vLLM v0.26.0: New Inkling models, DeepSeek-V4 speed boosts, and KV offloading maturity

vLLM v0.26.0 is a landmark release with 411 commits from 212 contributors, including 61 first-time contributors. The release introduces full stack support for the new Inkling model family, covering base modeling, piecewise CUDA graphs, Hopper FA4 relative attention, MTP=1 speculative decoding, LoRA, and ModelOpt NVFP4 quantization. For DeepSeek-V4, the team delivered a specialized routing kernel (2.94% E2E TPOT improvement), fused_topk_bias kernel (1.5–2x speedup), and redundant repeat/copy removal (1.8% E2E TPOT). Additional performance work includes ROCm two-stage compressor for HCA prefill, sparse decode/prefill optimizations, and DSpark speculative decoding on AMD and XPU. The fp32 lm_head for generation models (with new head_dtype option) extends to the LoRA path and gains a ROCm torch.mm fast path.

Attention backends are now selectable per KV-cache group, and sliding-window support is an explicit backend capability—improving hybrid model support. KV offloading and tiered secondary storage matured substantially: new offloading metrics, tier-owned event handling, object-store secondary tier with workload identity, DP-replica-aware tiering, and encoder-cache connectors (including CPU offloading). The Rust frontend now handles multimodal video and audio, a Seed-OSS tool parser, and a native vllm-bench port. Model support expands greatly: Inkling family, BertForMaskedLM, RobertaForTokenClassification, LongCat-Flash-Lite n-gram embedding, Cosmos3 Edge Reasoner, TranslateGemma, and more. Transformers 5.13.0 brings migrations for Olmo, MistralLarge3, and HunyuanVL. LoRA improvements include FlashInfer MoE LoRA for BF16, LlavaNextVideo tower/connector LoRA, and optimized TrtLlmLoRAExperts. The engine core gains lower memory for large CUDA graphs, opt-in memory-profiling persistence, and improved tensor loading.