| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Test the piecewise compilation with a simple model so that we |
| 4 | can exactly calculate the expected output and side effects. |
| 5 | """ |
| 6 | |
| 7 | import pytest |
| 8 | import torch |
| 9 | from torch import nn |
| 10 | |
| 11 | from vllm.compilation.counter import compilation_counter |
| 12 | from vllm.compilation.decorators import support_torch_compile |
| 13 | from vllm.config import ( |
| 14 | CompilationConfig, |
| 15 | CompilationMode, |
| 16 | CUDAGraphMode, |
| 17 | VllmConfig, |
| 18 | set_current_vllm_config, |
| 19 | ) |
| 20 | from vllm.forward_context import BatchDescriptor, set_forward_context |
| 21 | from vllm.utils.torch_utils import is_torch_equal_or_newer |
| 22 | |
| 23 | from ...utils import create_new_process_for_each_test |
| 24 | |
| 25 | # This import automatically registers `torch.ops.silly.attention` |
| 26 | from ..silly_attention import get_global_counter, reset_global_counter |
| 27 | |
| 28 | |
| 29 | # Custom op that returns an unbacked symint during graph capture |
| 30 | @torch.library.custom_op("mylib::foo", mutates_args=()) |
| 31 | def foo(x: torch.Tensor) -> int: |
| 32 | return 3 |
| 33 | |
| 34 | |
| 35 | @foo.register_fake |
| 36 | def _(x): |
| 37 | return torch.library.get_ctx().new_dynamic_size() |
| 38 | |
| 39 | |
| 40 | @support_torch_compile |
| 41 | class SillyModel(nn.Module): |
| 42 | def __init__( |
| 43 | self, |
| 44 | *, |
| 45 | vllm_config: VllmConfig, |
| 46 | prefix: str = "", |
| 47 | intermediate_unbacked=False, |
| 48 | **kwargs, |
| 49 | ) -> None: |
| 50 | super().__init__() |
| 51 | self.intermediate_unbacked = intermediate_unbacked |
| 52 | |
| 53 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 54 | """Overall effect: |
| 55 | x = 3 * x + 19 |
| 56 | global_counter += 2 |
| 57 | """ |
| 58 | x = x + 1 |
| 59 | x = x + 2 |
| 60 | out = torch.empty_like(x) |
| 61 | torch.ops.silly.attention(x, x, x, out) |
| 62 | x = out |
| 63 | x = x - 2 |
| 64 | |
| 65 | if self.intermediate_unbacked: |
| 66 | # Test for unbacked symints: the following is a fancy way to multiply by 1 |
| 67 | u0 = foo(x) |
| 68 | ones = x.new_ones(x.shape[0], u0).sum(-1) / 3 |
| 69 | x = x * ones |
| 70 | |
| 71 | x = x - 1 |
| 72 | out = torch.empty_like(x) |
| 73 | torch.ops.silly.attention(x, x, x, out) |
| 74 | x = out |
| 75 | x = x + 1 |
| 76 | return x |
| 77 | |
| 78 | |
| 79 | @torch._dynamo.config.patch(capture_dynamic_output_shape_ops=True) |
| 80 | def _run_simple_model( |
| 81 | splitting_ops, |
| 82 | use_inductor_graph_partition, |
| 83 | backend, |
| 84 | expected_num_piecewise_graphs_seen, |
| 85 | expected_num_piecewise_capturable_graphs_seen, |
| 86 | expected_num_backend_compilations, |
| 87 | expected_num_cudagraph_captured, |
| 88 | *, |
| 89 | intermediate_unbacked=False, |
| 90 | ): |
| 91 | vllm_config = VllmConfig( |
| 92 | compilation_config=CompilationConfig( |
| 93 | mode=CompilationMode.VLLM_COMPILE, |
| 94 | backend=backend, |
| 95 | splitting_ops=splitting_ops, |
| 96 | use_inductor_graph_partition=use_inductor_graph_partition, |
| 97 | cudagraph_copy_inputs=True, |
| 98 | cudagraph_capture_sizes=[1, 2], |
| 99 | ) |
| 100 | ) |
| 101 | with set_current_vllm_config(vllm_config): |
| 102 | model = SillyModel( |
| 103 | vllm_config=vllm_config, |
| 104 | prefix="", |
| 105 | intermediate_unbacked=intermediate_unbacked, |
| 106 | ) |
| 107 | |
| 108 | inputs = torch.randn(100).cuda() |
| 109 | |
| 110 | with ( |
| 111 | compilation_counter.expect( |
| 112 | num_graphs_seen=1, # one graph for the model |
| 113 | num_piecewise_graphs_seen=expected_num_piecewise_graphs_seen, |
| 114 | num_piecewise_capturable_graphs_seen=expected_num_piecewise_capturable_graphs_seen, |
| 115 | num_backend_compilations=expected_num_backend_compilations, |
| 116 | num_cudagraph_captured=expected_num_cudagraph_captured, |
| 117 | ), |
| 118 | set_forward_context(None, vllm_config=vllm_config), |
| 119 | ): # background context |
| 120 | # warm up with background context |
| 121 | model(inputs) |
| 122 | |
| 123 | # capturing/replaying should under context of cudagraph dispatching |
| 124 | with set_forward_context( |
| 125 | None, |
| 126 | vllm_config=vllm_config, |
| 127 | cudagraph_runtime_mode=CUDAGraphMode.PIECEWISE, |
| 128 | batch_descriptor=BatchDescriptor( |
| 129 | num_tokens=2, |
| 130 | ), |
| 131 | ): |
| 132 | model(torch.randn(2).cuda()) |
| 133 | with set_forward_context( |
| 134 | None, |
| 135 | vllm_config=vllm_config, |
| 136 | cudagraph_runtime_mode=CUDAGraphMode.PIECEWISE, |
| 137 | batch_descriptor=BatchDescriptor( |
| 138 | num_tokens=1, |
| 139 | ), |
| 140 | ): |
| 141 | model(torch.randn(1).cuda()) |
| 142 | |
| 143 | input = torch.zeros(2).cuda() |
| 144 | reset_global_counter() |
| 145 | with set_forward_context( |
| 146 | None, |
| 147 | vllm_config=vllm_config, |
| 148 | cudagraph_runtime_mode=CUDAGraphMode.PIECEWISE, |
| 149 | batch_descriptor=BatchDescriptor( |
| 150 | num_tokens=2, |
| 151 | ), |
| 152 | ): |
| 153 | output = model(input) |
| 154 | assert get_global_counter() == 2 |
| 155 | assert torch.allclose(output.cpu(), torch.tensor([19.0, 19.0])) |
| 156 | |
| 157 | |
| 158 | @pytest.mark.parametrize("backend", ["inductor", "eager"]) |
| 159 | @pytest.mark.parametrize("intermediate_unbacked", [True, False]) |
| 160 | @torch.inference_mode() |
| 161 | @create_new_process_for_each_test("spawn") |
| 162 | def test_simple_piecewise_compile(backend, intermediate_unbacked, monkeypatch): |
| 163 | # `intermediate_unbacked` flips a control-flow branch inside |
| 164 | # `SillyModel.forward`, but the AOT-compile cache key only hashes the |
| 165 | # forward function's qualname + line number, so both parametrize variants |
| 166 | # share the same cache slot. Disabling the cache forces each variant to |
| 167 | # compile fresh; otherwise the second-running variant loads the first's |
| 168 | # artifact and segfaults with an illegal memory access. |
| 169 | monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1") |
| 170 | _run_simple_model( |
| 171 | splitting_ops=["silly::attention"], |
| 172 | use_inductor_graph_partition=False, |
| 173 | backend=backend, |
| 174 | # 2 * num_layers + 1 |
| 175 | expected_num_piecewise_graphs_seen=5, |
| 176 | # 1 + num_layers |
| 177 | expected_num_piecewise_capturable_graphs_seen=3, |
| 178 | # num_piecewise_capturable_graphs_seen |
| 179 | expected_num_backend_compilations=3, |
| 180 | # num_cudagraph_sizes * num_piecewise_capturable_graphs_seen |
| 181 | expected_num_cudagraph_captured=6, |
| 182 | intermediate_unbacked=intermediate_unbacked, |
| 183 | ) |
| 184 | |
| 185 | |
| 186 | @torch.inference_mode() |
| 187 | def test_simple_inductor_graph_partition(monkeypatch): |
| 188 | if not is_torch_equal_or_newer("2.9.0.dev"): |
| 189 | pytest.skip("inductor graph partition is only available in PyTorch 2.9+") |
| 190 | |
| 191 | # disable compile cache so that we run separately for different splitting_ops |
| 192 | # and get the expected number of cudagraphs captured. |
| 193 | monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1") |
| 194 | |
| 195 | _run_simple_model( |
| 196 | splitting_ops=["silly::attention"], |
| 197 | use_inductor_graph_partition=True, |
| 198 | backend="inductor", |
| 199 | # Since not splitting at fx graph level |
| 200 | expected_num_piecewise_graphs_seen=1, |
| 201 | # Since not splitting at fx graph level |
| 202 | expected_num_piecewise_capturable_graphs_seen=1, |
| 203 | # Since not splitting at fx graph level |
| 204 | expected_num_backend_compilations=1, |
| 205 | # Inductor graph partition still captures 6 graph, same as fx graph partition |
| 206 | expected_num_cudagraph_captured=6, |
| 207 | ) |