|
| 1 | +# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved. |
| 2 | +# |
| 3 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | +# you may not use this file except in compliance with the License. |
| 5 | +# You may obtain a copy of the License at |
| 6 | +# |
| 7 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | +# |
| 9 | +# Unless required by applicable law or agreed to in writing, |
| 10 | +# software distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | +# See the License for the specific language governing permissions and |
| 13 | +# limitations under the License. |
| 14 | + |
| 15 | +""" |
| 16 | +End-to-end tests for asymmetric quantization with zero-point decompression. |
| 17 | +""" |
| 18 | + |
| 19 | +import pytest |
| 20 | +import torch |
| 21 | +from compressed_tensors.compressors.model_compressors.model_compressor import ( |
| 22 | + ModelCompressor, |
| 23 | +) |
| 24 | +from compressed_tensors.config import CompressionFormat |
| 25 | +from compressed_tensors.quantization import ( |
| 26 | + QuantizationArgs, |
| 27 | + QuantizationConfig, |
| 28 | + QuantizationScheme, |
| 29 | + QuantizationStrategy, |
| 30 | + apply_quantization_config, |
| 31 | +) |
| 32 | +from torch.nn import Linear, Module |
| 33 | + |
| 34 | + |
| 35 | +class SimpleModel(Module): |
| 36 | + """Simple model for testing""" |
| 37 | + |
| 38 | + def __init__(self, input_dim=512, hidden_dim=256, output_dim=128): |
| 39 | + super().__init__() |
| 40 | + self.layer1 = Linear(input_dim, hidden_dim, bias=False) |
| 41 | + self.layer2 = Linear(hidden_dim, output_dim, bias=False) |
| 42 | + |
| 43 | + def forward(self, x): |
| 44 | + x = self.layer1(x) |
| 45 | + x = torch.relu(x) |
| 46 | + x = self.layer2(x) |
| 47 | + return x |
| 48 | + |
| 49 | + |
| 50 | +def create_asymmetric_quant_config( |
| 51 | + num_bits=4, strategy=QuantizationStrategy.GROUP, group_size=128 |
| 52 | +) -> QuantizationConfig: |
| 53 | + """Create an asymmetric quantization config""" |
| 54 | + config_groups = { |
| 55 | + "group_1": QuantizationScheme( |
| 56 | + targets=["Linear"], |
| 57 | + weights=QuantizationArgs( |
| 58 | + num_bits=num_bits, |
| 59 | + strategy=strategy.value, |
| 60 | + group_size=( |
| 61 | + group_size if strategy == QuantizationStrategy.GROUP else None |
| 62 | + ), |
| 63 | + symmetric=False, |
| 64 | + ), |
| 65 | + ), |
| 66 | + } |
| 67 | + return QuantizationConfig(config_groups=config_groups) |
| 68 | + |
| 69 | + |
| 70 | +@pytest.mark.parametrize( |
| 71 | + "strategy,group_size", |
| 72 | + [ |
| 73 | + (QuantizationStrategy.GROUP, 128), |
| 74 | + (QuantizationStrategy.CHANNEL, None), |
| 75 | + ], |
| 76 | +) |
| 77 | +def test_end_to_end_asymmetric_quantization( |
| 78 | + strategy, |
| 79 | + group_size, |
| 80 | + mock_per_group_calibration, |
| 81 | + mock_per_channel_calibration, |
| 82 | +): |
| 83 | + """ |
| 84 | + Test end-to-end workflow: quantize -> compress -> decompress in memory |
| 85 | + """ |
| 86 | + model = SimpleModel() |
| 87 | + original_weights = { |
| 88 | + "layer1": model.layer1.weight.detach().clone(), |
| 89 | + "layer2": model.layer2.weight.detach().clone(), |
| 90 | + } |
| 91 | + |
| 92 | + quant_config = create_asymmetric_quant_config( |
| 93 | + num_bits=4, strategy=strategy, group_size=group_size |
| 94 | + ) |
| 95 | + # Set pack-quantized format for ModelCompressor usage |
| 96 | + quant_config.format = CompressionFormat.pack_quantized.value |
| 97 | + apply_quantization_config(model, quant_config) |
| 98 | + |
| 99 | + if strategy == QuantizationStrategy.GROUP: |
| 100 | + mock_per_group_calibration( |
| 101 | + model.layer1, "weight", model.layer1.weight, group_size |
| 102 | + ) |
| 103 | + mock_per_group_calibration( |
| 104 | + model.layer2, "weight", model.layer2.weight, group_size |
| 105 | + ) |
| 106 | + else: |
| 107 | + mock_per_channel_calibration(model.layer1, "weight", model.layer1.weight) |
| 108 | + mock_per_channel_calibration(model.layer2, "weight", model.layer2.weight) |
| 109 | + |
| 110 | + # Compress and decompress in memory using ModelCompressor |
| 111 | + mc = ModelCompressor(quantization_config=quant_config) |
| 112 | + mc.compress_model(model) |
| 113 | + |
| 114 | + # Verify compression created zero-point parameters |
| 115 | + assert hasattr(model.layer1, "weight_zero_point") |
| 116 | + assert hasattr(model.layer2, "weight_zero_point") |
| 117 | + assert model.layer1.weight_zero_point.dtype == torch.int32 |
| 118 | + assert model.layer2.weight_zero_point.dtype == torch.int32 |
| 119 | + |
| 120 | + # Decompress in memory |
| 121 | + mc.decompress_model(model) |
| 122 | + |
| 123 | + # Verify decompression restored weights correctly |
| 124 | + assert model.layer1.weight.shape == original_weights["layer1"].shape |
| 125 | + assert model.layer2.weight.shape == original_weights["layer2"].shape |
| 126 | + assert model.layer1.weight.dtype.is_floating_point |
| 127 | + assert model.layer2.weight.dtype.is_floating_point |
| 128 | + assert not torch.isnan(model.layer1.weight).any() |
| 129 | + assert not torch.isnan(model.layer2.weight).any() |
| 130 | + assert not torch.isinf(model.layer1.weight).any() |
| 131 | + assert not torch.isinf(model.layer2.weight).any() |
| 132 | + |
| 133 | + |
| 134 | +@pytest.mark.parametrize("num_bits", [4, 8]) |
| 135 | +def test_asymmetric_quantization_accuracy(num_bits, mock_per_group_calibration): |
| 136 | + """ |
| 137 | + Test that asymmetric quantization with zero-point preserves accuracy better |
| 138 | + than symmetric quantization for biased weight distributions. |
| 139 | + """ |
| 140 | + shape = (256, 512) |
| 141 | + biased_weights = torch.randn(shape) + 2.0 |
| 142 | + |
| 143 | + quant_config = create_asymmetric_quant_config( |
| 144 | + num_bits=num_bits, |
| 145 | + strategy=QuantizationStrategy.GROUP, |
| 146 | + group_size=128, |
| 147 | + ) |
| 148 | + quant_config.format = CompressionFormat.pack_quantized.value |
| 149 | + |
| 150 | + class SingleLayer(Module): |
| 151 | + def __init__(self): |
| 152 | + super().__init__() |
| 153 | + self.layer = Linear(shape[1], shape[0], bias=False) |
| 154 | + |
| 155 | + model = SingleLayer() |
| 156 | + apply_quantization_config(model, quant_config) |
| 157 | + |
| 158 | + with torch.no_grad(): |
| 159 | + model.layer.weight.copy_(biased_weights) |
| 160 | + mock_per_group_calibration(model.layer, "weight", model.layer.weight, 128) |
| 161 | + |
| 162 | + # Compress and decompress in memory using ModelCompressor |
| 163 | + mc = ModelCompressor(quantization_config=quant_config) |
| 164 | + mc.compress_model(model) |
| 165 | + mc.decompress_model(model) |
| 166 | + |
| 167 | + decompressed_weights = model.layer.weight |
| 168 | + assert decompressed_weights.shape == shape |
| 169 | + assert not torch.isnan(decompressed_weights).any() |
| 170 | + assert not torch.isinf(decompressed_weights).any() |
| 171 | + threshold = torch.std(torch.rand(shape) - torch.rand(shape)) |
| 172 | + assert torch.std(biased_weights - decompressed_weights) < threshold |
0 commit comments