| 1 | #!/usr/bin/env python3 |
| 2 | # SPDX-License-Identifier: Apache-2.0 |
| 3 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 4 | import sys |
| 5 | from dataclasses import dataclass, field |
| 6 | |
| 7 | import regex as re |
| 8 | |
| 9 | # Hub entry points that must go through the vLLM-tagged repo_utils helpers. |
| 10 | _HF_NAMES = ( |
| 11 | r"HfApi|HfFileSystem|hf_hub_download|snapshot_download" |
| 12 | r"|list_repo_files|file_exists|try_to_load_from_cache" |
| 13 | r"|list_repo_refs|repo_exists" |
| 14 | ) |
| 15 | |
| 16 | # Transformers v4 names kept as aliases in v5. |
| 17 | _TRANSFORMERS_LEGACY_NAMES = ( |
| 18 | r"PretrainedConfig|PreTrainedTokenizer|PreTrainedTokenizerFast" |
| 19 | r"|BaseImageProcessorFast|\w+ImageProcessorFast" |
| 20 | ) |
| 21 | _TRANSFORMERS_LEGACY_MODULES = ( |
| 22 | r"tokenization_utils|tokenization_utils_fast|image_processing_utils_fast" |
| 23 | r"|models\.\w+\.\w+_fast" |
| 24 | ) |
| 25 | |
| 26 | |
| 27 | @dataclass |
| 28 | class ForbiddenImport: |
| 29 | pattern: str |
| 30 | tip: str |
| 31 | allowed_pattern: re.Pattern = re.compile(r"^$") # matches nothing by default |
| 32 | allowed_files: set[str] = field(default_factory=set) |
| 33 | allowed_dirs: set[str] = field(default_factory=set) |
| 34 | |
| 35 | |
| 36 | CHECK_IMPORTS = { |
| 37 | "pickle/cloudpickle": ForbiddenImport( |
| 38 | pattern=( |
| 39 | r"^\s*(import\s+(pickle|cloudpickle)(\s|$|\sas)" |
| 40 | r"|from\s+(pickle|cloudpickle)\s+import\b)" |
| 41 | ), |
| 42 | tip=( |
| 43 | "Avoid using pickle or cloudpickle or add this file to " |
| 44 | "tools/pre_commit/check_forbidden_imports.py." |
| 45 | ), |
| 46 | allowed_files={ |
| 47 | # pickle |
| 48 | "vllm/multimodal/hasher.py", |
| 49 | "vllm/transformers_utils/config.py", |
| 50 | "vllm/model_executor/models/registry.py", |
| 51 | "vllm/compilation/caching.py", |
| 52 | "vllm/env_override.py", |
| 53 | "vllm/compilation/piecewise_backend.py", |
| 54 | "vllm/distributed/utils.py", |
| 55 | "vllm/distributed/parallel_state.py", |
| 56 | "vllm/distributed/device_communicators/all_reduce_utils.py", |
| 57 | "vllm/distributed/device_communicators/shm_broadcast.py", |
| 58 | "vllm/distributed/device_communicators/shm_object_storage.py", |
| 59 | "vllm/distributed/weight_transfer/ipc_engine.py", |
| 60 | "vllm/distributed/weight_transfer/clients.py", |
| 61 | "vllm/model_executor/model_loader/weight_cache/protocol.py", |
| 62 | "tests/distributed/test_shm_broadcast.py", |
| 63 | "tests/distributed/test_weight_transfer.py", |
| 64 | "vllm/utils/hashing.py", |
| 65 | "tests/multimodal/media/test_base.py", |
| 66 | "tests/tokenizers_/test_hf.py", |
| 67 | "tests/utils_/test_hashing.py", |
| 68 | "tests/compile/test_aot_compile.py", |
| 69 | "benchmarks/kernels/graph_machete_bench.py", |
| 70 | "benchmarks/kernels/benchmark_lora.py", |
| 71 | "benchmarks/kernels/benchmark_machete.py", |
| 72 | "benchmarks/fused_kernels/layernorm_rms_benchmarks.py", |
| 73 | # cloudpickle |
| 74 | "vllm/v1/executor/multiproc_executor.py", |
| 75 | "vllm/v1/executor/ray_executor.py", |
| 76 | "vllm/entrypoints/llm.py", |
| 77 | "tests/utils.py", |
| 78 | # pickle and cloudpickle |
| 79 | "vllm/v1/serial_utils.py", |
| 80 | }, |
| 81 | ), |
| 82 | "base64": ForbiddenImport( |
| 83 | pattern=r"^\s*(?:import\s+base64(?:$|\s|,)|from\s+base64\s+import)", |
| 84 | tip=( |
| 85 | "Replace 'import base64' with 'import pybase64' " |
| 86 | "or 'import pybase64 as base64'." |
| 87 | ), |
| 88 | allowed_pattern=re.compile(r"^\s*import\s+pybase64(\s*|\s+as\s+base64\s*)$"), |
| 89 | ), |
| 90 | "re": ForbiddenImport( |
| 91 | pattern=r"^\s*(?:import\s+re(?:$|\s|,)|from\s+re\s+import)", |
| 92 | tip="Replace 'import re' with 'import regex as re' or 'import regex'.", |
| 93 | allowed_pattern=re.compile(r"^\s*import\s+regex(\s*|\s+as\s+re\s*)$"), |
| 94 | allowed_files={"setup.py"}, |
| 95 | ), |
| 96 | "triton": ForbiddenImport( |
| 97 | pattern=r"^(from|import)\s+triton(\s|\.|$)", |
| 98 | tip="Use 'from vllm.triton_utils import triton' instead.", |
| 99 | allowed_pattern=re.compile( |
| 100 | "from vllm.triton_utils import (triton|tl|tl, triton)" |
| 101 | ), |
| 102 | allowed_files={"vllm/triton_utils/importing.py"}, |
| 103 | ), |
| 104 | "tilelang": ForbiddenImport( |
| 105 | pattern=r"^(from|import)\s+tilelang(\s|\.|$)", |
| 106 | tip="Use 'from vllm.tilelang_utils import tilelang, T' instead.", |
| 107 | allowed_pattern=re.compile( |
| 108 | r"from\s+vllm\.tilelang_utils\s+import\s+" |
| 109 | r"(tilelang|T|T, tilelang|tilelang, T)\b" |
| 110 | ), |
| 111 | allowed_files={"vllm/tilelang_utils/__init__.py"}, |
| 112 | ), |
| 113 | "huggingface_hub repo API": ForbiddenImport( |
| 114 | # Catch `from huggingface_hub import <fn>`, including parenthesized, |
| 115 | # multi-line imports. |
| 116 | pattern=( |
| 117 | r"^\s*from\s+huggingface_hub\s+import\s*\([^)]*\b(?:" + _HF_NAMES + r")\b" |
| 118 | r"|" |
| 119 | r"^\s*from\s+huggingface_hub\s+import\b[^\n]*\b(?:" + _HF_NAMES + r")\b" |
| 120 | ), |
| 121 | tip=( |
| 122 | "Use the shared, vLLM-tagged helpers from " |
| 123 | "vllm.transformers_utils.repo_utils (e.g. hf_api(), hf_fs(), " |
| 124 | "list_repo_files, file_exists) instead of calling " |
| 125 | "huggingface_hub directly." |
| 126 | ), |
| 127 | allowed_files={"vllm/transformers_utils/repo_utils.py"}, |
| 128 | allowed_dirs={"examples/"}, |
| 129 | ), |
| 130 | "transformers legacy names": ForbiddenImport( |
| 131 | pattern=( |
| 132 | r"^\s*from\s+transformers(?:\.[\w.]+)?\s+import\s*\([^)]*\b(?:" |
| 133 | + _TRANSFORMERS_LEGACY_NAMES |
| 134 | + r")\b" |
| 135 | r"|" |
| 136 | r"^\s*from\s+transformers(?:\.[\w.]+)?\s+import\b[^\n]*\b(?:" |
| 137 | + _TRANSFORMERS_LEGACY_NAMES |
| 138 | + r")\b" |
| 139 | r"|" |
| 140 | r"^\s*(?:from|import)\s+transformers\.(?:" |
| 141 | + _TRANSFORMERS_LEGACY_MODULES |
| 142 | + r")\b" |
| 143 | ), |
| 144 | tip=( |
| 145 | "Use the Transformers v5 names, e.g. PreTrainedConfig, PythonBackend, " |
| 146 | "TokenizersBackend, TorchvisionBackend, tokenization_utils_tokenizers, " |
| 147 | "tokenization_utils_sentencepiece and image_processing_backends." |
| 148 | ), |
| 149 | ), |
| 150 | } |
| 151 | |
| 152 | |
| 153 | def check_file(path: str) -> int: |
| 154 | with open(path, encoding="utf-8") as f: |
| 155 | content = f.read() |
| 156 | return_code = 0 |
| 157 | # Check all patterns in the whole file |
| 158 | for import_name, forbidden_import in CHECK_IMPORTS.items(): |
| 159 | # Skip files that are allowed for this import |
| 160 | if path in forbidden_import.allowed_files: |
| 161 | continue |
| 162 | # Skip directories that are allowed for this import |
| 163 | if any(path.startswith(prefix) for prefix in forbidden_import.allowed_dirs): |
| 164 | continue |
| 165 | # Search for forbidden imports |
| 166 | for match in re.finditer(forbidden_import.pattern, content, re.MULTILINE): |
| 167 | # Check if it's allowed |
| 168 | if forbidden_import.allowed_pattern.match(match.group()): |
| 169 | continue |
| 170 | # Skip matches inside a comment |
| 171 | line_start = content.rfind("\n", 0, match.start()) + 1 |
| 172 | if "#" in content[line_start : match.start()]: |
| 173 | continue |
| 174 | # Calculate line number from match position |
| 175 | line_num = content[: match.start() + 1].count("\n") + 1 |
| 176 | print( |
| 177 | f"{path}:{line_num}: " |
| 178 | "\033[91merror:\033[0m " # red color |
| 179 | f"Found forbidden import: {import_name}. {forbidden_import.tip}" |
| 180 | ) |
| 181 | return_code = 1 |
| 182 | return return_code |
| 183 | |
| 184 | |
| 185 | def main(): |
| 186 | returncode = 0 |
| 187 | for path in sys.argv[1:]: |
| 188 | returncode |= check_file(path) |
| 189 | return returncode |
| 190 | |
| 191 | |
| 192 | def test_regex(): |
| 193 | def matches(rule: str, content: str) -> bool: |
| 194 | return bool(re.search(CHECK_IMPORTS[rule].pattern, content, re.MULTILINE)) |
| 195 | |
| 196 | pickle_cases = [ |
| 197 | # Should match |
| 198 | ("import pickle", True), |
| 199 | ("import cloudpickle", True), |
| 200 | ("import pickle as pkl", True), |
| 201 | ("import cloudpickle as cpkl", True), |
| 202 | ("from pickle import *", True), |
| 203 | ("from cloudpickle import dumps", True), |
| 204 | ("from pickle import dumps, loads", True), |
| 205 | ("from cloudpickle import (dumps, loads)", True), |
| 206 | (" import pickle", True), |
| 207 | ("\timport cloudpickle", True), |
| 208 | ("from pickle import loads", True), |
| 209 | # Should not match |
| 210 | ("import somethingelse", False), |
| 211 | ("from somethingelse import pickle", False), |
| 212 | ("# import pickle", False), |
| 213 | ("print('import pickle')", False), |
| 214 | ("import pickleas as asdf", False), |
| 215 | ] |
| 216 | for i, (content, should_match) in enumerate(pickle_cases): |
| 217 | result = matches("pickle/cloudpickle", content) |
| 218 | assert result == should_match, ( |
| 219 | f"pickle case {i} failed: {content!r} " |
| 220 | f"(expected {should_match}, got {result})" |
| 221 | ) |
| 222 | |
| 223 | tilelang_cases = [ |
| 224 | # Should match |
| 225 | ("import tilelang", True), |
| 226 | ("import tilelang.language as T", True), |
| 227 | ("from tilelang.jit import JITImpl", True), |
| 228 | ("from tilelang.jit.kernel import JITKernel", True), |
| 229 | # Should not match: indented (local) imports are allowed, mirroring |
| 230 | # the "triton" rule, so mocked-module test imports are not flagged. |
| 231 | (" import tilelang", False), |
| 232 | (" from tilelang.jit import JITImpl", False), |
| 233 | ("from vllm.tilelang_utils import tilelang", False), |
| 234 | ("from vllm.tilelang_utils import T", False), |
| 235 | ("from vllm.tilelang_utils import T, tilelang", False), |
| 236 | ("from vllm.tilelang_utils import tilelang, T", False), |
| 237 | ("import tilelang_kernels", False), |
| 238 | ] |
| 239 | for i, (content, should_match) in enumerate(tilelang_cases): |
| 240 | result = matches("tilelang", content) |
| 241 | assert result == should_match, ( |
| 242 | f"tilelang case {i} failed: {content!r} " |
| 243 | f"(expected {should_match}, got {result})" |
| 244 | ) |
| 245 | |
| 246 | hf_cases = [ |
| 247 | # Should match |
| 248 | ("from huggingface_hub import snapshot_download", True), |
| 249 | ("from huggingface_hub import hf_hub_download", True), |
| 250 | ("from huggingface_hub import HfApi", True), |
| 251 | ("from huggingface_hub import HfFileSystem", True), |
| 252 | ("from huggingface_hub import list_repo_files", True), |
| 253 | ("from huggingface_hub import try_to_load_from_cache", True), |
| 254 | (" from huggingface_hub import snapshot_download", True), |
| 255 | ("from huggingface_hub import PyTorchModelHubMixin, hf_hub_download", True), |
| 256 | ("from huggingface_hub import (snapshot_download)", True), |
| 257 | # Parenthesized multi-line import must not bypass the hook |
| 258 | ("from huggingface_hub import (\n snapshot_download,\n)", True), |
| 259 | ( |
| 260 | "from huggingface_hub import (\n PyTorchModelHubMixin,\n HfApi,\n)", |
| 261 | True, |
| 262 | ), |
| 263 | # Should not match |
| 264 | ("import huggingface_hub", False), |
| 265 | ("import huggingface_hub as hf", False), |
| 266 | ("from huggingface_hub import PyTorchModelHubMixin", False), |
| 267 | ("from huggingface_hub.constants import HF_HUB_CACHE", False), |
| 268 | ("from huggingface_hub.utils import EntryNotFoundError", False), |
| 269 | ("from vllm.transformers_utils.repo_utils import hf_api", False), |
| 270 | ("from huggingface_hub import (\n PyTorchModelHubMixin,\n)", False), |
| 271 | ("# from huggingface_hub import snapshot_download", False), |
| 272 | ] |
| 273 | for i, (content, should_match) in enumerate(hf_cases): |
| 274 | result = matches("huggingface_hub repo API", content) |
| 275 | assert result == should_match, ( |
| 276 | f"huggingface_hub case {i} failed: {content!r} " |
| 277 | f"(expected {should_match}, got {result})" |
| 278 | ) |
| 279 | |
| 280 | transformers_cases = [ |
| 281 | # Should match |
| 282 | ("from transformers import PretrainedConfig", True), |
| 283 | ("from transformers import AutoConfig, PretrainedConfig", True), |
| 284 | ("from transformers import PreTrainedTokenizer", True), |
| 285 | ("from transformers import PreTrainedTokenizerFast", True), |
| 286 | ("from transformers.models.siglip import SiglipImageProcessorFast", True), |
| 287 | ("from transformers import (\n PreTrainedTokenizerFast,\n)", True), |
| 288 | ("from transformers.configuration_utils import PretrainedConfig", True), |
| 289 | (" from transformers import PretrainedConfig", True), |
| 290 | ("from transformers.tokenization_utils import AddedToken", True), |
| 291 | ("from transformers.tokenization_utils_fast import X", True), |
| 292 | ("from transformers.image_processing_utils_fast import X", True), |
| 293 | ( |
| 294 | "from transformers.models.qwen2_vl.image_processing_qwen2_vl_fast import X", |
| 295 | True, |
| 296 | ), |
| 297 | ("import transformers.tokenization_utils_fast", True), |
| 298 | # Should not match |
| 299 | ("from transformers import PreTrainedConfig", False), |
| 300 | ("from transformers import PreTrainedTokenizerBase", False), |
| 301 | ("from transformers import PythonBackend, TokenizersBackend", False), |
| 302 | ("from transformers.tokenization_utils_base import BatchEncoding", False), |
| 303 | ("from transformers.tokenization_utils_tokenizers import X", False), |
| 304 | ( |
| 305 | "from transformers.image_processing_backends import TorchvisionBackend", |
| 306 | False, |
| 307 | ), |
| 308 | ("from transformers.models.qwen2_vl import Qwen2VLImageProcessor", False), |
| 309 | ("from vllm.foo import GLM4VImageProcessorFast", False), |
| 310 | ("from transformers import (\n PreTrainedConfig,\n)", False), |
| 311 | ("# from transformers import PretrainedConfig", False), |
| 312 | ] |
| 313 | for i, (content, should_match) in enumerate(transformers_cases): |
| 314 | result = matches("transformers legacy names", content) |
| 315 | assert result == should_match, ( |
| 316 | f"transformers case {i} failed: {content!r} " |
| 317 | f"(expected {should_match}, got {result})" |
| 318 | ) |
| 319 | |
| 320 | print("All regex tests passed.") |
| 321 | |
| 322 | |
| 323 | if __name__ == "__main__": |
| 324 | if "--test-regex" in sys.argv: |
| 325 | test_regex() |
| 326 | else: |
| 327 | sys.exit(main()) |