vLLM 0.27.0 adds Kimi K3, Qwen3.5, 2x DeepSeek-V4 kernel gains

vLLM v0.27.0, the latest release of the open-source LLM inference engine, brings 561 commits from 242 contributors (64 new). The marquee feature is full Kimi K3 support landing in one release: core model files and kernels, Python and Rust frontends, AttnRes kernels, DeepGEMM support, compressed-tensors quantized checkpoints, DSpark AR fusion, and optional shared-expert sharding instead of replication. That means vLLM users can deploy Moonshot's Kimi K3 at production scale with native optimizations from day one.

Beyond Kimi, the release adds Qwen3.5 text-only dense and MoE models with EVS video token pruning, K-EXAONE-2.0-750B-A37B, VaultGemma via the Transformers backend, plus jina-embeddings-v5-text-nano for embeddings. Performance-focused upgrades target DeepSeek-V4: sequence parallelism, a ~2x kernel speedup by skipping empty c128 launches, 3.4% end-to-end TTFT reduction from skipping unneeded topk/router calls, and 3.9% TTFT gains from workspace reuse. Memory savings include 448 MiB freed in the PP buffer and a compact MXFP4 indexer KV cache. The framework also upgrades to PyTorch 2.13.0, torchvision 0.28.0, and Triton 3.7.1 (a breaking environment change). FlashAttention 4 integration deepens on SM100 with FP8 KV cache and headdim-256 support, backed by new JIT warmup infrastructure that removes first-request compilation stalls. Model Runner V2 now supports encoder-only attention, sequence pooling for embeddings/classification, and BGE-M3 pooling. For resilience, a simplified fault tolerance framework for DP+EP external load-balancer deployments and async elastic EP scaling are introduced. Early next-gen hardware enablement includes NVIDIA Rubin sm_107 targets with NVLink all-reduce paths and ROCm gfx1250 support.