v0.5.0 - third engine measured
TensorRT-LLM activated as measured third engine (2026-07-16): PyTorch and compiled-TRT backends with backend as a swept axis, in-framework trtllm-build cache-by-hash, quantisation preflight, GPU compute-capability gates, per-engine GPU CI matrix. Historical release; no distribution artifacts.