vllm.cpp

A C++ reimplementation of the vLLM engine (continuous batching, paged KV cache, GGUF loading) with CUDA, CPU, Metal, Vulkan and ROCm backends plus an opt-in Tenstorrent backend under `src/vt/tenstorrent`. The TT backend is a thin adapter over TTNN and TT-Metalium built with `-DVLLM_CPP_TENSTORRENT=ON`, adding paged attention, Qwen3.5 gated-delta-net, trace capture and quant-preserving (`keepquant`) matmul paths. Status is correctness-first on Blackhole: OPT-125m passes a strict token-exact gate, Qwen3-0.6B has committed goldens with a full rerun pending, and 27B GGUF decode is in smoke-measurement stage.