#vllm — Hans Christian Thjømøe

Close-up of GPU server racks with cabling in a data center, illustrating production inference deployment.

August 15, 2026

vLLM 0.27 Removes First-Request Stalls — But Torch 2.13 Is Breaking

vLLM 0.27.0 (Aug 10) ships JIT warmup and runner-owned Triton warmup that eliminate first-request compile stalls, plus Kimi K3 support — but the torch 2.13.0 upgrade is a breaking change that needs migration testing before you bump.

Close-up of a black Gigabyte graphics card — the kind of GPU that runs Qwen3.8-27B at 4-bit in 24 GB of VRAM

August 11, 2026

Hybrid-Attention Models Finally Run Offline in vLLM

vLLM v0.27.0 lets hybrid SWA+full attention models run through offline inference, closing the gap between eval and serving — plus a breaking torch 2.13 upgrade.

GPU server racks in a dim data center with cooling fans spinning

August 5, 2026

Retune Draft Models Without Restarting vLLM

vLLM v0.26.0 adds runtime draft-weight updates, so you can retune speculative decoding without restarting the engine and flushing the KV cache.

An hourglass counting down on a desk, sand falling through the narrow neck

May 18, 2026

Your vLLM Thinking Budget Was Doing Nothing With MTP On

vLLM 0.21.0 shipped Friday with a quiet fix: thinking_token_budget was being silently ignored when MTP speculative decoding was enabled. If you serve reasoning models with spec decode, you have been paying for it.

← All posts