I am trying to run local inference on a 7B parameter open-source LLM using `transformers` and `torch` on a GPU with 12GB VRAM.
When loading the standard model weights in FP16/FP32 precision using `AutoModelForCausalLM.from_pretrained()`, the execution fails immediately with `torch.cuda.OutOfMemoryError: CUDA out of memory`.
What is the standard configuration required to load and run inference on these models efficiently without exceeding VRAM limits?