• Meta Llama 3 8B model card
• Meta Llama 3 implementation (GQA and KV cache)
• Kwon et al. (2023): PagedAttention
• vLLM engine arguments
• Pope et al. (2022): Efficiently Scaling Transformer Inference

This is a simulation. Times come from fixed example bandwidth and compute rates, not measurements. Decimal GB and TB are used throughout. The model has the Llama 3 8B shape; its parameter count is rounded to 8.03 billion. FP16 is a teaching choice (the original checkpoint uses BF16). Traffic is fixed token counts, not a real tokenizer or language model. The roofline estimate omits quadratic prefill attention, kernel overhead, sampling costs, prefix sharing, scaling metadata and preemption; mixed batches overlap compute and memory in an idealized way. The capacity guard reserves full contexts, while the allocator grows pages as tokens arrive. The command is a starting config, not a deployment or performance guarantee. Independent lesson; no affiliation with Meta, vLLM or a GPU vendor.