|
1 |
| -import unittest |
| 1 | +from contextlib import nullcontext |
2 | 2 |
|
3 | 3 | import pytest
|
4 | 4 |
|
5 |
| -from llmcompressor.modifiers.factory import ModifierFactory |
6 |
| -from llmcompressor.modifiers.quantization import QuantizationModifier |
7 |
| -from tests.llmcompressor.modifiers.conf import setup_modifier_factory |
| 5 | +from llmcompressor.modifiers.quantization import GPTQModifier |
8 | 6 |
|
9 | 7 |
|
10 |
| -@pytest.mark.unit |
11 |
| -class TestQuantizationRegistered(unittest.TestCase): |
12 |
| - def setUp(self): |
13 |
| - setup_modifier_factory() |
14 |
| - self.kwargs = dict( |
15 |
| - index=0, group="quantization", start=2.0, end=-1.0, config_groups={} |
| 8 | +@pytest.fixture |
| 9 | +def q_config_kwargs(config_0, config_1): |
| 10 | + return dict( |
| 11 | + config_groups=dict( |
| 12 | + group_0=dict( |
| 13 | + targets=["Linear"], |
| 14 | + input_activations=dict(num_bits=8, symmetric=False, strategy="token"), |
| 15 | + weights=dict( |
| 16 | + num_bits=4, |
| 17 | + symmetric=True, |
| 18 | + strategy="group", |
| 19 | + group_size=128, |
| 20 | + actorder=config_0, |
| 21 | + ), |
| 22 | + ), |
| 23 | + group_1=dict( |
| 24 | + targets=["Linear"], |
| 25 | + input_activations=dict(num_bits=8, symmetric=False, strategy="token"), |
| 26 | + weights=dict( |
| 27 | + num_bits=4, |
| 28 | + symmetric=True, |
| 29 | + strategy="group", |
| 30 | + group_size=128, |
| 31 | + actorder=config_1, |
| 32 | + ), |
| 33 | + ), |
16 | 34 | )
|
| 35 | + ) |
17 | 36 |
|
18 |
| - def test_quantization_registered(self): |
19 |
| - quant_obj = ModifierFactory.create( |
20 |
| - type_="QuantizationModifier", |
21 |
| - allow_experimental=False, |
22 |
| - allow_registered=True, |
23 |
| - **self.kwargs, |
24 |
| - ) |
25 | 37 |
|
26 |
| - self.assertIsInstance(quant_obj, QuantizationModifier) |
| 38 | +@pytest.mark.parametrize( |
| 39 | + "has_actorder,actorder,config_0,config_1,expected_0,expected_1", |
| 40 | + [ |
| 41 | + # defaults to None if nothing provided |
| 42 | + (False, None, None, None, None, None), |
| 43 | + # modifier overrides config if no config provided |
| 44 | + (True, "group", None, None, "group", "group"), |
| 45 | + # modifier overrides if config partially matches anyways |
| 46 | + (True, "group", None, "group", "group", "group"), |
| 47 | + (True, "group", "group", None, "group", "group"), |
| 48 | + # modifier errors if conflict with config |
| 49 | + (True, "group", None, "static", "error", "error"), |
| 50 | + (True, "group", "static", None, "error", "error"), |
| 51 | + # modifier does not override if not provided |
| 52 | + (False, "N/A", None, None, None, None), |
| 53 | + (False, "N/A", None, "static", None, "static"), |
| 54 | + (False, "N/A", "static", None, "static", None), |
| 55 | + (False, "N/A", "static", "static", "static", "static"), |
| 56 | + (False, "N/A", None, "group", None, "group"), |
| 57 | + (False, "N/A", "group", None, "group", None), |
| 58 | + (False, "N/A", "group", "group", "group", "group"), |
| 59 | + ], |
| 60 | +) |
| 61 | +def test_actorder_resolution( |
| 62 | + has_actorder, actorder, q_config_kwargs, expected_0, expected_1 |
| 63 | +): |
| 64 | + if has_actorder: |
| 65 | + modifier = GPTQModifier(**q_config_kwargs, actorder=actorder) |
| 66 | + else: |
| 67 | + modifier = GPTQModifier(**q_config_kwargs) |
| 68 | + |
| 69 | + with pytest.raises(ValueError) if expected_0 == "error" else nullcontext(): |
| 70 | + resolved = modifier.resolve_quantization_config() |
| 71 | + |
| 72 | + if expected_0 != "error": |
| 73 | + assert resolved.config_groups["group_0"].input_activations.actorder is None |
| 74 | + assert resolved.config_groups["group_0"].weights.actorder == expected_0 |
| 75 | + assert resolved.config_groups["group_1"].input_activations.actorder is None |
| 76 | + assert resolved.config_groups["group_1"].weights.actorder == expected_1 |
0 commit comments