vllm

Key Features

• Supports the full Kimi K3 stack. This includes core model files and kernels, Python and Rust frontends, AttnRes kernels, DeepGEMM, compressed-tensors quantized checkpoints, and DSpark AR fusion.
• Provides an option to shard shared experts in Kimi K3 instead of replicating them.
• Supports Qwen3.5 text-only dense and MoE models, and adds EVS video token pruning.
• Supports K-EXAONE-2.0-750B-A37B, VaultGemma via the Transformers modeling backend, and jina-embeddings-v5-text-nano.
• Model Runner V2 now supports encoder-only attention, sequence pooling for embedding and classification, encoder token classification, token embedding, and BGE-M3 pooling, extending beyond generative tasks.
• Adds CPU multimodal processing and multi-layer MTP speculator support to Model Runner V2.

Compatibility & Breaking Changes

• Upgraded to PyTorch 2.13.0, torchvision 0.28.0, and Triton 3.7.1.
• This PyTorch upgrade is a breaking change to environment compatibility. XPU and CPU have also been updated to torch 2.13.

Performance Improvements

• Strengthened FlashAttention 4's SM100 support to support FP8 KV cache and headdim-256.
• Introduced JIT warmup infrastructure and runner-owned Triton kernel warmup to reduce the compilation delay that occurred on the first request.
• Added sequence parallelism to DeepSeek-V4.
• Improved kernel and E2E TTFT performance for DeepSeek-V4 by skipping empty c128 kernel launches, removing unnecessary topk/router computation, reusing workspace, eliminating duplicate full kernel execution, and applying adaptive topk width.
• Saved 448 MiB of GPU memory in DeepSeek-V4's PP buffer and added a compact MXFP4 indexer KV cache.
• Removed sparse-MLA q-head padding on FlashInfer >= 0.6.14.

Other

• The provided changelog does not specify any separate major deprecation or removal items.