This tutorial details the deployment of Llama 3 on NVIDIA H100 (Hopper) and RTX Pro 6000 (Ada Lovelace) architectures using the TensorRT-LLM library to optimize inference performance. The process leverages native FP8 precision via fourth-generation Tensor Cores, which halves the model's memory footprint compared to FP16, thereby increasing model density and expanding the paged KV cache for larger batch sizes. The deployment workflow involves pulling official NGC containers, downloading weights via the Hugging Face CLI, and a two-step optimization process: converting checkpoints to an optimized format using a calibration size of 1024 to maintain accuracy, and compiling the final engine using the trtllm-build tool. Key technical configurations include setting the gemm_plugin to fp8, a maximum batch size of 64, and sequence lengths up to 8192. Once the engine is built, the model is served through the NVIDIA Triton Inference Server to handle HTTP/gRPC requests. This stack utilizes "in-flight batching" to maximize GPU utilization by inserting new requests into the batch during every generation step. While the H100 offers up to 3,958 FP8 TFLOPS for massive MoE models, the RTX Pro 6000 provides a cost-effective 48GB VRAM alternative for smaller models like Llama 3 8B.