V100 unlocked NVFP4, GLM5.3F insane output 145 toks
10.6K
2
2026-09-23 20:59:16
186
17
183
48
课程总结
收起全部章节
关闭课程总结
暂无课程总结
8 V100s, 256GB of VRAM, we got GLM-5.3-Flash-NVFP4 320B running.
This time, actual testing was performed based on 1Cat-vLLM + DFlash2:
About 82 tok/s in low acceptance rate scenarios, up to 145 tok/s in code scenarios.
More importantly, this is only the initial preliminary adaptation version; the entire optimization space has not yet been fully utilized.
A batch of what many consider "old cards" is regaining the ability to run new-generation large models.
The value of existing GPUs may have only just begun to be recalculated.
