from transformers import AutoModelForCausalLM, AutoTokenizer
repo_id = "RthItalia/NanoLLM-Qwen2.5-7B-v3.1"
tokenizer = AutoTokenizer.from_pretrained(repo_id, subfolder="nano_compact", use_fast=True)
model = AutoModelForCausalLM.from_pretrained(repo_id, subfolder="nano_compact", trust_remote_code=True, device_map="auto")

Validation against 8-bit reference:

avg cosine: 0.990625
min cosine: 0.98046875
gate: avg >= 0.985

nano_compact/model.safetensors contains Nano quantized tensors and does not require downloading the Qwen base weights.