CUDA NVRTC (runtime compilation library)

Distinct from CUDA itself and from nvcc: NVRTC is the LIBRARY that lets a host program compile CUDA C++ at run time, which is what makes shape- or config-specialised kernels possible without shipping a prebuilt binary per shape. That property is load-bearing for inference engines precisely because decode is shape-variable. The cost lands on the cold-start path, and this repo already owns the measurement: ev-trtllm-pr-18542-fmha-jit-warmup records TensorRT-LLM compiling FMHA generation kernels with NVRTC synchronously inside the first warmup forward, ~5-11 s per kernel variant depending on host CPU and 80-160 s in total on DeepSeek-family MLA fp8-KV configs, with an opt-in async-warmup flag cutting an internal B200 measurement from 163 s to ~26-30 s. THREE FIELDS ARE DELIBERATELY EMPTY, and each is an honest gap rather than a missing value: (1) input_languages is empty because the schema's vocabulary enumerates model and DSL front ends (python, jax, pytorch, tensorflow, onnx, tflite, cutlass-cpp, triton, mlir, xla-hlo) and has no term for C++/CUDA C++ source - NVRTC sits BELOW that vocabulary. Inventing a term would change an existing field's meaning, which AGENTS.md forbids; the enum needs extending instead. (2) compile_time_seconds is null rather than a midpoint of the 5-11 s per-kernel range above, because that range is a single vendor-reported measurement in one engine on one class of config, and this schema has no *_basis field in which to record the derivation of a number taken from a range. (3) first_release_year is null: no reachable primary source states when NVRTC first shipped, and CUDA-version-to-year mapping is exactly the kind of inference AGENTS.md's contract exists to prevent. category is compiler-framework as the least-wrong of the six available values, but it is genuinely imprecise: NVRTC is neither a framework (it composes nothing) nor a DSL, template library, graph optimizer, or runtime. The enum describes compiler SHAPES and has no slot for a plain single-language code generator, which is what a deployment planner most needs to distinguish. autotuning is false - NVRTC compiles the program it is given and never searches tactics.