tests/test_quantization.py

import os, sys; sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "..")))
import torch
from bit_transformer import BitTransformerLM, quantize_dynamic, prepare_qat_fx, convert_qat_fx
from bit_transformer.training import train_loop

def test_qat_matches_dynamic_quant():
    data = torch.randint(0, 2, (16, 8), dtype=torch.long)
    base = BitTransformerLM(d_model=16, nhead=4, num_layers=1, dim_feedforward=32, max_seq_len=8)
    train_loop(base, data, epochs=1, log=False)
    dyn = quantize_dynamic(base)
    qat_model = BitTransformerLM(d_model=16, nhead=4, num_layers=1, dim_feedforward=32, max_seq_len=8)
    qat_model.load_state_dict(base.state_dict())
    prepare_qat_fx(qat_model)
    convert_qat_fx(qat_model)
    inp = torch.randint(0, 2, (10, 8), dtype=torch.long)
    out_dyn, _ = dyn(inp)
    out_qat, _ = qat_model(inp)
    diff = (out_dyn - out_qat).abs().max().item()
    assert diff < 0.6