cd /news/ai-infrastructure/vllm-cuts-a-deepseek-v4-1-engram-loo… · home › topics › ai-infrastructure › article
[ARTICLE · art-170248] src=runtimewire.com ↗ pub= topic=ai-infrastructure verified=true sentiment=↑ positive

vLLM cuts a DeepSeek V4.1 Engram lookup from 54 ms to 5.1 ms on eight MI355X GPUs

An AMD contributor filing as Fangzhou Ai merged a vLLM change that cut a DeepSeek V4.1 Engram lookup from 54 ms to 5.1 ms, a 10.6x speedup, on eight MI355X GPUs. The change covers a 589,824-row lookup against a 94 GiB shared table and traces the prefill stall to the order in which GPUs fetched rows from a host-resident table, ordering those requests before they reached host memory. The sort path is enabled on ROCm from 32,768 rows.

read1 min views1 publishedNov 16, 2026
vLLM cuts a DeepSeek V4.1 Engram lookup from 54 ms to 5.1 ms on eight MI355X GPUs
Image: Runtimewire (auto-discovered)

Contributor Fangzhou Ai's 10.6x result covers a 589,824-row lookup against a 94 GiB shared table on eight MI355X GPUs, with the sort path enabled on ROCm from 32,768 rows.

        By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff)
        · Published 

Fangzhou Ai, an AMD contributor filing as [email protected], traced a DeepSeek V4.1 prefill stall on MI355X to the order in which GPUs fetched rows from a host-resident table. The merged vLLM change cut the largest tested lookup from 54 ms to 5.1 ms by ordering those requests before they reached host memory.…

── more in #ai-infrastructure 4 stories · sorted by recency
── more on @vllm 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/vllm-cuts-a-deepseek…] indexed:0 read:1min 2026-11-16 · —