llamaperf

DeepSeek V4 Flash 284B (13B active)

on GH200 · vLLM · 1,048,576 ctx

Tone: positive
Sep 7, 2026
Throughput
171.0 t/s gen
Quant
FP4+FP8
KV cache
FP8
System RAM
480 GB
VRAM reported
144 GB

Summary

User reports 64 t/s baseline and 171 t/s with DSpark k=5, a 2.7x speedup. Setup is vLLM nightly 0.27.2rc1.dev122, with 88 GB of experts offloaded to LPDDR5x via UVA.