💥 Dual 3090s. No NVLink. Still ate 256K context. Qwen3.8-27B INT4 + MTP on 2× … · @QwenAIFans · TwKit

@QwenAIFans

💥 Dual 3090s. No NVLink. Still ate 256K context. Qwen3.8-27B INT4 + MTP on 2× RTX 3090 over plain PCIe: Prefill ~2,010 tok/s at 256K. Decode ~123 tok/s after. Old cards, huge window. Would you skip NVLink for this? via QEngine 👇 https://github.com/bpx03/QEngine

曝光 564 · 评论 6 · 点赞 14 · 书签 6