| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import pytest |
| 4 | import torch |
| 5 | from torch import nn |
| 6 | |
| 7 | from vllm.compilation.counter import compilation_counter |
| 8 | from vllm.compilation.decorators import ignore_torch_compile, support_torch_compile |
| 9 | from vllm.config import ( |
| 10 | CacheConfig, |
| 11 | CompilationConfig, |
| 12 | CompilationMode, |
| 13 | CUDAGraphMode, |
| 14 | VllmConfig, |
| 15 | set_current_vllm_config, |
| 16 | ) |
| 17 | from vllm.forward_context import BatchDescriptor, set_forward_context |
| 18 | from vllm.utils.torch_utils import is_torch_equal_or_newer |
| 19 | |
| 20 | # This import automatically registers `torch.ops.silly.attention` |
| 21 | from . import silly_attention # noqa: F401 |
| 22 | |
| 23 | BATCH_SIZE = 32 |
| 24 | MLP_SIZE = 128 |
| 25 | |
| 26 | |
| 27 | @torch.inference_mode |
| 28 | def run_model( |
| 29 | vllm_config: VllmConfig, model: nn.Module, cudagraph_runtime_mode: CUDAGraphMode |
| 30 | ): |
| 31 | with set_forward_context({}, vllm_config=vllm_config): |
| 32 | # warmup for the model with cudagraph_mode NONE |
| 33 | model(torch.randn(BATCH_SIZE, MLP_SIZE).cuda()) |
| 34 | |
| 35 | # simulate cudagraphs capturing |
| 36 | with set_forward_context( |
| 37 | {}, |
| 38 | vllm_config=vllm_config, |
| 39 | cudagraph_runtime_mode=cudagraph_runtime_mode, |
| 40 | batch_descriptor=BatchDescriptor( |
| 41 | num_tokens=2, |
| 42 | ), |
| 43 | ): |
| 44 | model(torch.randn(2, MLP_SIZE).cuda()) |
| 45 | with set_forward_context( |
| 46 | {}, |
| 47 | vllm_config=vllm_config, |
| 48 | cudagraph_runtime_mode=cudagraph_runtime_mode, |
| 49 | batch_descriptor=BatchDescriptor( |
| 50 | num_tokens=1, |
| 51 | ), |
| 52 | ): |
| 53 | model(torch.randn(1, MLP_SIZE).cuda()) |
| 54 | |
| 55 | # simulate cudagraphs replay |
| 56 | with set_forward_context( |
| 57 | {}, |
| 58 | vllm_config=vllm_config, |
| 59 | cudagraph_runtime_mode=cudagraph_runtime_mode, |
| 60 | batch_descriptor=BatchDescriptor( |
| 61 | num_tokens=2, |
| 62 | ), |
| 63 | ): |
| 64 | output = model(torch.randn(2, MLP_SIZE).cuda()) |
| 65 | |
| 66 | output = output.cpu() |
| 67 | return output.cpu() |
| 68 | |
| 69 | |
| 70 | @pytest.mark.parametrize("use_inductor_graph_partition", [True, False]) |
| 71 | def test_ignore_torch_compile_decorator(use_inductor_graph_partition, monkeypatch): |
| 72 | # disable compile cache so that we can count the number of compilations |
| 73 | # appropriately |
| 74 | monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1") |
| 75 | |
| 76 | if use_inductor_graph_partition and not is_torch_equal_or_newer("2.9.0.dev"): |
| 77 | pytest.skip("inductor graph partition is only available in PyTorch 2.9+") |
| 78 | |
| 79 | # piecewise |
| 80 | vllm_config = VllmConfig( |
| 81 | compilation_config=CompilationConfig( |
| 82 | mode=CompilationMode.VLLM_COMPILE, |
| 83 | splitting_ops=["silly::attention"], |
| 84 | cudagraph_capture_sizes=[1, 2], |
| 85 | use_inductor_graph_partition=use_inductor_graph_partition, |
| 86 | ) |
| 87 | ) |
| 88 | cudagraph_runtime_mode = CUDAGraphMode.PIECEWISE |
| 89 | |
| 90 | expected_num_graphs_seen = 1 |
| 91 | expected_num_cudagraph_captured = ( |
| 92 | 4 # num_cudagraph_sizes * num cudagraphs to capture |
| 93 | ) |
| 94 | if use_inductor_graph_partition: |
| 95 | expected_num_piecewise_graphs_seen = 1 |
| 96 | expected_num_piecewise_capturable_graphs_seen = 1 |
| 97 | expected_num_backend_compilations = 1 |
| 98 | else: |
| 99 | expected_num_piecewise_graphs_seen = 3 |
| 100 | expected_num_piecewise_capturable_graphs_seen = 2 |
| 101 | expected_num_backend_compilations = 2 |
| 102 | |
| 103 | @support_torch_compile |
| 104 | class A(nn.Module): |
| 105 | def __init__( |
| 106 | self, *, vllm_config: VllmConfig, prefix: str = "", **kwargs |
| 107 | ) -> None: |
| 108 | super().__init__() |
| 109 | |
| 110 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 111 | x = x + x |
| 112 | attn_output = torch.empty_like(x) |
| 113 | torch.ops.silly.attention(x, x, x, attn_output) |
| 114 | x = attn_output |
| 115 | x = x * 3 |
| 116 | return x |
| 117 | |
| 118 | @ignore_torch_compile |
| 119 | class B(A): ... |
| 120 | |
| 121 | @support_torch_compile |
| 122 | class C(B): ... |
| 123 | |
| 124 | with set_current_vllm_config(vllm_config): |
| 125 | mod_A = A(vllm_config=vllm_config, prefix="").eval().cuda() |
| 126 | |
| 127 | # A has support_torch_compile |
| 128 | with compilation_counter.expect( |
| 129 | num_graphs_seen=expected_num_graphs_seen, |
| 130 | num_piecewise_graphs_seen=expected_num_piecewise_graphs_seen, |
| 131 | num_piecewise_capturable_graphs_seen=expected_num_piecewise_capturable_graphs_seen, |
| 132 | num_backend_compilations=expected_num_backend_compilations, |
| 133 | num_cudagraph_captured=expected_num_cudagraph_captured, |
| 134 | ): |
| 135 | run_model(vllm_config, mod_A, cudagraph_runtime_mode) |
| 136 | |
| 137 | with set_current_vllm_config(vllm_config): |
| 138 | mod_B = B(vllm_config=vllm_config, prefix="").eval().cuda() |
| 139 | |
| 140 | # B's ignore_torch_compile should override A's support_torch_compile |
| 141 | with compilation_counter.expect( |
| 142 | num_graphs_seen=0, |
| 143 | num_piecewise_graphs_seen=0, |
| 144 | num_piecewise_capturable_graphs_seen=0, |
| 145 | num_backend_compilations=0, |
| 146 | num_cudagraph_captured=0, |
| 147 | ): |
| 148 | run_model(vllm_config, mod_B, cudagraph_runtime_mode) |
| 149 | |
| 150 | with set_current_vllm_config(vllm_config): |
| 151 | mod_C = C(vllm_config=vllm_config, prefix="").eval().cuda() |
| 152 | |
| 153 | # C's support_torch_compile should override B's ignore_torch_compile |
| 154 | with compilation_counter.expect( |
| 155 | num_graphs_seen=expected_num_graphs_seen, |
| 156 | num_piecewise_graphs_seen=expected_num_piecewise_graphs_seen, |
| 157 | num_piecewise_capturable_graphs_seen=expected_num_piecewise_capturable_graphs_seen, |
| 158 | num_backend_compilations=expected_num_backend_compilations, |
| 159 | num_cudagraph_captured=expected_num_cudagraph_captured, |
| 160 | ): |
| 161 | run_model(vllm_config, mod_C, cudagraph_runtime_mode) |
| 162 | |
| 163 | |
| 164 | # Only enable torch.compile if |
| 165 | # vllm_config.cache_config.kv_sharing_fast_prefill=True |
| 166 | @support_torch_compile( |
| 167 | enable_if=lambda vllm_config: vllm_config.cache_config.kv_sharing_fast_prefill |
| 168 | ) |
| 169 | class B(nn.Module): |
| 170 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "", **kwargs) -> None: |
| 171 | super().__init__() |
| 172 | |
| 173 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 174 | x = x + x |
| 175 | attn_output = torch.empty_like(x) |
| 176 | torch.ops.silly.attention(x, x, x, attn_output) |
| 177 | x = attn_output |
| 178 | x = x + x |
| 179 | return x |
| 180 | |
| 181 | |
| 182 | # Only enable torch.compile if |
| 183 | # vllm_config.cache_config.kv_sharing_fast_prefill=False |
| 184 | @support_torch_compile( |
| 185 | enable_if=lambda vllm_config: not vllm_config.cache_config.kv_sharing_fast_prefill |
| 186 | ) |
| 187 | class A(nn.Module): |
| 188 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "", **kwargs) -> None: |
| 189 | super().__init__() |
| 190 | self.mod1 = B(vllm_config=vllm_config, prefix=prefix, **kwargs) |
| 191 | self.mod2 = B(vllm_config=vllm_config, prefix=prefix, **kwargs) |
| 192 | |
| 193 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 194 | x = self.mod1(x) |
| 195 | attn_output = torch.empty_like(x) |
| 196 | torch.ops.silly.attention(x, x, x, attn_output) |
| 197 | x = attn_output |
| 198 | x = self.mod2(x) |
| 199 | return x |
| 200 | |
| 201 | |
| 202 | @pytest.mark.parametrize("use_inductor_graph_partition", [True, False]) |
| 203 | def test_conditional_compile_enable_if(use_inductor_graph_partition, monkeypatch): |
| 204 | # disable compile cache so that we can count the number of compilations |
| 205 | # appropriately |
| 206 | monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1") |
| 207 | |
| 208 | if use_inductor_graph_partition and not is_torch_equal_or_newer("2.9.0.dev"): |
| 209 | pytest.skip("inductor graph partition is only available in PyTorch 2.9+") |
| 210 | |
| 211 | vllm_config = VllmConfig( |
| 212 | cache_config=CacheConfig( |
| 213 | kv_sharing_fast_prefill=True, |
| 214 | ), |
| 215 | compilation_config=CompilationConfig( |
| 216 | mode=CompilationMode.VLLM_COMPILE, |
| 217 | splitting_ops=["silly::attention"], |
| 218 | cudagraph_capture_sizes=[1, 2], |
| 219 | use_inductor_graph_partition=use_inductor_graph_partition, |
| 220 | ), |
| 221 | ) |
| 222 | cudagraph_runtime_mode = CUDAGraphMode.PIECEWISE |
| 223 | |
| 224 | with set_current_vllm_config(vllm_config): |
| 225 | mod_A = A(vllm_config=vllm_config, prefix="").eval().cuda() |
| 226 | |
| 227 | if use_inductor_graph_partition: |
| 228 | expected_num_piecewise_graphs_seen = 2 |
| 229 | expected_num_piecewise_capturable_graphs_seen = 2 |
| 230 | expected_num_backend_compilations = 2 |
| 231 | else: |
| 232 | expected_num_piecewise_graphs_seen = 6 |
| 233 | expected_num_piecewise_capturable_graphs_seen = 4 |
| 234 | expected_num_backend_compilations = 4 |
| 235 | |
| 236 | # A has support_torch_compile but enable_if fn returns False |
| 237 | # enable_if will be True for B, so we expect mod1 and mod2 |
| 238 | # to be compiled |
| 239 | with compilation_counter.expect( |
| 240 | num_graphs_seen=2, |
| 241 | num_piecewise_graphs_seen=expected_num_piecewise_graphs_seen, |
| 242 | # 3 piecewise graphs per instance of B() |
| 243 | num_piecewise_capturable_graphs_seen=expected_num_piecewise_capturable_graphs_seen, |
| 244 | num_backend_compilations=expected_num_backend_compilations, |
| 245 | num_cudagraph_captured=8, |
| 246 | # num_cudagraph_sizes * num cudagraphable graphs to capture |
| 247 | ): |
| 248 | run_model(vllm_config, mod_A, cudagraph_runtime_mode) |
| 249 | |
| 250 | # Set kv_sharing_fast_prefill=False |
| 251 | # which will cause A to be compiled and B to not be compiled |
| 252 | vllm_config = VllmConfig( |
| 253 | cache_config=CacheConfig( |
| 254 | kv_sharing_fast_prefill=False, |
| 255 | ), |
| 256 | compilation_config=CompilationConfig( |
| 257 | mode=CompilationMode.VLLM_COMPILE, |
| 258 | splitting_ops=["silly::attention"], |
| 259 | cudagraph_capture_sizes=[1, 2], |
| 260 | use_inductor_graph_partition=use_inductor_graph_partition, |
| 261 | ), |
| 262 | ) |
| 263 | |
| 264 | with set_current_vllm_config(vllm_config): |
| 265 | mod_A = A(vllm_config=vllm_config, prefix="").eval().cuda() |
| 266 | |
| 267 | if use_inductor_graph_partition: |
| 268 | expected_num_piecewise_graphs_seen = 1 |
| 269 | expected_num_piecewise_capturable_graphs_seen = 1 |
| 270 | expected_num_backend_compilations = 1 |
| 271 | else: |
| 272 | # 3 attn ops and 4 non-attn ops |
| 273 | expected_num_piecewise_graphs_seen = 7 |
| 274 | expected_num_piecewise_capturable_graphs_seen = 4 |
| 275 | expected_num_backend_compilations = 4 |
| 276 | |
| 277 | with compilation_counter.expect( |
| 278 | num_graphs_seen=1, |
| 279 | num_piecewise_graphs_seen=expected_num_piecewise_graphs_seen, |
| 280 | # 3 attn ops and 4 non-attn ops |
| 281 | num_piecewise_capturable_graphs_seen=expected_num_piecewise_capturable_graphs_seen, |
| 282 | num_backend_compilations=expected_num_backend_compilations, |
| 283 | num_cudagraph_captured=8, |
| 284 | # num_cudagraph_sizes * num cudagraphable graphs to capture |
| 285 | ): |
| 286 | run_model(vllm_config, mod_A, cudagraph_runtime_mode) |