May 2026 performance sprint delivered key features in TPU inference path and mamba caching, with crucial bug fixes that improved correctness, memory efficiency, and distributed processing reliability. Notable outcomes include bf16-optimized fused GDN kernels on TPU, DP attention batching/routing improvements for hybrid attention+mamba workloads, bf16 support in mamba ssm cache, and stability fixes in recurrent scan and mosaic tile alignment. These efforts reduce memory footprint, speed up inference, and improve correctness under high-load scenarios across the vLLM ecosystem.

In 2026-04, delivered two major improvements for vllm-project/tpu-inference: (1) Mamba KV cache robustness and memory efficiency, and (2) GDN kernel correctness and stability. These changes enhance inference reliability, memory usage, and accuracy for Qwen3.5 models, with safe concurrent request handling and proper state management across fused kernels.