X · @teortaxesTex
· X / Twitter
> Blackwell's increased memory bandwidth and communication bandwidth, plus DeepSeek's megamoe operator (though it seems to only support fp4?), make ac…
> Blackwell's increased memory bandwidth and communication bandwidth, plus DeepSeek's megamoe operator (though it seems to only support fp4?), make achieving 300tps on GLM 5.2 not that difficult.I hope to see at least 150 as the new normalsusun: https://huggingface.co/RedHatAI/GLM-5.2-speculator.dspark 如果按这个数据计算，我们 aim