Full threadtikkun·TensorRT-LLM plus int8 quantization = cheaper, minimal quality loss. Some nice tables in the postView on HN