vLLM AFD tests show 11.3% higher decode throughput and 47% lower TTFT
vllm_project · x · 2026-07-24
This follow-up explains why attention and FFN should be split in MoE serving: attention capacity tracks request state and KV-cache pressure, while expert capacity tracks token routing and load.
The post adds early controlled results on DeepSeek-V3.2 W8A8 with Ascend 910C:
- A 64A/16F layout improves per-die decode throughput by +11.3% at 16K and +9.0% at 32K versus an EP64 baseline.
- CAM async prefill reduces median TTFT by about 47% — from 15.1s to 8.0s — at 12 req/s.
It also notes an important caveat: disaggregation is not always beneficial, and the attention-to-FFN ratio matters; 48A16F falls below baseline.
Related event: vLLM Releases Experimental AFD Plugin to Boost MoE Inference(3 posts)→
More from Infra
- Cloud release velocity scores: AWS 95, Azure 90, Google Cloud 82 — DavidLinthicum · 2026-10-27
- Selling a 5090 to scale local inference: one business weighs 5 hardware paths — Anarchaotic · 2026-10-27
- AWS tutorial: fleet-wide edge CV model lifecycle with SageMaker SDK v3, ~$0.07 per run — AWS ML Blog · 2026-10-27
- Qwen3.8-27B hits 40 tok/s on Jetson Thor with INT4 + MTP, accuracy beats FP16 — No-Dragonfly6246 · 2026-10-27
- Celesto launches CI runners nearly 90% cheaper than GitHub Actions, with agent sandboxes — aniketmaurya · 2026-10-27
- Supabase founders on rebuilding the backend for a billion AI-built apps — Latent Space · 2026-10-27