import json import mlx.core as mx import numpy as np import pytest from h3_mlx.config import DiTConfig from h3_mlx.dit import DiTBlock from h3_mlx.weights import Linear, WeightStore @pytest.mark.parametrize("bits,word", [(4, 0x11111111), (8, 0x01010101)]) def test_preconverted_quantized_linear_loads_without_online_quantization( tmp_path, monkeypatch, bits, word ): # Fixed synthetic quantized matrix: each decoded weight is 0.5. tensors = { "proj.weight": mx.full((4, 64 // (32 // bits)), word, mx.uint32), "proj.scales": mx.full((4, 1), 0.5, mx.bfloat16), "proj.biases": mx.zeros((4, 1), mx.bfloat16), "proj.bias": mx.array([0, 0.25, 0.5, 1], mx.bfloat16), } mx.save_safetensors(str(tmp_path / "core.safetensors"), tensors) (tmp_path / "quantization.json").write_text( json.dumps({"bits": bits, "group_size": 64}) ) def forbidden(*args, **kwargs): pytest.fail("Inference must never quantize weights") monkeypatch.setattr(mx, "quantize", forbidden) layer = Linear.load(WeightStore(tmp_path), "proj") x = mx.ones((3, 64), mx.bfloat16) expected = mx.array([[32, 32.25, 32.5, 33]] * 3, mx.bfloat16) np.testing.assert_array_equal( np.asarray(layer(x).view(mx.uint16)), np.asarray(expected.view(mx.uint16)) ) def test_missing_packed_block_cannot_trigger_conversion(tmp_path): class NormStore: def get(self, key): assert ( key.endswith("norm1.weight") or key.endswith("norm2.weight") or key.endswith("_norm.weight") ) return mx.ones(32, mx.bfloat16) missing = tmp_path / "missing" with pytest.raises(FileNotFoundError, match="preconverted"): DiTBlock(NormStore(), "blocks.0", DiTConfig(), packed_cache=missing) assert not missing.exists()