vLLM v0.26.0 adds Inkling model family support, DeepSeek-V4 optimizations, and Rust frontend multimodal features

The vLLM project published release v0.26.0 with 411 commits from 212 contributors, including 61 new contributors. The update introduces native support for the Inkling model family, including base modeling, piecewise CUDA graph support, Hopper FA4 relative attention, MTP=1 speculative decoding, LoRA, and ModelOpt NVFP4 quantization.

Performance optimizations for DeepSeek-V4 include a specialized routing kernel (2.94% end-to-end token-per-second improvement), fused_topk_bias kernels (1.5–2x speedup), and redundant repeat/copy removal (1.8% end-to-end token-per-second improvement). Additional hardware-specific improvements include ROCm two-stage compressor support for HCA prefill, sparse decode/prefill optimizations, and DSpark speculative decoding on AMD and XPU.

The release adds flexible attention backends, allowing selection per KV-cache group and explicit sliding-window support as a backend capability, which improves support for hybrid models. KV offloading and tiered secondary storage matured with features such as object-store secondary tier with workload identity, DP-replica-aware tiering, and encoder-cache connectors including CPU offloading.

Rust frontend gained multimodal capabilities with video and audio support, a Seed-OSS tool parser, and a native vllm-bench port. Transformers backend support was updated to include Olmo/Olmo2, MistralLarge3, and HunyuanVL models. Additional engine core improvements include fp32 lm_head for generation models, memory profiling persistence, and KV-cache layout refactoring.