Gemma 4 Runs 1.79x Faster on Tesla T4 Using QAT Int4 Weights Over bf16

A developer has published a step-by-step guide for deploying Google's Gemma 4 2B model on a Tesla T4 GPU hosted on a Google Compute Engine VM. Two model checkpoints were benchmarked: the standard bf16 version and a QAT int4-weight variant, using vLLM 0.29.0 on a modest n1-standard-2 instance with 2 vCPUs and 7.8 GB RAM. The QAT model decoded at 72.31 tokens per second in a single stream compared to 40.44 for bf16, and reached 215.91 tokens per second across eight streams versus 164.62. A suite of Python MCP tools was built to manage the deployment pipeline, handling steps such as GPU architecture verification and a required Turing-specific patch. The full setup and source code are available on GitHub, offering a practical reference for running modern LLMs on older, lower-cost GPU hardware.