| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from typing import Any |
| 4 | |
| 5 | import torch |
| 6 | from torch import fx as fx |
| 7 | from torch import nn |
| 8 | |
| 9 | # This import automatically registers `torch.ops.silly.attention` |
| 10 | import tests.compile.silly_attention # noqa |
| 11 | from vllm.compilation.counter import compilation_counter |
| 12 | from vllm.compilation.decorators import support_torch_compile |
| 13 | from vllm.compilation.passes.inductor_pass import ( |
| 14 | InductorPass, |
| 15 | get_pass_context, |
| 16 | ) |
| 17 | from vllm.config import ( |
| 18 | VllmConfig, |
| 19 | set_current_vllm_config, |
| 20 | ) |
| 21 | from vllm.config.compilation import CompilationConfig, CompilationMode |
| 22 | from vllm.config.scheduler import SchedulerConfig |
| 23 | from vllm.config.utils import Range |
| 24 | from vllm.forward_context import set_forward_context |
| 25 | |
| 26 | BATCH_SIZE = 64 |
| 27 | MLP_SIZE = 128 |
| 28 | |
| 29 | |
| 30 | @support_torch_compile |
| 31 | class TestModel(nn.Module): |
| 32 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "", **kwargs) -> None: |
| 33 | super().__init__() |
| 34 | |
| 35 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 36 | x = x + x |
| 37 | attn_output = torch.empty_like(x) |
| 38 | torch.ops.silly.attention(x, x, x, attn_output) |
| 39 | x = attn_output |
| 40 | x = x * 3 |
| 41 | return x |
| 42 | |
| 43 | |
| 44 | @torch.inference_mode |
| 45 | def run_model(vllm_config: VllmConfig, model: nn.Module, batch_sizes: list[int]): |
| 46 | with set_forward_context({}, vllm_config=vllm_config): |
| 47 | model(torch.randn(BATCH_SIZE, MLP_SIZE)) |
| 48 | for batch_size in batch_sizes: |
| 49 | model(torch.randn(batch_size, MLP_SIZE)) |
| 50 | |
| 51 | |
| 52 | class PostGradRangeChecker(InductorPass): |
| 53 | def __init__(self, ranges: list[Range]): |
| 54 | self.ranges = ranges |
| 55 | self.num_calls = 0 |
| 56 | |
| 57 | def __call__(self, graph: fx.Graph): |
| 58 | compile_range = get_pass_context().compile_range |
| 59 | assert compile_range in self.ranges, ( |
| 60 | f"Compile range {compile_range} not in {self.ranges}" |
| 61 | ) |
| 62 | self.num_calls += 1 |
| 63 | |
| 64 | def uuid(self) -> str: |
| 65 | state: dict[str, Any] = {} |
| 66 | return InductorPass.hash_dict(state) |
| 67 | |
| 68 | |
| 69 | def test_compile_ranges(disable_vllm_compile_cache): |
| 70 | post_grad_range_checker = PostGradRangeChecker( |
| 71 | [ |
| 72 | Range(start=1, end=8), |
| 73 | Range(start=16, end=16), |
| 74 | Range(start=9, end=32), |
| 75 | Range(start=64, end=64), |
| 76 | Range(start=128, end=128), |
| 77 | Range(start=33, end=8192), |
| 78 | ] |
| 79 | ) |
| 80 | torch.set_default_device("cuda") |
| 81 | vllm_config = VllmConfig( |
| 82 | scheduler_config=SchedulerConfig( |
| 83 | max_num_batched_tokens=8192, |
| 84 | max_model_len=8192, |
| 85 | is_encoder_decoder=False, |
| 86 | ), |
| 87 | compilation_config=CompilationConfig( |
| 88 | mode=CompilationMode.VLLM_COMPILE, |
| 89 | compile_ranges_endpoints=[8, 32], |
| 90 | compile_sizes=[16, 64, 128], |
| 91 | inductor_compile_config={ |
| 92 | "post_grad_custom_post_pass": post_grad_range_checker, |
| 93 | }, |
| 94 | ), |
| 95 | ) |
| 96 | |
| 97 | with set_current_vllm_config(vllm_config): |
| 98 | model = TestModel(vllm_config=vllm_config, prefix="").eval() |
| 99 | # Number of compilations: 3 compile ranges + 3 compile sizes |
| 100 | batch_sizes = [1, 4, 16, 24, 48, 64, 8192] |
| 101 | |
| 102 | with compilation_counter.expect( |
| 103 | num_graphs_seen=1, |
| 104 | num_piecewise_graphs_seen=1, |
| 105 | num_backend_compilations=6, |
| 106 | ): |
| 107 | run_model(vllm_config, model, batch_sizes) |
| 108 | assert post_grad_range_checker.num_calls == 6 |
| 109 | |
| 110 | |
| 111 | def test_compile_config_get_compile_ranges(): |
| 112 | compilation_config = CompilationConfig( |
| 113 | compile_ranges_endpoints=[8, 32], |
| 114 | ) |
| 115 | VllmConfig( |
| 116 | scheduler_config=SchedulerConfig( |
| 117 | max_num_batched_tokens=8192, |
| 118 | max_model_len=8192, |
| 119 | is_encoder_decoder=False, |
| 120 | ), |
| 121 | compilation_config=compilation_config, |
| 122 | ) |
| 123 | assert compilation_config.get_compile_ranges() == [ |
| 124 | Range(start=1, end=8), |
| 125 | Range(start=9, end=32), |
| 126 | Range(start=33, end=8192), |
| 127 | ] |
| 128 | |
| 129 | |
| 130 | class PostGradStaticShapeChecker(InductorPass): |
| 131 | """Asserts that compile_sizes entries produce graphs with fully concrete |
| 132 | (non-symbolic) shapes, and compile_ranges entries have symbolic shapes.""" |
| 133 | |
| 134 | def __init__(self): |
| 135 | self.num_static_calls = 0 |
| 136 | self.num_dynamic_calls = 0 |
| 137 | |
| 138 | def __call__(self, graph: fx.Graph): |
| 139 | from torch.fx.experimental.symbolic_shapes import is_symbolic |
| 140 | |
| 141 | compile_range = get_pass_context().compile_range |
| 142 | is_single = compile_range.is_single_size() |
| 143 | |
| 144 | for node in graph.nodes: |
| 145 | val = node.meta.get("val") |
| 146 | if val is None: |
| 147 | val = node.meta.get("example_value") |
| 148 | if isinstance(val, torch.Tensor): |
| 149 | has_symbolic = any(is_symbolic(d) for d in val.shape) |
| 150 | if is_single: |
| 151 | assert not has_symbolic, ( |
| 152 | f"compile_sizes entry {compile_range}: " |
| 153 | f"node '{node.name}' has symbolic shape " |
| 154 | f"{val.shape}" |
| 155 | ) |
| 156 | else: |
| 157 | # compile_ranges should have at least some |
| 158 | # symbolic shapes (the batch dimension) |
| 159 | if has_symbolic: |
| 160 | self.num_dynamic_calls += 1 |
| 161 | return |
| 162 | |
| 163 | if is_single: |
| 164 | self.num_static_calls += 1 |
| 165 | |
| 166 | def uuid(self) -> str: |
| 167 | state: dict[str, Any] = {} |
| 168 | return InductorPass.hash_dict(state) |
| 169 | |
| 170 | |
| 171 | def test_compile_sizes_produce_static_shapes(disable_vllm_compile_cache): |
| 172 | """Verify that compile_sizes entries are compiled with fully concrete |
| 173 | shapes (no SymInts), while compile_ranges entries retain dynamic shapes.""" |
| 174 | checker = PostGradStaticShapeChecker() |
| 175 | torch.set_default_device("cuda") |
| 176 | vllm_config = VllmConfig( |
| 177 | scheduler_config=SchedulerConfig( |
| 178 | max_num_batched_tokens=8192, |
| 179 | max_model_len=8192, |
| 180 | is_encoder_decoder=False, |
| 181 | ), |
| 182 | compilation_config=CompilationConfig( |
| 183 | mode=CompilationMode.VLLM_COMPILE, |
| 184 | compile_ranges_endpoints=[8], |
| 185 | compile_sizes=[16], |
| 186 | inductor_compile_config={ |
| 187 | "post_grad_custom_post_pass": checker, |
| 188 | }, |
| 189 | ), |
| 190 | ) |
| 191 | |
| 192 | with set_current_vllm_config(vllm_config): |
| 193 | model = TestModel(vllm_config=vllm_config, prefix="").eval() |
| 194 | # 3 compilations: Range(1,8), Range(9,8192), single-size 16 |
| 195 | with compilation_counter.expect( |
| 196 | num_graphs_seen=1, |
| 197 | num_piecewise_graphs_seen=1, |
| 198 | num_backend_compilations=3, |
| 199 | ): |
| 200 | run_model(vllm_config, model, [1, 16, 64]) |
| 201 | |
| 202 | # compile_sizes=16 should produce static shapes |
| 203 | assert checker.num_static_calls == 1, ( |
| 204 | f"Expected 1 static compilation, got {checker.num_static_calls}" |
| 205 | ) |
| 206 | # compile_ranges should produce dynamic shapes |
| 207 | assert checker.num_dynamic_calls == 2, ( |
| 208 | f"Expected 2 dynamic compilations, got {checker.num_dynamic_calls}" |
| 209 | ) |
| 210 | |
| 211 | |
| 212 | def test_inductor_cache_compile_ranges(disable_vllm_compile_cache): |
| 213 | # disable_vllm_compile_cache sets VLLM_DISABLE_COMPILE_CACHE=1 to force |
| 214 | # multiple compilations by disabling vLLM's on-disk compile cache. |
| 215 | post_grad_range_checker = PostGradRangeChecker( |
| 216 | ranges=[ |
| 217 | Range(start=1, end=8), |
| 218 | Range(start=9, end=8192), |
| 219 | ] |
| 220 | ) |
| 221 | scheduler_config = SchedulerConfig( |
| 222 | max_num_batched_tokens=8192, |
| 223 | max_model_len=8192, |
| 224 | is_encoder_decoder=False, |
| 225 | ) |
| 226 | torch.set_default_device("cuda") |
| 227 | |
| 228 | def create_vllm_config(): |
| 229 | return VllmConfig( |
| 230 | scheduler_config=scheduler_config, |
| 231 | compilation_config=CompilationConfig( |
| 232 | mode=CompilationMode.VLLM_COMPILE, |
| 233 | compile_ranges_endpoints=[8], |
| 234 | inductor_compile_config={ |
| 235 | "post_grad_custom_post_pass": post_grad_range_checker, |
| 236 | }, |
| 237 | ), |
| 238 | ) |
| 239 | |
| 240 | vllm_config_1 = create_vllm_config() |
| 241 | with set_current_vllm_config(vllm_config_1): |
| 242 | model1 = TestModel(vllm_config=vllm_config_1, prefix="").eval() |
| 243 | batch_sizes = [1, 16] |
| 244 | run_model(vllm_config_1, model1, batch_sizes) |
| 245 | assert post_grad_range_checker.num_calls == 2 |
| 246 | |
| 247 | post_grad_range_checker.num_calls = 0 |
| 248 | # Create a new vllm config with the new pass context |
| 249 | vllm_config_2 = create_vllm_config() |
| 250 | with set_current_vllm_config(vllm_config_2): |
| 251 | model2 = TestModel(vllm_config=vllm_config_2, prefix="").eval() |
| 252 | batch_sizes = [4, 32] |
| 253 | run_model(vllm_config_2, model2, batch_sizes) |
| 254 | # Check that cache is used, so the number of calls |
| 255 | # should be 0 |
| 256 | assert post_grad_range_checker.num_calls == 0 |