|
| 1 | +import itertools |
| 2 | + |
| 3 | +import pytest |
| 4 | +import torch |
| 5 | +import torch.utils.checkpoint |
| 6 | +from torch._dynamo.utils import counters |
| 7 | +from torch._inductor import config |
| 8 | +from torch._inductor.test_case import TestCase, run_tests |
| 9 | +from torch._inductor.utils import run_and_get_code |
| 10 | +from torch.testing._internal.common_utils import IS_LINUX, skipIfRocm |
| 11 | +from torch.testing._internal.inductor_utils import HAS_CPU |
| 12 | +from torch.utils.cpp_extension import IS_WINDOWS |
| 13 | + |
| 14 | +import torchao |
| 15 | +from torchao.prototype.inductor.fx_passes.int8_sdpa_fusion import _int8_sdpa_init |
| 16 | +from torchao.utils import TORCH_VERSION_AT_LEAST_2_7 |
| 17 | + |
| 18 | + |
| 19 | +class SelfAttnLikeModule(torch.nn.Module): |
| 20 | + def __init__( |
| 21 | + self, |
| 22 | + input_dim, |
| 23 | + has_mask, |
| 24 | + num_attention_heads=None, |
| 25 | + attention_head_size=None, |
| 26 | + ) -> None: |
| 27 | + super().__init__() |
| 28 | + self.input_dim = input_dim |
| 29 | + self.q_proj = torch.nn.Linear(input_dim, input_dim, bias=False) |
| 30 | + self.k_proj = torch.nn.Linear(input_dim, input_dim, bias=False) |
| 31 | + self.v_proj = torch.nn.Linear(input_dim, input_dim, bias=False) |
| 32 | + self.softmax = torch.nn.Softmax(dim=-1) |
| 33 | + assert num_attention_heads is not None |
| 34 | + assert attention_head_size is not None |
| 35 | + self.num_attention_heads = num_attention_heads |
| 36 | + self.attention_head_size = attention_head_size |
| 37 | + self.all_head_size = self.num_attention_heads * self.attention_head_size |
| 38 | + self.dense = torch.nn.Linear(self.all_head_size, self.all_head_size) |
| 39 | + self.dropout = torch.nn.Dropout(0) |
| 40 | + self.has_mask = has_mask |
| 41 | + |
| 42 | + def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor: |
| 43 | + new_x_shape = x.size()[:-1] + ( |
| 44 | + self.num_attention_heads, |
| 45 | + self.attention_head_size, |
| 46 | + ) |
| 47 | + x = x.view(new_x_shape) |
| 48 | + return x.permute([0, 2, 1, 3]) |
| 49 | + |
| 50 | + def forward(self, x, mask): |
| 51 | + q = self.q_proj(x) |
| 52 | + k = self.k_proj(x) |
| 53 | + v = self.v_proj(x) |
| 54 | + q = self.transpose_for_scores(q) |
| 55 | + k = self.transpose_for_scores(k) |
| 56 | + v = self.transpose_for_scores(v) |
| 57 | + scores = torch.matmul(q, k.transpose(-1, -2)) / (self.input_dim**0.5) |
| 58 | + if self.has_mask and mask.dtype != scores.dtype: |
| 59 | + scores = scores + mask |
| 60 | + attention = self.softmax(scores) |
| 61 | + attention = self.dropout(attention) |
| 62 | + context_layer = torch.matmul(attention, v) |
| 63 | + context_layer = context_layer.permute(0, 2, 1, 3).contiguous() |
| 64 | + context_layer = context_layer.view( |
| 65 | + context_layer.size()[:-2] + (self.all_head_size,) |
| 66 | + ) |
| 67 | + return self.dense(context_layer) |
| 68 | + |
| 69 | + |
| 70 | +class TestSDPAPatternRewriterTemplate(TestCase): |
| 71 | + def _clone_inputs(self, inputs): |
| 72 | + def clone(x): |
| 73 | + if not isinstance(x, torch.Tensor): |
| 74 | + return x |
| 75 | + return x.clone() |
| 76 | + |
| 77 | + return [clone(x) for x in inputs] |
| 78 | + |
| 79 | + def _check_common( |
| 80 | + self, |
| 81 | + dot_prod_attention, |
| 82 | + args1=None, |
| 83 | + contains=True, |
| 84 | + atol=1e-5, |
| 85 | + has_fuse_pattern=True, |
| 86 | + has_dropout=False, |
| 87 | + check_train=True, |
| 88 | + override_check_equal=False, |
| 89 | + dtype=torch.float, |
| 90 | + rtol=1.3e-6, |
| 91 | + ): |
| 92 | + if args1 is None: |
| 93 | + tensor_shape = (4, 2, 16, 32) |
| 94 | + args1 = [ |
| 95 | + torch.randn(tensor_shape, device=self.device, dtype=dtype), |
| 96 | + torch.randn(tensor_shape, device=self.device, dtype=dtype), |
| 97 | + torch.randn(tensor_shape, device=self.device, dtype=dtype), |
| 98 | + ] |
| 99 | + else: |
| 100 | + args1 = list(args1) |
| 101 | + args2 = self._clone_inputs(args1) |
| 102 | + |
| 103 | + for training in [False, True] if check_train else [False]: |
| 104 | + for x in itertools.chain(args1[:], args2[:]): |
| 105 | + if isinstance(x, torch.Tensor) and x.is_floating_point(): |
| 106 | + x.requires_grad = training |
| 107 | + |
| 108 | + dropout_arg = [training] if has_dropout else [] |
| 109 | + torch.manual_seed(1234) |
| 110 | + result1 = dot_prod_attention(*(args1 + dropout_arg)) |
| 111 | + |
| 112 | + counters.clear() |
| 113 | + torch.manual_seed(1234) |
| 114 | + compiled_model = torch.compile(dot_prod_attention, fullgraph=True) |
| 115 | + result2, source_code = run_and_get_code( |
| 116 | + compiled_model, |
| 117 | + *(args2 + dropout_arg), |
| 118 | + ) |
| 119 | + source_code = "\n".join(source_code) |
| 120 | + if has_fuse_pattern: |
| 121 | + self.assertGreaterEqual(counters["inductor"]["int8_fuse_attention"], 1) |
| 122 | + if contains: |
| 123 | + # many of the patterns get re-expanded in dispatcher |
| 124 | + self.assertIn( |
| 125 | + "torchao.scaled_dot_product_int8", |
| 126 | + source_code, |
| 127 | + ) |
| 128 | + |
| 129 | + # some tests configured with very low dropout where we still want to check equality |
| 130 | + if not has_dropout or override_check_equal: |
| 131 | + self.assertEqual(result1, result2, atol=atol, rtol=1.3e-6) |
| 132 | + |
| 133 | + if training: |
| 134 | + result1.sum().backward() |
| 135 | + result2.sum().backward() |
| 136 | + for arg1, arg2 in zip(args1, args2): |
| 137 | + if ( |
| 138 | + isinstance(arg1, torch.Tensor) |
| 139 | + and arg1.is_floating_point() |
| 140 | + and (not has_dropout or override_check_equal) |
| 141 | + ): |
| 142 | + self.assertEqual(arg1.grad, arg2.grad, atol=atol, rtol=rtol) |
| 143 | + |
| 144 | + @skipIfRocm |
| 145 | + @pytest.mark.skipif( |
| 146 | + not TORCH_VERSION_AT_LEAST_2_7, reason="int8 sdpa requires torch 2.7 or later" |
| 147 | + ) |
| 148 | + @pytest.mark.skipif(IS_WINDOWS, reason="int8 sdpa does not support windows yet") |
| 149 | + @config.patch({"freezing": True}) |
| 150 | + def _test_sdpa_int8_rewriter(self): |
| 151 | + from torch.export import export_for_training |
| 152 | + |
| 153 | + import torchao.quantization.pt2e.quantizer.x86_inductor_quantizer as xiq |
| 154 | + from torchao.quantization.pt2e.quantize_pt2e import convert_pt2e, prepare_pt2e |
| 155 | + from torchao.quantization.pt2e.quantizer.x86_inductor_quantizer import ( |
| 156 | + X86InductorQuantizer, |
| 157 | + ) |
| 158 | + |
| 159 | + # pattern is different for bs=1 |
| 160 | + torch.manual_seed(1234) |
| 161 | + for dtype, has_mask, bs in itertools.product( |
| 162 | + [torch.float32, torch.bfloat16], [True, False], [56, 1] |
| 163 | + ): |
| 164 | + seqlen, numhead, headsize = 197, 16, 64 |
| 165 | + mod = SelfAttnLikeModule( |
| 166 | + input_dim=headsize * numhead, |
| 167 | + has_mask=has_mask, |
| 168 | + num_attention_heads=numhead, |
| 169 | + attention_head_size=headsize, |
| 170 | + ).eval() |
| 171 | + inputs = ( |
| 172 | + torch.randn( |
| 173 | + (bs, seqlen, headsize * numhead), device=self.device, dtype=dtype |
| 174 | + ), |
| 175 | + torch.randn((bs, 1, 1, seqlen), device=self.device) |
| 176 | + if has_mask |
| 177 | + else None, |
| 178 | + ) |
| 179 | + enable_autocast = dtype == torch.bfloat16 |
| 180 | + with ( |
| 181 | + torch.no_grad(), |
| 182 | + torch.amp.autocast( |
| 183 | + self.device, enabled=enable_autocast, dtype=torch.bfloat16 |
| 184 | + ), |
| 185 | + ): |
| 186 | + _int8_sdpa_init() |
| 187 | + quantizer = X86InductorQuantizer() |
| 188 | + quantizer.set_global(xiq.get_default_x86_inductor_quantization_config()) |
| 189 | + quantizer.set_function_type_qconfig( |
| 190 | + torch.matmul, quantizer.get_global_quantization_config() |
| 191 | + ) |
| 192 | + export_model = export_for_training( |
| 193 | + mod, |
| 194 | + inputs, |
| 195 | + strict=True, |
| 196 | + ).module() |
| 197 | + prepare_model = prepare_pt2e(export_model, quantizer) |
| 198 | + prepare_model(*inputs) |
| 199 | + convert_model = convert_pt2e(prepare_model) |
| 200 | + torchao.quantization.pt2e.move_exported_model_to_eval(convert_model) |
| 201 | + self._check_common( |
| 202 | + convert_model, args1=inputs, check_train=False, atol=1.0 |
| 203 | + ) |
| 204 | + |
| 205 | + |
| 206 | +if HAS_CPU: |
| 207 | + |
| 208 | + class SDPAPatternRewriterCpuTests(TestSDPAPatternRewriterTemplate): |
| 209 | + device = "cpu" |
| 210 | + test_sdpa_int8_rewriter_cpu = ( |
| 211 | + TestSDPAPatternRewriterTemplate._test_sdpa_int8_rewriter |
| 212 | + ) |
| 213 | + |
| 214 | + |
| 215 | +if __name__ == "__main__": |
| 216 | + if IS_LINUX: |
| 217 | + run_tests() |
0 commit comments