Thanks, @Stefano Fiorucci - anakin87, your suggestion solved the problem. The right way to run the generator is:

generator = HuggingFaceLocalGenerator(model="meta-llama/Meta-Llama-3-70B-Instruct",
                                      generation_kwargs={"max_new_tokens": 100},
                                      huggingface_pipeline_kwargs={"device_map": "auto" 
                                                                   })

That way, i can see in the nvidia-smi output that the load is distributed among all GPUs (all are loaded to ±50%), and the model runs without crushing.