llama.cpp b10419: OpenVINO adds Qwen3.5, fixes Phi-3 perplexity 4000x

The latest llama.cpp release, b10419 from ggml-org, is a substantial update focused on the OpenVINO backend, which runs LLMs on Intel CPUs, GPUs, and NPUs. Key additions include support for Qwen3.5, enabling gpt-oss mixture-of-experts (MoE) models, and the new mxFP4 (microscaling 4-bit floating point) format for efficient low-precision inference. The release also streamlines OpenVINO graph compilation with dynamic shape inference, unique node naming, and expanded operator coverage (FILL, SIGMOID, SQR, SQRT), plus dozens of bug fixes for architectures like Gemma3n, MPT, Kimi-linear, and MiniMax-M3.

The most dramatic fix addresses a perplexity disaster in the NEOX RoPE rotary position embedding path on GPU stateful mode. A mixed-rank Multiply fed rank-3 data against rank-4 cos/sin tables, causing the OpenVINO GPU plugin to miscalculate the broadcast and corrupt Q/K rotation. On Phi-3-mini Q4_K_M, wiki.test perplexity plummeted from 27120.43 to 6.2263 — just a hair above the CPU reference of 6.2251. Additionally, the new GGML_OPENVINO_RELEASE_WEIGHTS environment variable lets users reclaim host memory after OpenVINO uploads weights to the GPU, significantly reducing RSS for long-running inference servers. For developers deploying on Intel hardware, this release makes local generation both more accurate and more memory-efficient.