Benchmark of the new unsloth/Qwen3.6-27B-NVFP4 on 4x 5060 ti's with P2P and PP=4 at 1,4,8,12, and 16 concurrency.
英文标题：Benchmark of the new unsloth/Qwen3.6-27B-NVFP4 on 4x 5060 ti's with P2P and PP=4 at 1,4,8,12, and 16 concurrency.
Been trying to troubleshoot prefill around when I saw the newer (supposibly faster) nvfp4 quant. For those curious of prefill issues on 4x5060 ti's, hopefully this helps paint a picture for at what concurrency prefill/ttft starts being an issue. ## info: PP=4 (new unsloth NVFP4 quant) Pipeline para
打开原文相关 2 · 机会 0
中文
Been trying to troubleshoot prefill around when I saw the newer (supposibly faster) nvfp4 quant. For those curious of prefill issues on 4x5060 ti's, hopefully this helps paint a picture for at what concurrency prefill/ttft starts being an issue. ## info: PP=4 (new unsloth NVFP4 quant) Pipeline para
中文全文翻译正在补齐，可先查看下方英文原文。
English / 原文
Been trying to troubleshoot prefill around when I saw the newer (supposibly faster) nvfp4 quant. For those curious of prefill issues on 4x5060 ti's, hopefully this helps paint a picture for at what concurrency prefill/ttft starts being an issue. ## info: PP=4 (new unsloth NVFP4 quant) Pipeline para