|
| 1 | +# Owner(s): ["module: inductor"] |
| 2 | + |
| 3 | +import logging |
| 4 | + |
| 5 | +import numpy as np |
| 6 | + |
| 7 | +import torch |
| 8 | +import torch._inductor |
| 9 | +import torch._inductor.fx_passes.group_batch_fusion |
| 10 | +from torch._dynamo.utils import counters |
| 11 | +from torch._inductor.test_case import run_tests, TestCase |
| 12 | +from torch.testing._internal.common_utils import IS_LINUX |
| 13 | +from torch.testing._internal.inductor_utils import GPU_TYPE, HAS_GPU, requires_gpu |
| 14 | + |
| 15 | + |
| 16 | +log = logging.getLogger(__name__) |
| 17 | + |
| 18 | + |
| 19 | +class TargetCPModule(torch.nn.Module): |
| 20 | + def __init__(self): |
| 21 | + super().__init__() |
| 22 | + |
| 23 | + def forward(self, x1, x2): |
| 24 | + relued = torch.relu(x1) |
| 25 | + tanhed = torch.tanh(relued) |
| 26 | + tensor = torch.matmul( |
| 27 | + tanhed, |
| 28 | + x2, |
| 29 | + ) |
| 30 | + return tensor |
| 31 | + |
| 32 | + |
| 33 | +class FeedforwardNN(torch.nn.Module): |
| 34 | + def __init__(self): |
| 35 | + super().__init__() |
| 36 | + self.fc1 = torch.nn.Linear(1, 64) |
| 37 | + self.fc2 = torch.nn.Linear(64, 64) |
| 38 | + self.fc3 = torch.nn.Linear(64, 64) |
| 39 | + self.fc4 = torch.nn.Linear(64, 1) |
| 40 | + |
| 41 | + def forward(self, x): |
| 42 | + x = torch.relu(self.fc1(x)) |
| 43 | + tanh_x = torch.tanh(x) |
| 44 | + x = torch.relu(self.fc2(x)) |
| 45 | + x = torch.relu(self.fc3(tanh_x)) |
| 46 | + x = self.fc4(x) |
| 47 | + return x |
| 48 | + |
| 49 | + |
| 50 | +class TestQuantization(TestCase): |
| 51 | + def compare_dict_tensors(self, ref_dict, res_dict, rtol=1e-3, atol=1e-3): |
| 52 | + if len(set(ref_dict.keys())) != len(set(res_dict.keys())): |
| 53 | + return False |
| 54 | + for key1 in ref_dict.keys(): |
| 55 | + key2 = "_orig_mod." + key1 |
| 56 | + assert key2 in res_dict, f"{key1} does not exist in traced module" |
| 57 | + # if both of them are None, continue |
| 58 | + if ( |
| 59 | + not isinstance(ref_dict[key1], torch.Tensor) |
| 60 | + and not isinstance(res_dict[key2], torch.Tensor) |
| 61 | + and ref_dict[key1] is None |
| 62 | + and res_dict[key2] is None |
| 63 | + ): |
| 64 | + log.info( |
| 65 | + "None found with key1 and value 1: %s, %s, key2 and value2 %s, %s", |
| 66 | + key1, |
| 67 | + ref_dict[key1], |
| 68 | + key2, |
| 69 | + res_dict[key2], |
| 70 | + ) |
| 71 | + continue |
| 72 | + elif not torch.allclose( |
| 73 | + ref_dict[key1], res_dict[key2], rtol=rtol, atol=atol, equal_nan=True |
| 74 | + ): |
| 75 | + log.info( |
| 76 | + "gradient mismatch for eager and compiled modules, with eager: %s and compiled: %s", |
| 77 | + ref_dict[key1], |
| 78 | + res_dict[key2], |
| 79 | + ) |
| 80 | + return False |
| 81 | + return True |
| 82 | + |
| 83 | + def compare_pred(self, module, traced, input, rtol=1e-3, atol=1e-3): |
| 84 | + ref = module(*input) |
| 85 | + res = traced(*input) |
| 86 | + self.assertEqual(ref, res, rtol=rtol, atol=atol) |
| 87 | + |
| 88 | + def compare_parameters(self, module, traced, rtol=1e-3, atol=1e-3): |
| 89 | + ref_params = dict(module.named_parameters()) |
| 90 | + res_params = dict(traced.named_parameters()) |
| 91 | + self.assertTrue(self.compare_dict_tensors(ref_params, res_params, rtol, atol)) |
| 92 | + |
| 93 | + def compare_gradients(self, module, traced, rtol=1e-3, atol=1e-3): |
| 94 | + ref_grad = {key: param.grad for key, param in module.named_parameters()} |
| 95 | + res_grad = {key: param.grad for key, param in traced.named_parameters()} |
| 96 | + self.assertTrue( |
| 97 | + self.compare_dict_tensors(ref_grad, res_grad, rtol=rtol, atol=atol) |
| 98 | + ) |
| 99 | + |
| 100 | + @requires_gpu() |
| 101 | + @torch._inductor.config.patch( |
| 102 | + pre_grad_fusion_options={}, |
| 103 | + post_grad_fusion_options={ |
| 104 | + "activation_quantization_aten_pass": { |
| 105 | + "quant_type": "torch.float8_e5m2", |
| 106 | + "size_in_mb": 0.0, |
| 107 | + }, |
| 108 | + }, |
| 109 | + ) |
| 110 | + def test_activation_quantization_aten(self): |
| 111 | + counters.clear() |
| 112 | + module = TargetCPModule().to(GPU_TYPE) |
| 113 | + input = [ |
| 114 | + torch.rand( |
| 115 | + (16, 10), requires_grad=True, device=GPU_TYPE, dtype=torch.bfloat16 |
| 116 | + ), |
| 117 | + torch.rand( |
| 118 | + (10, 16), requires_grad=True, device=GPU_TYPE, dtype=torch.bfloat16 |
| 119 | + ), |
| 120 | + ] |
| 121 | + traced = torch.compile(module) |
| 122 | + ref = module(*input) |
| 123 | + res = traced(*input) |
| 124 | + self.compare_pred(module, traced, input) |
| 125 | + ref.sum().backward() |
| 126 | + res.sum().backward() |
| 127 | + self.compare_parameters(module, traced) |
| 128 | + self.compare_gradients(module, traced) |
| 129 | + self.assertEqual( |
| 130 | + counters["inductor"]["activation_quantization_fwd_aten_pass"], 1 |
| 131 | + ) |
| 132 | + self.assertEqual( |
| 133 | + counters["inductor"]["activation_quantization_bwd_aten_pass"], 1 |
| 134 | + ) |
| 135 | + self.assertTrue(torch.allclose(ref, res)) |
| 136 | + counters.clear() |
| 137 | + |
| 138 | + module = FeedforwardNN().to(GPU_TYPE) |
| 139 | + X = np.linspace(-10, 10, 100).reshape(-1, 1).astype(np.float32) |
| 140 | + input = [ |
| 141 | + torch.from_numpy(X).to(GPU_TYPE), |
| 142 | + ] |
| 143 | + traced = torch.compile(module) |
| 144 | + ref = module(*input) |
| 145 | + res = traced(*input) |
| 146 | + self.compare_pred(module, traced, input) |
| 147 | + ref.sum().backward() |
| 148 | + res.sum().backward() |
| 149 | + self.compare_parameters(module, traced) |
| 150 | + self.compare_gradients(module, traced) |
| 151 | + self.assertEqual( |
| 152 | + counters["inductor"]["activation_quantization_fwd_aten_pass"], 1 |
| 153 | + ) |
| 154 | + self.assertEqual( |
| 155 | + counters["inductor"]["activation_quantization_bwd_aten_pass"], 1 |
| 156 | + ) |
| 157 | + self.assertTrue(torch.allclose(ref, res)) |
| 158 | + counters.clear() |
| 159 | + |
| 160 | + |
| 161 | +if __name__ == "__main__": |
| 162 | + if IS_LINUX and HAS_GPU: |
| 163 | + run_tests() |
0 commit comments