From RuntimeWire · RuntimeWire · · 30s
vLLM speeds DeepSeek V4.1 lookup
vLLM cuts DeepSeek V4.1 engram lookup from 54ms to 5.1ms on eight MI355X GPUs.
In brief
vLLM cuts DeepSeek V4.1 engram lookup from 54ms to 5.1ms on eight MI355X GPUs. vLLM reduced DeepSeek V4.1 engram lookup time from 54ms to 5.1ms using eight MI355X GPUs. Originally reported by RuntimeWire.
vLLM change

vLLM cut DeepSeek V4.1 engram lookup from 54 ms to 5.1 ms on eight MI355X GPUs.
Fangzhou Ai fix
Fangzhou Ai traced DeepSeek V4.1 prefill stall on MI355X to GPU row fetch order.
vLLM solution
vLLM ordered GPU requests before host memory to cut lookup time.
The bottom line
vLLM reduced DeepSeek V4.1 engram lookup time from 54ms to 5.1ms using eight MI355X GPUs.
Read the TL;DR — this story on one card.





