Dynamo released v1.5.1, a patch on top of v1.5.0, fixing router overload recovery, min_tokens on tokenizer-free SGLang decode workers, Qwen3-VL video routing across mixed workers, and reinforcement learning worker discovery on Kubernetes1. The release bounds client-supplied multimodal input across SGLang, TensorRT-LLM, vLLM, and vLLM-Omni paths, and consolidates media fetching onto a single aiohttp client. Frontend runtime libraries were upgraded to fix Anthropic Messages API requests from Claude Code against strict chat templates. Two breaking changes require action on upgrade, including a new DYN_MM_TRUST_EGRESS_PROXY=1 opt-in for workers fetching media through an HTTP proxy.
Dynamo v1.5.1 Patches Router Overload, Multimodal Input Bounds
Dynamo v1.5.1 fixes router overload recovery, bounds multimodal input across SGLang, TensorRT-LLM, vLLM, and vLLM-Omni, and updates frontend libraries…