Vulkan top_k radix select arrives for long rows

What shipped: The standout is llama.cpp's Vulkan top_k radix select for k >= 1024, targeting Qwen 3.8 Flash Next [1]. Backend work also includes HIP Q2_0 dot-product optimization for gfx1201 [2], RDNA3 mmq tuning [3], and a new SWIGLU_CLAMP ggml op [4]. vLLM shipped a security fix that bounds the validation-error response body [5], alongside AutoRound Format Block-Wise FP8 support [6] and speculative decoding CUDA graph memory-safety fixes [7]. Multiple diffusion model fusions landed in SGLang: Qwen-Image output projection bias absorption [8], FLUX.2 gated residual normalization on Blackwell [9], and Qwen-Image FP8 QKV projection [10].

Why it matters: The vLLM security fix is the only action item; bound your error response sizes. For performance, the Vulkan top_k radix select promises faster sampling for long-row cases, and the HIP Q2_0 optimization benefits RDNA3 users. The diffusion fusions in SGLang are prebuilt optimizations that should speed up inference for those models. Ollama now honors model generation defaults [11], which means the server may override client-provided settings unless you explicitly send them.

Also: vLLM added LoRA and fusion support in its kernels (not listed above, but referenced), speculative decoding under structured output in Ollama , and a fast-start zero-copy mapping . Git was not in scope, but note: vLLM's ROCm dual-stream decode with hipgraphs improves throughput on AMD. SGLang extended speculative decoding with LFM2 and LFM2-MoE , and llama.cpp added SYCL free-memory reporting .