| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import hashlib |
| 5 | import os |
| 6 | import pickle |
| 7 | import tempfile |
| 8 | from contextlib import contextmanager |
| 9 | from pathlib import Path |
| 10 | from unittest.mock import Mock, patch |
| 11 | |
| 12 | import pytest |
| 13 | import torch |
| 14 | |
| 15 | import vllm.envs as envs |
| 16 | import vllm.kernels # noqa: F401 to register kernels |
| 17 | from vllm import ir |
| 18 | from vllm.compilation.backends import VllmBackend |
| 19 | from vllm.compilation.caching import ( |
| 20 | StandaloneCompiledArtifacts, |
| 21 | VllmSerializableFunction, |
| 22 | ) |
| 23 | from vllm.compilation.counter import compilation_counter |
| 24 | from vllm.compilation.decorators import support_torch_compile |
| 25 | from vllm.config import ( |
| 26 | CompilationConfig, |
| 27 | CompilationMode, |
| 28 | VllmConfig, |
| 29 | set_current_vllm_config, |
| 30 | ) |
| 31 | from vllm.envs import disable_envs_cache |
| 32 | from vllm.forward_context import set_forward_context |
| 33 | from vllm.ir import ops |
| 34 | from vllm.model_executor.layers.activation import GeluAndMulSparse |
| 35 | from vllm.platforms import current_platform |
| 36 | from vllm.utils.torch_utils import is_torch_equal_or_newer |
| 37 | |
| 38 | from ..utils import create_new_process_for_each_test |
| 39 | |
| 40 | |
| 41 | @pytest.fixture |
| 42 | def vllm_tmp_cache(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> Path: |
| 43 | """Fixture that sets VLLM_CACHE_ROOT to a temporary directory.""" |
| 44 | monkeypatch.setenv("VLLM_CACHE_ROOT", str(tmp_path / "vllm_cache")) |
| 45 | return tmp_path |
| 46 | |
| 47 | |
| 48 | def reference_fn(x: torch.Tensor): |
| 49 | assert x.shape[0] <= 42 |
| 50 | assert x.shape[0] % 2 == 0 |
| 51 | for _ in range(30): |
| 52 | x = x + x.shape[0] |
| 53 | return x |
| 54 | |
| 55 | |
| 56 | def reference_fn_tuple(x: torch.Tensor): |
| 57 | """Reference function that returns a tuple of tensors.""" |
| 58 | assert x.shape[0] <= 42 |
| 59 | assert x.shape[0] % 2 == 0 |
| 60 | for _ in range(30): |
| 61 | x = x + x.shape[0] |
| 62 | return x, x * 2 |
| 63 | |
| 64 | |
| 65 | @support_torch_compile |
| 66 | class CompiledMod(torch.nn.Module): |
| 67 | def __init__(self, **kwargs): |
| 68 | super().__init__() |
| 69 | |
| 70 | def forward(self, x: torch.Tensor): |
| 71 | return reference_fn(x) |
| 72 | |
| 73 | |
| 74 | @support_torch_compile |
| 75 | class CompiledModTuple(torch.nn.Module): |
| 76 | """A compiled module that returns a tuple of tensors.""" |
| 77 | |
| 78 | def __init__(self, **kwargs): |
| 79 | super().__init__() |
| 80 | |
| 81 | def forward(self, x: torch.Tensor): |
| 82 | return reference_fn_tuple(x) |
| 83 | |
| 84 | |
| 85 | @support_torch_compile(dynamic_arg_dims={"x": 0}) |
| 86 | class SparseActivationAotModel(torch.nn.Module): |
| 87 | def __init__(self, **kwargs): |
| 88 | super().__init__() |
| 89 | self.up_proj = torch.nn.Linear(32, 256, bias=False) |
| 90 | self.activation = GeluAndMulSparse(0.95, "tanh") |
| 91 | self.down_proj = torch.nn.Linear(128, 16, bias=False) |
| 92 | |
| 93 | def forward(self, x: torch.Tensor): |
| 94 | return self.down_proj(self.activation(self.up_proj(x))) |
| 95 | |
| 96 | |
| 97 | def make_vllm_config() -> VllmConfig: |
| 98 | return VllmConfig( |
| 99 | compilation_config=CompilationConfig( |
| 100 | mode=CompilationMode.VLLM_COMPILE, |
| 101 | backend="inductor", |
| 102 | ) |
| 103 | ) |
| 104 | |
| 105 | |
| 106 | @contextmanager |
| 107 | def use_vllm_config(vllm_config: VllmConfig): |
| 108 | with set_forward_context({}, vllm_config), set_current_vllm_config(vllm_config): |
| 109 | yield |
| 110 | |
| 111 | |
| 112 | @pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10") |
| 113 | def test_no_dynamo_cache_entry(monkeypatch: pytest.MonkeyPatch): |
| 114 | with monkeypatch.context() as m: |
| 115 | vllm_config = make_vllm_config() |
| 116 | args = (torch.randn(10, 10),) |
| 117 | expected = reference_fn(*args) |
| 118 | with use_vllm_config(vllm_config): |
| 119 | m.setenv("VLLM_USE_AOT_COMPILE", "0") |
| 120 | m.setenv("VLLM_USE_MEGA_AOT_ARTIFACT", "1") |
| 121 | m.setenv("VLLM_USE_STANDALONE_COMPILE", "1") |
| 122 | with ( |
| 123 | pytest.raises(RuntimeError, match="Detected recompile"), |
| 124 | torch.compiler.set_stance("fail_on_recompile"), |
| 125 | ): |
| 126 | CompiledMod(vllm_config=vllm_config)(*args) |
| 127 | disable_envs_cache() |
| 128 | |
| 129 | m.setenv("VLLM_USE_AOT_COMPILE", "1") |
| 130 | torch._dynamo.reset() |
| 131 | with torch.compiler.set_stance("fail_on_recompile"): |
| 132 | actual = CompiledMod(vllm_config=vllm_config)(*args) |
| 133 | assert torch.allclose(actual, expected) |
| 134 | |
| 135 | |
| 136 | @pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10") |
| 137 | def test_force_aot_load(monkeypatch: pytest.MonkeyPatch): |
| 138 | with tempfile.TemporaryDirectory() as tmpdirname, monkeypatch.context() as m: |
| 139 | args = (torch.randn(10, 10),) |
| 140 | m.setenv("VLLM_USE_AOT_COMPILE", "1") |
| 141 | m.setenv("VLLM_USE_MEGA_AOT_ARTIFACT", "1") |
| 142 | m.setenv("VLLM_USE_STANDALONE_COMPILE", "1") |
| 143 | m.setenv("VLLM_FORCE_AOT_LOAD", "1") |
| 144 | m.setenv("VLLM_CACHE_ROOT", tmpdirname) |
| 145 | vllm_config = make_vllm_config() |
| 146 | with use_vllm_config(vllm_config), pytest.raises(FileNotFoundError): |
| 147 | CompiledMod(vllm_config=vllm_config)(*args) |
| 148 | |
| 149 | |
| 150 | @pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10") |
| 151 | def test_save_and_load(monkeypatch: pytest.MonkeyPatch): |
| 152 | with monkeypatch.context() as m: |
| 153 | args = (torch.randn(10, 10),) |
| 154 | |
| 155 | with tempfile.TemporaryDirectory() as tmpdirname: |
| 156 | m.setenv("VLLM_CACHE_ROOT", tmpdirname) |
| 157 | m.setenv("VLLM_USE_AOT_COMPILE", "1") |
| 158 | m.setenv("VLLM_USE_MEGA_AOT_ARTIFACT", "1") |
| 159 | m.setenv("VLLM_USE_STANDALONE_COMPILE", "1") |
| 160 | disable_envs_cache() |
| 161 | vllm_config = make_vllm_config() |
| 162 | with ( |
| 163 | use_vllm_config(vllm_config), |
| 164 | compilation_counter.expect( |
| 165 | num_aot_compiles=1, |
| 166 | num_aot_artifacts_saved=1, |
| 167 | num_aot_artifacts_loaded=0, |
| 168 | ), |
| 169 | ): |
| 170 | compiled_mod = CompiledMod(vllm_config=vllm_config) |
| 171 | expected = compiled_mod(*args) |
| 172 | assert isinstance(expected, torch.Tensor) |
| 173 | |
| 174 | disable_envs_cache() |
| 175 | |
| 176 | m.setenv("VLLM_FORCE_AOT_LOAD", "1") |
| 177 | vllm_config = make_vllm_config() |
| 178 | with ( |
| 179 | use_vllm_config(vllm_config), |
| 180 | compilation_counter.expect( |
| 181 | num_aot_compiles=0, |
| 182 | num_aot_artifacts_saved=0, |
| 183 | num_aot_artifacts_loaded=1, |
| 184 | ), |
| 185 | ): |
| 186 | cached_mod = CompiledMod(vllm_config=vllm_config) |
| 187 | ret = cached_mod(*args) |
| 188 | assert isinstance(ret, torch.Tensor) |
| 189 | assert cached_mod.was_aot_compile_fn_loaded_from_disk, ( |
| 190 | "Expected was_aot_compile_fn_loaded_from_disk to be True" |
| 191 | ) |
| 192 | assert torch.allclose(ret, expected) |
| 193 | |
| 194 | |
| 195 | @pytest.mark.skipif( |
| 196 | not current_platform.is_cuda() or not is_torch_equal_or_newer("2.10.0"), |
| 197 | reason="requires CUDA and torch 2.10", |
| 198 | ) |
| 199 | def test_sparse_activation_standalone_aot_preserves_dependencies( |
| 200 | monkeypatch: pytest.MonkeyPatch, vllm_tmp_cache: Path |
| 201 | ): |
| 202 | def native_reference(model: SparseActivationAotModel, x: torch.Tensor): |
| 203 | projected = torch.nn.functional.linear(x, model.up_proj.weight) |
| 204 | activated = ops.gelu_and_mul_sparse.impls["native"].impl_fn( |
| 205 | projected, model.activation.std_multiplier, "tanh" |
| 206 | ) |
| 207 | return torch.nn.functional.linear(activated, model.down_proj.weight) |
| 208 | |
| 209 | with monkeypatch.context() as m: |
| 210 | m.setenv("VLLM_USE_AOT_COMPILE", "1") |
| 211 | m.setenv("VLLM_USE_MEGA_AOT_ARTIFACT", "1") |
| 212 | m.setenv("VLLM_USE_STANDALONE_COMPILE", "1") |
| 213 | m.setenv("VLLM_CACHE_ROOT", str(vllm_tmp_cache / "vllm_cache")) |
| 214 | |
| 215 | vllm_config = make_vllm_config() |
| 216 | torch.manual_seed(0) |
| 217 | inputs = [ |
| 218 | torch.randn(4, 32, device="cuda", dtype=torch.bfloat16), |
| 219 | torch.randn(7, 32, device="cuda", dtype=torch.bfloat16), |
| 220 | ] |
| 221 | |
| 222 | with ( |
| 223 | use_vllm_config(vllm_config), |
| 224 | vllm_config.kernel_config.ir_op_priority.set_priority(), |
| 225 | ir.enable_torch_wrap(True), |
| 226 | ): |
| 227 | compiled = SparseActivationAotModel(vllm_config=vllm_config).to( |
| 228 | device="cuda", dtype=torch.bfloat16 |
| 229 | ) |
| 230 | compiled.requires_grad_(False) |
| 231 | expected = [native_reference(compiled, x) for x in inputs] |
| 232 | actual = [compiled(x) for x in inputs] |
| 233 | |
| 234 | for result, reference in zip(actual, expected): |
| 235 | assert torch.isfinite(result).all() |
| 236 | torch.testing.assert_close( |
| 237 | result, |
| 238 | reference, |
| 239 | **ops.gelu_and_mul_sparse.get_tolerance(result.dtype), |
| 240 | ) |
| 241 | |
| 242 | |
| 243 | @pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10") |
| 244 | def test_save_and_load_slice(monkeypatch: pytest.MonkeyPatch): |
| 245 | from torch._subclasses import FakeTensorMode |
| 246 | from torch.fx.experimental.symbolic_shapes import ShapeEnv |
| 247 | |
| 248 | def foo(x: torch.Tensor): |
| 249 | return x[slice(0, x.shape[0])] |
| 250 | |
| 251 | vllm_config = make_vllm_config() |
| 252 | |
| 253 | example_input = torch.randn(10, 10) |
| 254 | torch._dynamo.mark_dynamic(example_input, 0) |
| 255 | gm = torch.fx.symbolic_trace(foo) |
| 256 | assert "getitem_1 = x[slice(0, getitem, None)]" in gm.code |
| 257 | with use_vllm_config(vllm_config): |
| 258 | payload = VllmSerializableFunction.serialize_graph_module(gm) |
| 259 | fake_mode = FakeTensorMode(shape_env=ShapeEnv()) |
| 260 | loaded_gm = VllmSerializableFunction.deserialize_graph_module( |
| 261 | payload, fake_mode |
| 262 | ) |
| 263 | |
| 264 | assert gm.code == loaded_gm.code |
| 265 | |
| 266 | |
| 267 | @pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10") |
| 268 | def test_cache_load_returns_tuple_consistency_tuple_output( |
| 269 | monkeypatch: pytest.MonkeyPatch, |
| 270 | ): |
| 271 | """Test that cache loading correctly handles models that return tuples. |
| 272 | |
| 273 | This verifies that when a model returns a tuple of tensors, the output |
| 274 | type is preserved as a tuple between fresh compilation and cache load. |
| 275 | """ |
| 276 | with monkeypatch.context() as m: |
| 277 | args = (torch.randn(10, 10),) |
| 278 | |
| 279 | with tempfile.TemporaryDirectory() as tmpdirname: |
| 280 | m.setenv("VLLM_CACHE_ROOT", tmpdirname) |
| 281 | m.setenv("VLLM_USE_AOT_COMPILE", "1") |
| 282 | m.setenv("VLLM_USE_MEGA_AOT_ARTIFACT", "1") |
| 283 | m.setenv("VLLM_USE_STANDALONE_COMPILE", "1") |
| 284 | vllm_config = make_vllm_config() |
| 285 | |
| 286 | # Fresh compilation with tuple-returning model |
| 287 | with use_vllm_config(vllm_config): |
| 288 | compiled_mod = CompiledModTuple(vllm_config=vllm_config) |
| 289 | fresh_result = compiled_mod(*args) |
| 290 | fresh_result_type = type(fresh_result) |
| 291 | |
| 292 | # Verify fresh result is a tuple |
| 293 | assert isinstance(fresh_result, tuple), ( |
| 294 | f"Fresh compile should return tuple, got {fresh_result_type}" |
| 295 | ) |
| 296 | assert len(fresh_result) == 2, ( |
| 297 | f"Fresh compile should return 2-tuple, got {len(fresh_result)}" |
| 298 | ) |
| 299 | |
| 300 | disable_envs_cache() |
| 301 | |
| 302 | # Load from cache |
| 303 | m.setenv("VLLM_FORCE_AOT_LOAD", "1") |
| 304 | vllm_config = make_vllm_config() |
| 305 | with use_vllm_config(vllm_config): |
| 306 | cached_mod = CompiledModTuple(vllm_config=vllm_config) |
| 307 | cached_result = cached_mod(*args) |
| 308 | cached_result_type = type(cached_result) |
| 309 | |
| 310 | # Verify cache was actually loaded |
| 311 | assert cached_mod.was_aot_compile_fn_loaded_from_disk, ( |
| 312 | "Expected was_aot_compile_fn_loaded_from_disk to be True after " |
| 313 | "loading from cache" |
| 314 | ) |
| 315 | |
| 316 | # Verify cached result is also a tuple |
| 317 | assert isinstance(cached_result, tuple), ( |
| 318 | f"Cache load should return tuple, got {cached_result_type}. " |
| 319 | "This indicates the returns_tuple logic is not preserving " |
| 320 | "tuple outputs when loading from cache." |
| 321 | ) |
| 322 | assert len(cached_result) == 2, ( |
| 323 | f"Cache load should return 2-tuple, got {len(cached_result)}" |
| 324 | ) |
| 325 | |
| 326 | # Verify values match |
| 327 | assert torch.allclose(cached_result[0], fresh_result[0]), ( |
| 328 | "Cached result[0] values should match fresh compilation" |
| 329 | ) |
| 330 | assert torch.allclose(cached_result[1], fresh_result[1]), ( |
| 331 | "Cached result[1] values should match fresh compilation" |
| 332 | ) |
| 333 | |
| 334 | |
| 335 | @pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10") |
| 336 | def test_shape_env(monkeypatch: pytest.MonkeyPatch): |
| 337 | """Test that the shape environment is correctly serialized and preserved |
| 338 | when loading from cache. |
| 339 | """ |
| 340 | with monkeypatch.context() as m: |
| 341 | args = (torch.randn(10, 10),) |
| 342 | |
| 343 | with tempfile.TemporaryDirectory() as tmpdirname: |
| 344 | m.setenv("VLLM_CACHE_ROOT", tmpdirname) |
| 345 | m.setenv("VLLM_USE_AOT_COMPILE", "1") |
| 346 | m.setenv("VLLM_USE_MEGA_AOT_ARTIFACT", "1") |
| 347 | m.setenv("VLLM_USE_STANDALONE_COMPILE", "1") |
| 348 | vllm_config = make_vllm_config() |
| 349 | with use_vllm_config(vllm_config): |
| 350 | compiled_mod = CompiledMod(vllm_config=vllm_config) |
| 351 | compiled_mod(*args) |
| 352 | artifacts = compiled_mod.aot_compiled_fn._artifacts |
| 353 | guards_string = artifacts.compiled_fn.shape_env.format_guards() |
| 354 | assert guards_string == " - s77 <= 42\n - Eq(Mod(s77, 2), 0)" |
| 355 | |
| 356 | disable_envs_cache() |
| 357 | |
| 358 | m.setenv("VLLM_FORCE_AOT_LOAD", "1") |
| 359 | vllm_config = make_vllm_config() |
| 360 | with use_vllm_config(vllm_config): |
| 361 | compiled_mod = CompiledMod(vllm_config=vllm_config) |
| 362 | compiled_mod(*args) |
| 363 | assert compiled_mod.was_aot_compile_fn_loaded_from_disk, ( |
| 364 | "Expected was_aot_compile_fn_loaded_from_disk to be True" |
| 365 | ) |
| 366 | artifacts = compiled_mod.aot_compiled_fn._artifacts |
| 367 | guards_string = artifacts.compiled_fn.shape_env.format_guards() |
| 368 | assert guards_string == " - s77 <= 42\n - Eq(Mod(s77, 2), 0)" |
| 369 | |
| 370 | |
| 371 | @pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10") |
| 372 | def test_partition_wrapper_applied_on_aot_load( |
| 373 | monkeypatch: pytest.MonkeyPatch, vllm_tmp_cache: Path, mocker |
| 374 | ): |
| 375 | """Test that partition wrappers are applied when loading AOT cached functions. |
| 376 | |
| 377 | This test verifies the fix for GitHub issue #31439 where AOT compile |
| 378 | caused 2x latency regression when use_inductor_graph_partition=True. |
| 379 | The root cause was that partition wrapper context was bypassed when |
| 380 | loading from AOT cache. |
| 381 | """ |
| 382 | from vllm.config import CUDAGraphMode |
| 383 | |
| 384 | args = (torch.randn(10, 10),) |
| 385 | monkeypatch.setenv("VLLM_USE_AOT_COMPILE", "1") |
| 386 | |
| 387 | # Create config with partition enabled |
| 388 | vllm_config = VllmConfig( |
| 389 | compilation_config=CompilationConfig( |
| 390 | mode=CompilationMode.VLLM_COMPILE, |
| 391 | use_inductor_graph_partition=True, |
| 392 | cudagraph_mode=CUDAGraphMode.PIECEWISE, |
| 393 | ) |
| 394 | ) |
| 395 | |
| 396 | # First compilation - save to cache |
| 397 | with use_vllm_config(vllm_config): |
| 398 | compiled_mod = CompiledMod(vllm_config=vllm_config) |
| 399 | compiled_mod(*args) |
| 400 | |
| 401 | disable_envs_cache() |
| 402 | |
| 403 | # Second run - load from cache, verify partition wrapper applied |
| 404 | monkeypatch.setenv("VLLM_FORCE_AOT_LOAD", "1") |
| 405 | vllm_config = VllmConfig( |
| 406 | compilation_config=CompilationConfig( |
| 407 | mode=CompilationMode.VLLM_COMPILE, |
| 408 | use_inductor_graph_partition=True, |
| 409 | cudagraph_mode=CUDAGraphMode.PIECEWISE, |
| 410 | ) |
| 411 | ) |
| 412 | |
| 413 | # Use mocker to spy on set_customized_partition_wrappers |
| 414 | spy = mocker.spy(torch._inductor.utils, "set_customized_partition_wrappers") |
| 415 | |
| 416 | with use_vllm_config(vllm_config): |
| 417 | compiled_mod = CompiledMod(vllm_config=vllm_config) |
| 418 | |
| 419 | # First call after restart: loads from AOT cache. |
| 420 | # This tests the fix for the first call after a restart. |
| 421 | compiled_mod(*args) |
| 422 | |
| 423 | # Verify cache was loaded |
| 424 | assert compiled_mod.was_aot_compile_fn_loaded_from_disk, ( |
| 425 | "Expected was_aot_compile_fn_loaded_from_disk to be True" |
| 426 | ) |
| 427 | |
| 428 | # Verify partition wrapper was called on AOT load. |
| 429 | assert spy.call_count >= 2, ( |
| 430 | "Expected partition wrapper to be set and cleared on AOT load, " |
| 431 | f"got {spy.call_count} calls" |
| 432 | ) |
| 433 | # First call should set a wrapper, last call should clear it |
| 434 | assert spy.call_args_list[0][0][0] is not None, ( |
| 435 | "First call on AOT load should set a wrapper function" |
| 436 | ) |
| 437 | assert spy.call_args_list[-1][0][0] is None, ( |
| 438 | "Last call on AOT load should clear the wrapper" |
| 439 | ) |
| 440 | |
| 441 | # Reset for the next check. |
| 442 | spy.reset_mock() |
| 443 | |
| 444 | # Subsequent call: uses the cached `aot_compiled_fn`. |
| 445 | # This tests the fix for subsequent calls. |
| 446 | compiled_mod(*args) |
| 447 | |
| 448 | # Verify partition wrapper was called on the subsequent call. |
| 449 | assert spy.call_count >= 2, ( |
| 450 | "Expected partition wrapper set and cleared on subsequent " |
| 451 | f"call, got {spy.call_count} calls" |
| 452 | ) |
| 453 | assert spy.call_args_list[0][0][0] is not None, ( |
| 454 | "First call on subsequent call should set a wrapper function" |
| 455 | ) |
| 456 | assert spy.call_args_list[-1][0][0] is None, ( |
| 457 | "Last call on subsequent call should clear the wrapper" |
| 458 | ) |
| 459 | |
| 460 | |
| 461 | @create_new_process_for_each_test("spawn") |
| 462 | def test_standalone_compile_correctness(): |
| 463 | """Outputs must match regardless of VLLM_USE_STANDALONE_COMPILE.""" |
| 464 | import json |
| 465 | |
| 466 | from ..utils import compare_two_settings |
| 467 | |
| 468 | compilation_config = json.dumps( |
| 469 | { |
| 470 | "mode": CompilationMode.VLLM_COMPILE, |
| 471 | } |
| 472 | ) |
| 473 | |
| 474 | common_args = [ |
| 475 | "--dtype", |
| 476 | "float16", |
| 477 | "--max-model-len", |
| 478 | "256", |
| 479 | "--compilation_config", |
| 480 | compilation_config, |
| 481 | ] |
| 482 | |
| 483 | compare_two_settings( |
| 484 | "facebook/opt-125m", |
| 485 | common_args, |
| 486 | common_args, |
| 487 | env1={"VLLM_USE_STANDALONE_COMPILE": "1"}, |
| 488 | env2={ |
| 489 | "VLLM_USE_STANDALONE_COMPILE": "0", |
| 490 | "VLLM_USE_MEGA_AOT_ARTIFACT": "0", |
| 491 | }, |
| 492 | ) |
| 493 | |
| 494 | |
| 495 | @pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10") |
| 496 | @create_new_process_for_each_test("spawn") |
| 497 | def test_gpt2_cache_hit(monkeypatch: pytest.MonkeyPatch): |
| 498 | """Test that compiling gpt2 twice results in a cache hit. |
| 499 | |
| 500 | Counter values are read from the EngineCore subprocess via |
| 501 | ``LLM.collective_rpc`` so the test works under default V1 |
| 502 | multiprocessing (no shared memory between test and engine). |
| 503 | """ |
| 504 | from vllm import LLM |
| 505 | |
| 506 | def _snap(self): |
| 507 | from vllm.compilation.counter import compilation_counter |
| 508 | |
| 509 | return ( |
| 510 | compilation_counter.num_aot_compiles, |
| 511 | compilation_counter.num_aot_artifacts_saved, |
| 512 | compilation_counter.num_aot_artifacts_loaded, |
| 513 | ) |
| 514 | |
| 515 | # collective_rpc(callable) requires pickle-based serialization. |
| 516 | monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1") |
| 517 | |
| 518 | with monkeypatch.context() as m, tempfile.TemporaryDirectory() as tmpdirname: |
| 519 | m.setenv("VLLM_CACHE_ROOT", tmpdirname) |
| 520 | m.setenv("VLLM_USE_AOT_COMPILE", "1") |
| 521 | # First compilation - initialize model and generate |
| 522 | llm_model = LLM( |
| 523 | model="openai-community/gpt2", |
| 524 | compilation_config=CompilationConfig( |
| 525 | mode=CompilationMode.VLLM_COMPILE, |
| 526 | ), |
| 527 | max_model_len=256, |
| 528 | ) |
| 529 | |
| 530 | llm_model.generate("Hello, my name is") |
| 531 | assert llm_model.collective_rpc(_snap)[0] == (1, 1, 0) |
| 532 | |
| 533 | # Clean up first model |
| 534 | del llm_model |
| 535 | disable_envs_cache() |
| 536 | |
| 537 | # Second compilation - should hit cache |
| 538 | m.setenv("VLLM_FORCE_AOT_LOAD", "1") |
| 539 | llm_model = LLM( |
| 540 | model="openai-community/gpt2", |
| 541 | compilation_config=CompilationConfig( |
| 542 | mode=CompilationMode.VLLM_COMPILE, |
| 543 | ), |
| 544 | max_model_len=256, |
| 545 | ) |
| 546 | llm_model.generate("Hello, my name is") |
| 547 | assert llm_model.collective_rpc(_snap)[0] == (0, 0, 1) |
| 548 | |
| 549 | |
| 550 | @pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10") |
| 551 | class TestStandaloneCompiledArtifacts: |
| 552 | def test_init(self): |
| 553 | cache = StandaloneCompiledArtifacts() |
| 554 | assert cache.submodule_bytes == {} |
| 555 | assert cache.submodule_bytes_store == {} |
| 556 | assert cache.loaded_submodule_store == {} |
| 557 | |
| 558 | def test_insert_new_artifact(self): |
| 559 | cache = StandaloneCompiledArtifacts() |
| 560 | test_data = b"test_artifact_data" |
| 561 | submod_name = "test_submod" |
| 562 | shape = "s1" |
| 563 | |
| 564 | hasher = hashlib.sha256() |
| 565 | hasher.update(test_data) |
| 566 | expected_hash = hasher.hexdigest() |
| 567 | |
| 568 | cache.insert(submod_name, shape, test_data) |
| 569 | |
| 570 | assert f"{submod_name}_{shape}" in cache.submodule_bytes |
| 571 | assert cache.submodule_bytes[f"{submod_name}_{shape}"] == expected_hash |
| 572 | assert expected_hash in cache.submodule_bytes_store |
| 573 | assert cache.submodule_bytes_store[expected_hash] == test_data |
| 574 | |
| 575 | def test_insert_duplicate_artifact(self): |
| 576 | cache = StandaloneCompiledArtifacts() |
| 577 | |
| 578 | test_data = b"duplicate_test_data" |
| 579 | submod_name1 = "submod1" |
| 580 | submod_name2 = "submod2" |
| 581 | shape = "s2" |
| 582 | |
| 583 | cache.insert(submod_name1, shape, test_data) |
| 584 | cache.insert(submod_name2, shape, test_data) |
| 585 | |
| 586 | hash1 = cache.submodule_bytes[f"{submod_name1}_{shape}"] |
| 587 | hash2 = cache.submodule_bytes[f"{submod_name2}_{shape}"] |
| 588 | assert hash1 == hash2 |
| 589 | |
| 590 | assert len(cache.submodule_bytes_store) == 1 |
| 591 | assert len(cache.submodule_bytes) == 2 |
| 592 | |
| 593 | def test_get_artifact(self): |
| 594 | cache = StandaloneCompiledArtifacts() |
| 595 | test_data = b"retrievable_data" |
| 596 | submod_name = "mod1" |
| 597 | shape = "shape16" |
| 598 | |
| 599 | cache.insert(submod_name, shape, test_data) |
| 600 | retrieved_data = cache.get(submod_name, shape) |
| 601 | |
| 602 | assert retrieved_data == test_data |
| 603 | |
| 604 | def test_get_nonexistent_artifact(self): |
| 605 | cache = StandaloneCompiledArtifacts() |
| 606 | |
| 607 | with pytest.raises(KeyError): |
| 608 | cache.get("nonexistent", "shape") |
| 609 | |
| 610 | def test_size_bytes(self): |
| 611 | cache = StandaloneCompiledArtifacts() |
| 612 | |
| 613 | assert cache.size_bytes() == 0 |
| 614 | |
| 615 | data1 = b"x" * 100 |
| 616 | data2 = b"y" * 200 |
| 617 | cache.insert("mod1", "shape1", data1) |
| 618 | cache.insert("mod2", "shape2", data2) |
| 619 | |
| 620 | assert cache.size_bytes() == 300 |
| 621 | |
| 622 | def test_num_artifacts_and_entries(self): |
| 623 | cache = StandaloneCompiledArtifacts() |
| 624 | |
| 625 | assert cache.num_artifacts() == 0 |
| 626 | assert cache.num_entries() == 0 |
| 627 | |
| 628 | cache.insert("mod1", "shape1", b"data1") |
| 629 | cache.insert("mod2", "shape2", b"data2") |
| 630 | assert cache.num_artifacts() == 2 |
| 631 | assert cache.num_entries() == 2 |
| 632 | |
| 633 | cache.insert("mod3", "shape3", b"data1") |
| 634 | assert cache.num_artifacts() == 2 |
| 635 | assert cache.num_entries() == 3 |
| 636 | |
| 637 | @patch("torch._inductor.standalone_compile.AOTCompiledArtifact.deserialize") |
| 638 | def test_load_all_success(self, mock_deserialize): |
| 639 | """Test successful loading of all artifacts.""" |
| 640 | cache = StandaloneCompiledArtifacts() |
| 641 | |
| 642 | mock_artifact1 = Mock() |
| 643 | mock_artifact2 = Mock() |
| 644 | mock_deserialize.side_effect = [mock_artifact1, mock_artifact2] |
| 645 | |
| 646 | cache.insert("mod1", "shape1", pickle.dumps(b"data1")) |
| 647 | cache.insert("mod2", "shape2", pickle.dumps(b"data2")) |
| 648 | |
| 649 | cache.load_all() |
| 650 | |
| 651 | assert len(cache.loaded_submodule_store) == 2 |
| 652 | assert mock_deserialize.call_count == 2 |
| 653 | |
| 654 | @patch("torch._inductor.standalone_compile.AOTCompiledArtifact.deserialize") |
| 655 | def test_load_all_already_loaded(self, mock_deserialize): |
| 656 | """Test that load_all skips if already loaded.""" |
| 657 | cache = StandaloneCompiledArtifacts() |
| 658 | |
| 659 | mock_artifact = Mock() |
| 660 | cache.submodule_bytes_store["hash1"] = pickle.dumps(b"data1") |
| 661 | cache.loaded_submodule_store["hash1"] = mock_artifact |
| 662 | |
| 663 | cache.load_all() |
| 664 | |
| 665 | mock_deserialize.assert_not_called() |
| 666 | |
| 667 | @patch("torch._inductor.standalone_compile.AOTCompiledArtifact.deserialize") |
| 668 | def test_get_loaded_artifact(self, mock_deserialize): |
| 669 | """Test retrieving loaded artifacts.""" |
| 670 | cache = StandaloneCompiledArtifacts() |
| 671 | |
| 672 | mock_artifact = Mock() |
| 673 | mock_deserialize.return_value = mock_artifact |
| 674 | |
| 675 | submod_name = "test_mod" |
| 676 | shape = "test_shape" |
| 677 | cache.insert(submod_name, shape, pickle.dumps(b"test_data")) |
| 678 | cache.load_all() |
| 679 | |
| 680 | retrieved_artifact = cache.get_loaded(submod_name, shape) |
| 681 | assert retrieved_artifact == mock_artifact |
| 682 | |
| 683 | def test_getstate_setstate(self): |
| 684 | cache = StandaloneCompiledArtifacts() |
| 685 | |
| 686 | cache.insert("mod1", "shape1", b"data1") |
| 687 | cache.insert("mod2", "shape2", b"data2") |
| 688 | |
| 689 | cache.loaded_submodule_store["hash1"] = Mock() |
| 690 | |
| 691 | state = cache.__getstate__() |
| 692 | |
| 693 | assert "submodule_bytes" in state |
| 694 | assert "submodule_bytes_store" in state |
| 695 | assert "loaded_submodule_store" not in state |
| 696 | |
| 697 | new_cache = StandaloneCompiledArtifacts() |
| 698 | new_cache.__setstate__(state) |
| 699 | |
| 700 | assert new_cache.submodule_bytes == cache.submodule_bytes |
| 701 | assert new_cache.submodule_bytes_store == cache.submodule_bytes_store |
| 702 | assert new_cache.loaded_submodule_store == {} |
| 703 | |
| 704 | def test_pickle_roundtrip(self): |
| 705 | cache = StandaloneCompiledArtifacts() |
| 706 | |
| 707 | test_data1 = b"pickle_test_data_1" |
| 708 | test_data2 = b"pickle_test_data_2" |
| 709 | cache.insert("mod1", "shape1", test_data1) |
| 710 | cache.insert("mod2", "shape2", test_data2) |
| 711 | |
| 712 | pickled_data = pickle.dumps(cache) |
| 713 | restored_cache = pickle.loads(pickled_data) |
| 714 | |
| 715 | assert restored_cache.get("mod1", "shape1") == test_data1 |
| 716 | assert restored_cache.get("mod2", "shape2") == test_data2 |
| 717 | assert restored_cache.num_artifacts() == cache.num_artifacts() |
| 718 | assert restored_cache.num_entries() == cache.num_entries() |
| 719 | assert restored_cache.size_bytes() == cache.size_bytes() |
| 720 | |
| 721 | assert len(restored_cache.loaded_submodule_store) == 0 |
| 722 | |
| 723 | |
| 724 | @pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10") |
| 725 | class TestStandaloneCompiledArtifactsIntegration: |
| 726 | def test_add_pickle_unpickle(self): |
| 727 | cache = StandaloneCompiledArtifacts() |
| 728 | |
| 729 | artifacts = { |
| 730 | ("mod1", "shape1"): b"m1s1_artifact", |
| 731 | ("mod1", "shape2"): b"m1s2_artifact", |
| 732 | ("mod2", "shape1"): b"m2s1_artifact", |
| 733 | ("mod2", "shape2"): b"m2s2_artifact", |
| 734 | } |
| 735 | |
| 736 | for (submod, shape), data in artifacts.items(): |
| 737 | cache.insert(submod, shape, data) |
| 738 | |
| 739 | assert cache.num_entries() == 4 |
| 740 | assert cache.num_artifacts() == 4 |
| 741 | |
| 742 | for (submod, shape), expected_data in artifacts.items(): |
| 743 | retrieved_data = cache.get(submod, shape) |
| 744 | assert retrieved_data == expected_data |
| 745 | |
| 746 | pickled = pickle.dumps(cache) |
| 747 | restored_cache = pickle.loads(pickled) |
| 748 | |
| 749 | for (submod, shape), expected_data in artifacts.items(): |
| 750 | retrieved_data = restored_cache.get(submod, shape) |
| 751 | assert retrieved_data == expected_data |
| 752 | |
| 753 | def test_deduplication(self): |
| 754 | cache = StandaloneCompiledArtifacts() |
| 755 | |
| 756 | shared_data = b"shared_artifact_data" * 1000 |
| 757 | |
| 758 | cache.insert("mod1", "shape1", shared_data) |
| 759 | cache.insert("mod2", "shape1", shared_data) |
| 760 | cache.insert("mod1", "shape2", shared_data) |
| 761 | cache.insert("mod3", "shape3", shared_data) |
| 762 | |
| 763 | assert cache.num_entries() == 4 |
| 764 | assert cache.num_artifacts() == 1 |
| 765 | assert cache.size_bytes() == len(shared_data) |
| 766 | |
| 767 | for submod, shape in [ |
| 768 | ("mod1", "shape1"), |
| 769 | ("mod2", "shape1"), |
| 770 | ("mod1", "shape2"), |
| 771 | ("mod3", "shape3"), |
| 772 | ]: |
| 773 | assert cache.get(submod, shape) == shared_data |
| 774 | |
| 775 | @pytest.mark.skipif( |
| 776 | envs.VLLM_USE_MEGA_AOT_ARTIFACT, |
| 777 | reason="There's no AOT Autograd run with mega artifact", |
| 778 | ) |
| 779 | def test_functorch_config(self): |
| 780 | vllm_config = make_vllm_config() |
| 781 | example_inputs = (torch.randn(10, 10),) |
| 782 | |
| 783 | def add_1(x: torch.Tensor): |
| 784 | return x + 1 |
| 785 | |
| 786 | gm = torch._dynamo.functional_export.dynamo_graph_capture_for_export(add_1)( |
| 787 | *example_inputs |
| 788 | ) |
| 789 | |
| 790 | gm.graph._codegen = torch.fx.graph.CodeGen() |
| 791 | gm._dynamo_bytecode_flatten = None |
| 792 | gm._dynamo_bytecode_unflatten = None |
| 793 | |
| 794 | with ( |
| 795 | torch._functorch.config.patch(bundled_autograd_cache=False), |
| 796 | set_current_vllm_config(vllm_config), |
| 797 | ): |
| 798 | with torch._functorch.config.patch(bundled_autograd_cache=True): |
| 799 | fn = VllmSerializableFunction(gm, example_inputs, "", add_1) |
| 800 | |
| 801 | payload = VllmSerializableFunction.serialize_compile_artifacts(fn) |
| 802 | |
| 803 | config = None |
| 804 | |
| 805 | def backend(*args, **kwargs) -> VllmSerializableFunction: |
| 806 | nonlocal config |
| 807 | # bundled_autograd_cache should be True even compiler backend |
| 808 | # runs with bundled_autograd_cache=False in ambient context. |
| 809 | config = torch._functorch.config.save_config_portable() |
| 810 | return fn |
| 811 | |
| 812 | loaded_fn = VllmSerializableFunction.deserialize_compile_artifacts(payload) |
| 813 | with patch.object(VllmBackend, "__call__", backend): |
| 814 | loaded_fn(*example_inputs) |
| 815 | |
| 816 | assert isinstance(config, dict) |
| 817 | assert "bundled_autograd_cache" in config |
| 818 | assert config["bundled_autograd_cache"] is True |
| 819 | |
| 820 | |
| 821 | @pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10") |
| 822 | def test_disable_compile_cache_skips_aot_save( |
| 823 | monkeypatch: pytest.MonkeyPatch, fresh_vllm_cache: str |
| 824 | ): |
| 825 | """When VLLM_DISABLE_COMPILE_CACHE=1, AOT artifacts must not be saved.""" |
| 826 | monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1") |
| 827 | monkeypatch.setenv("VLLM_USE_AOT_COMPILE", "1") |
| 828 | disable_envs_cache() |
| 829 | |
| 830 | args = (torch.randn(10, 10),) |
| 831 | expected = reference_fn(*args) |
| 832 | vllm_config = make_vllm_config() |
| 833 | |
| 834 | with ( |
| 835 | use_vllm_config(vllm_config), |
| 836 | compilation_counter.expect( |
| 837 | num_aot_compiles=1, |
| 838 | num_aot_artifacts_saved=0, |
| 839 | num_aot_artifacts_loaded=0, |
| 840 | ), |
| 841 | ): |
| 842 | mod = CompiledMod(vllm_config=vllm_config) |
| 843 | actual = mod(*args) |
| 844 | |
| 845 | assert torch.allclose(actual, expected) |
| 846 | |
| 847 | # No cached artifact should exist on disk |
| 848 | aot_dir = os.path.join(fresh_vllm_cache, "torch_compile_cache", "torch_aot_compile") |
| 849 | if os.path.isdir(aot_dir): |
| 850 | for root, _dirs, files in os.walk(aot_dir): |
| 851 | for f in files: |
| 852 | assert f != "model", ( |
| 853 | f"AOT artifact unexpectedly saved at {os.path.join(root, f)}" |
| 854 | ) |
| 855 | |
| 856 | |
| 857 | @pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10") |
| 858 | def test_disable_compile_cache_skips_aot_load( |
| 859 | monkeypatch: pytest.MonkeyPatch, fresh_vllm_cache: str |
| 860 | ): |
| 861 | """When VLLM_DISABLE_COMPILE_CACHE=1, AOT artifacts must not be loaded.""" |
| 862 | # Phase 1: compile and save with cache enabled |
| 863 | monkeypatch.setenv("VLLM_USE_AOT_COMPILE", "1") |
| 864 | disable_envs_cache() |
| 865 | |
| 866 | args = (torch.randn(10, 10),) |
| 867 | vllm_config = make_vllm_config() |
| 868 | |
| 869 | with ( |
| 870 | use_vllm_config(vllm_config), |
| 871 | compilation_counter.expect(num_aot_artifacts_saved=1), |
| 872 | ): |
| 873 | CompiledMod(vllm_config=vllm_config)(*args) |
| 874 | |
| 875 | # Phase 2: disable cache, compile again — should NOT load from disk |
| 876 | monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1") |
| 877 | disable_envs_cache() |
| 878 | torch._dynamo.reset() |
| 879 | |
| 880 | vllm_config = make_vllm_config() |
| 881 | with ( |
| 882 | use_vllm_config(vllm_config), |
| 883 | compilation_counter.expect( |
| 884 | num_aot_compiles=1, |
| 885 | num_aot_artifacts_saved=0, |
| 886 | num_aot_artifacts_loaded=0, |
| 887 | ), |
| 888 | ): |
| 889 | mod = CompiledMod(vllm_config=vllm_config) |
| 890 | mod(*args) |
| 891 | |
| 892 | assert not mod.was_aot_compile_fn_loaded_from_disk |