|
1 | 1 | # SPDX-License-Identifier: Apache-2.0 |
2 | 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
3 | 3 |
|
| 4 | +import hashlib |
| 5 | +import pickle |
4 | 6 | import tempfile |
5 | 7 | from contextlib import contextmanager |
| 8 | +from unittest.mock import Mock, patch |
6 | 9 |
|
7 | 10 | import pytest |
8 | 11 | import torch |
9 | 12 |
|
| 13 | +from vllm.compilation.backends import VllmCompiledFunction |
10 | 14 | from vllm.compilation.decorators import support_torch_compile |
11 | 15 | from vllm.config import (CompilationConfig, CompilationLevel, VllmConfig, |
12 | 16 | set_current_vllm_config) |
@@ -91,7 +95,251 @@ def test_basic(monkeypatch: pytest.MonkeyPatch): |
91 | 95 | m.setenv("VLLM_USE_AOT_COMPILE", "1") |
92 | 96 | vllm_config = make_vllm_config() |
93 | 97 | with use_vllm_config(vllm_config): |
94 | | - expected = CompiledMod(vllm_config=vllm_config)(*args) |
| 98 | + mod = CompiledMod(vllm_config=vllm_config) |
| 99 | + expected = mod(*args) |
95 | 100 | m.setenv("VLLM_FORCE_AOT_LOAD", "1") |
96 | | - ret = CompiledMod(vllm_config=vllm_config)(*args) |
| 101 | + ret = mod(*args) |
97 | 102 | assert torch.allclose(ret, expected) |
| 103 | + |
| 104 | + |
| 105 | +class TestInductorCache: |
| 106 | + |
| 107 | + def test_init(self): |
| 108 | + cache = VllmCompiledFunction.InductorCache() |
| 109 | + assert cache.submodule_bytes == {} |
| 110 | + assert cache.submodule_bytes_store == {} |
| 111 | + assert cache.loaded_submodule_store == {} |
| 112 | + |
| 113 | + def test_insert_new_artifact(self): |
| 114 | + cache = VllmCompiledFunction.InductorCache() |
| 115 | + test_data = b"test_artifact_data" |
| 116 | + submod_name = "test_submod" |
| 117 | + shape = "s1" |
| 118 | + |
| 119 | + hasher = hashlib.sha256() |
| 120 | + hasher.update(test_data) |
| 121 | + expected_hash = hasher.hexdigest() |
| 122 | + |
| 123 | + cache.insert(submod_name, shape, test_data) |
| 124 | + |
| 125 | + assert f"{submod_name}_{shape}" in cache.submodule_bytes |
| 126 | + assert cache.submodule_bytes[f"{submod_name}_{shape}"] == expected_hash |
| 127 | + assert expected_hash in cache.submodule_bytes_store |
| 128 | + assert cache.submodule_bytes_store[expected_hash] == test_data |
| 129 | + |
| 130 | + def test_insert_duplicate_artifact(self): |
| 131 | + cache = VllmCompiledFunction.InductorCache() |
| 132 | + |
| 133 | + test_data = b"duplicate_test_data" |
| 134 | + submod_name1 = "submod1" |
| 135 | + submod_name2 = "submod2" |
| 136 | + shape = "s2" |
| 137 | + |
| 138 | + cache.insert(submod_name1, shape, test_data) |
| 139 | + cache.insert(submod_name2, shape, test_data) |
| 140 | + |
| 141 | + hash1 = cache.submodule_bytes[f"{submod_name1}_{shape}"] |
| 142 | + hash2 = cache.submodule_bytes[f"{submod_name2}_{shape}"] |
| 143 | + assert hash1 == hash2 |
| 144 | + |
| 145 | + assert len(cache.submodule_bytes_store) == 1 |
| 146 | + assert len(cache.submodule_bytes) == 2 |
| 147 | + |
| 148 | + def test_get_artifact(self): |
| 149 | + |
| 150 | + cache = VllmCompiledFunction.InductorCache() |
| 151 | + test_data = b"retrievable_data" |
| 152 | + submod_name = "mod1" |
| 153 | + shape = "shape16" |
| 154 | + |
| 155 | + cache.insert(submod_name, shape, test_data) |
| 156 | + retrieved_data = cache.get(submod_name, shape) |
| 157 | + |
| 158 | + assert retrieved_data == test_data |
| 159 | + |
| 160 | + def test_get_nonexistent_artifact(self): |
| 161 | + cache = VllmCompiledFunction.InductorCache() |
| 162 | + |
| 163 | + with pytest.raises(KeyError): |
| 164 | + cache.get("nonexistent", "shape") |
| 165 | + |
| 166 | + def test_size_bytes(self): |
| 167 | + cache = VllmCompiledFunction.InductorCache() |
| 168 | + |
| 169 | + assert cache.size_bytes() == 0 |
| 170 | + |
| 171 | + data1 = b"x" * 100 |
| 172 | + data2 = b"y" * 200 |
| 173 | + cache.insert("mod1", "shape1", data1) |
| 174 | + cache.insert("mod2", "shape2", data2) |
| 175 | + |
| 176 | + assert cache.size_bytes() == 300 |
| 177 | + |
| 178 | + def test_num_artifacts_and_entries(self): |
| 179 | + cache = VllmCompiledFunction.InductorCache() |
| 180 | + |
| 181 | + assert cache.num_artifacts() == 0 |
| 182 | + assert cache.num_entries() == 0 |
| 183 | + |
| 184 | + cache.insert("mod1", "shape1", b"data1") |
| 185 | + cache.insert("mod2", "shape2", b"data2") |
| 186 | + assert cache.num_artifacts() == 2 |
| 187 | + assert cache.num_entries() == 2 |
| 188 | + |
| 189 | + cache.insert("mod3", "shape3", b"data1") |
| 190 | + assert cache.num_artifacts() == 2 |
| 191 | + assert cache.num_entries() == 3 |
| 192 | + |
| 193 | + @patch("torch._inductor.CompiledArtifact.from_bytes") |
| 194 | + def test_load_all_success(self, mock_from_bytes): |
| 195 | + """Test successful loading of all artifacts""" |
| 196 | + cache = VllmCompiledFunction.InductorCache() |
| 197 | + |
| 198 | + mock_artifact1 = Mock() |
| 199 | + mock_artifact2 = Mock() |
| 200 | + mock_from_bytes.side_effect = [mock_artifact1, mock_artifact2] |
| 201 | + |
| 202 | + cache.insert("mod1", "shape1", b"data1") |
| 203 | + cache.insert("mod2", "shape2", b"data2") |
| 204 | + |
| 205 | + cache.load_all() |
| 206 | + |
| 207 | + assert len(cache.loaded_submodule_store) == 2 |
| 208 | + assert mock_from_bytes.call_count == 2 |
| 209 | + |
| 210 | + @patch("torch._inductor.CompiledArtifact.from_bytes") |
| 211 | + def test_load_all_with_retry(self, mock_from_bytes): |
| 212 | + """Test loading with retries on initial failure""" |
| 213 | + cache = VllmCompiledFunction.InductorCache() |
| 214 | + |
| 215 | + mock_artifact = Mock() |
| 216 | + mock_from_bytes.side_effect = [ |
| 217 | + Exception("First attempt fails"), mock_artifact |
| 218 | + ] |
| 219 | + |
| 220 | + cache.insert("mod1", "shape1", b"data1") |
| 221 | + cache.load_all() |
| 222 | + |
| 223 | + assert len(cache.loaded_submodule_store) == 1 |
| 224 | + assert mock_from_bytes.call_count == 2 |
| 225 | + |
| 226 | + @patch("torch._inductor.CompiledArtifact.from_bytes") |
| 227 | + def test_load_all_already_loaded(self, mock_from_bytes): |
| 228 | + """Test that load_all skips if already loaded""" |
| 229 | + cache = VllmCompiledFunction.InductorCache() |
| 230 | + |
| 231 | + mock_artifact = Mock() |
| 232 | + cache.submodule_bytes_store["hash1"] = b"data1" |
| 233 | + cache.loaded_submodule_store["hash1"] = mock_artifact |
| 234 | + |
| 235 | + cache.load_all() |
| 236 | + |
| 237 | + mock_from_bytes.assert_not_called() |
| 238 | + |
| 239 | + @patch("torch._inductor.CompiledArtifact.from_bytes") |
| 240 | + def test_get_loaded_artifact(self, mock_from_bytes): |
| 241 | + """Test retrieving loaded artifacts""" |
| 242 | + cache = VllmCompiledFunction.InductorCache() |
| 243 | + |
| 244 | + mock_artifact = Mock() |
| 245 | + mock_from_bytes.return_value = mock_artifact |
| 246 | + |
| 247 | + submod_name = "test_mod" |
| 248 | + shape = "test_shape" |
| 249 | + cache.insert(submod_name, shape, b"test_data") |
| 250 | + cache.load_all() |
| 251 | + |
| 252 | + retrieved_artifact = cache.get_loaded(submod_name, shape) |
| 253 | + assert retrieved_artifact == mock_artifact |
| 254 | + |
| 255 | + def test_getstate_setstate(self): |
| 256 | + cache = VllmCompiledFunction.InductorCache() |
| 257 | + |
| 258 | + cache.insert("mod1", "shape1", b"data1") |
| 259 | + cache.insert("mod2", "shape2", b"data2") |
| 260 | + |
| 261 | + cache.loaded_submodule_store["hash1"] = Mock() |
| 262 | + |
| 263 | + state = cache.__getstate__() |
| 264 | + |
| 265 | + assert "submodule_bytes" in state |
| 266 | + assert "submodule_bytes_store" in state |
| 267 | + assert "loaded_submodule_store" not in state |
| 268 | + |
| 269 | + new_cache = VllmCompiledFunction.InductorCache() |
| 270 | + new_cache.__setstate__(state) |
| 271 | + |
| 272 | + assert new_cache.submodule_bytes == cache.submodule_bytes |
| 273 | + assert new_cache.submodule_bytes_store == cache.submodule_bytes_store |
| 274 | + assert new_cache.loaded_submodule_store == {} |
| 275 | + |
| 276 | + def test_pickle_roundtrip(self): |
| 277 | + cache = VllmCompiledFunction.InductorCache() |
| 278 | + |
| 279 | + test_data1 = b"pickle_test_data_1" |
| 280 | + test_data2 = b"pickle_test_data_2" |
| 281 | + cache.insert("mod1", "shape1", test_data1) |
| 282 | + cache.insert("mod2", "shape2", test_data2) |
| 283 | + |
| 284 | + pickled_data = pickle.dumps(cache) |
| 285 | + restored_cache = pickle.loads(pickled_data) |
| 286 | + |
| 287 | + assert restored_cache.get("mod1", "shape1") == test_data1 |
| 288 | + assert restored_cache.get("mod2", "shape2") == test_data2 |
| 289 | + assert restored_cache.num_artifacts() == cache.num_artifacts() |
| 290 | + assert restored_cache.num_entries() == cache.num_entries() |
| 291 | + assert restored_cache.size_bytes() == cache.size_bytes() |
| 292 | + |
| 293 | + assert len(restored_cache.loaded_submodule_store) == 0 |
| 294 | + |
| 295 | + |
| 296 | +class TestInductorCacheIntegration: |
| 297 | + |
| 298 | + def test_add_pickle_unpickle(self): |
| 299 | + cache = VllmCompiledFunction.InductorCache() |
| 300 | + |
| 301 | + artifacts = { |
| 302 | + ("mod1", "shape1"): b"m1s1_artifact", |
| 303 | + ("mod1", "shape2"): b"m1s2_artifact", |
| 304 | + ("mod2", "shape1"): b"m2s1_artifact", |
| 305 | + ("mod2", "shape2"): b"m2s2_artifact", |
| 306 | + } |
| 307 | + |
| 308 | + for (submod, shape), data in artifacts.items(): |
| 309 | + cache.insert(submod, shape, data) |
| 310 | + |
| 311 | + assert cache.num_entries() == 4 |
| 312 | + assert cache.num_artifacts() == 4 |
| 313 | + |
| 314 | + for (submod, shape), expected_data in artifacts.items(): |
| 315 | + retrieved_data = cache.get(submod, shape) |
| 316 | + assert retrieved_data == expected_data |
| 317 | + |
| 318 | + pickled = pickle.dumps(cache) |
| 319 | + restored_cache = pickle.loads(pickled) |
| 320 | + |
| 321 | + for (submod, shape), expected_data in artifacts.items(): |
| 322 | + retrieved_data = restored_cache.get(submod, shape) |
| 323 | + assert retrieved_data == expected_data |
| 324 | + |
| 325 | + def test_deduplication(self): |
| 326 | + cache = VllmCompiledFunction.InductorCache() |
| 327 | + |
| 328 | + shared_data = b"shared_artifact_data" * 1000 |
| 329 | + |
| 330 | + cache.insert("mod1", "shape1", shared_data) |
| 331 | + cache.insert("mod2", "shape1", shared_data) |
| 332 | + cache.insert("mod1", "shape2", shared_data) |
| 333 | + cache.insert("mod3", "shape3", shared_data) |
| 334 | + |
| 335 | + assert cache.num_entries() == 4 |
| 336 | + assert cache.num_artifacts() == 1 |
| 337 | + assert cache.size_bytes() == len(shared_data) |
| 338 | + |
| 339 | + for submod, shape in [ |
| 340 | + ("mod1", "shape1"), |
| 341 | + ("mod2", "shape1"), |
| 342 | + ("mod1", "shape2"), |
| 343 | + ("mod3", "shape3"), |
| 344 | + ]: |
| 345 | + assert cache.get(submod, shape) == shared_data |
0 commit comments