vLLM agentic serving details: packed KV saves 10% memory, Kimi K3 decode throughput up 2.7x

vllm_project · x · 2026-09-09

Parts 4-6 of vLLM's agentic serving thread detail the stack-level work: a packed KV layout for DeepSeek V4 saves 10% KV memory and reduces 92 tensors per block to 1; decode context parallelism delivers 2.7x decode throughput for Kimi K3 at the same TPOT, while prefill context parallelism runs DeepSeek V4's sparse MLA 2.65x faster than head sharding; simply capping prefill breaks head-of-line blocking for +93% TPGS and 2.3x better p90 interactivity. Real coding-agent traces show 43 median turns, 142K-token inputs vs. 444-token outputs, 96%+ prefix-cache hits, and 44% of sessions forking subagents.

Related event: vLLM Unveils AgentX Benchmark: Up to 106x Cost Advantage for Real-World Agentic Serving(5 posts)→

Original post →

More from coding & agent

coding & agent channel →