2.02x on MiniMax M2.7 under high prefix reuse against the same kernels on the same GPUs. This is what we changed 🧡
19
28
11
332
37,418