56531b79cc · tduncan/Nvidia-Optimization-Stack-VLLM

vllm/compilation/caching.py

+2 -2

@@ -1,7 +1,6 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4 -import hashlib

4 import inspect

5 import os

6 import pickle

@@ -14,6 +13,7 @@ import vllm.envs as envs

13 from vllm.config import VllmConfig, get_current_vllm_config

14 from vllm.config.utils import hash_factors

15 from vllm.logger import init_logger

16 +from vllm.utils.hashing import safe_hash

17

18 try:

19 from torch._dynamo.aot_compile import SerializableCallable

@@ -160,7 +160,7 @@ def _compute_code_hash_with_content(file_contents: dict[str, str]) -> str:

160 # e.g. exec(). We can't actually check these.

161 continue

162 hash_content.append(content)

163 - return hashlib.md5(

163 + return safe_hash(

164 "\n".join(hash_content).encode(), usedforsecurity=False

165 ).hexdigest()

166

vllm/compilation/compiler_interface.py

+7 -7

@@ -2,7 +2,6 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 import contextlib

4 import copy

5 -import hashlib

5 import os

6 from collections.abc import Callable

7 from contextlib import ExitStack

@@ -16,6 +15,7 @@ import torch.fx as fx

15 import vllm.envs as envs

16 from vllm.compilation.counter import compilation_counter

17 from vllm.config import VllmConfig

18 +from vllm.utils.hashing import safe_hash

19 from vllm.utils.torch_utils import is_torch_equal_or_newer

20

21

@@ -197,9 +197,9 @@ class InductorStandaloneAdaptor(CompilerInterface):

197

198 def compute_hash(self, vllm_config: VllmConfig) -> str:

199 factors = get_inductor_factors()

200 - hash_str = hashlib.md5(

201 - str(factors).encode(), usedforsecurity=False

202 - ).hexdigest()[:10]

200 + hash_str = safe_hash(str(factors).encode(), usedforsecurity=False).hexdigest()[

201 + :10

202 + ]

203 return hash_str

204

205 def initialize_cache(

@@ -286,9 +286,9 @@ class InductorAdaptor(CompilerInterface):

286

287 def compute_hash(self, vllm_config: VllmConfig) -> str:

288 factors = get_inductor_factors()

289 - hash_str = hashlib.md5(

290 - str(factors).encode(), usedforsecurity=False

291 - ).hexdigest()[:10]

289 + hash_str = safe_hash(str(factors).encode(), usedforsecurity=False).hexdigest()[

290 + :10

291 + ]

292 return hash_str

293

294 def initialize_cache(

vllm/config/device.py

+2 -2

@@ -1,7 +1,6 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4 -import hashlib

4 from dataclasses import field

5 from typing import Any, Literal

6

@@ -10,6 +9,7 @@ from pydantic import ConfigDict, SkipValidation

9 from pydantic.dataclasses import dataclass

10

11 from vllm.config.utils import config

12 +from vllm.utils.hashing import safe_hash

13

14 Device = Literal["auto", "cuda", "cpu", "tpu", "xpu"]

15

@@ -45,7 +45,7 @@ class DeviceConfig:

45 # the device/platform information will be summarized

46 # by torch/vllm automatically.

47 factors: list[Any] = []

48 - hash_str = hashlib.md5(str(factors).encode(), usedforsecurity=False).hexdigest()

48 + hash_str = safe_hash(str(factors).encode(), usedforsecurity=False).hexdigest()

49 return hash_str

50

51 def __post_init__(self):

vllm/config/kv_transfer.py

+2 -2

@@ -1,7 +1,6 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4 -import hashlib

4 import uuid

5 from dataclasses import field

6 from typing import Any, Literal, get_args

@@ -9,6 +8,7 @@ from typing import Any, Literal, get_args

8 from pydantic.dataclasses import dataclass

9

10 from vllm.config.utils import config

11 +from vllm.utils.hashing import safe_hash

12

13 KVProducer = Literal["kv_producer", "kv_both"]

14 KVConsumer = Literal["kv_consumer", "kv_both"]

@@ -79,7 +79,7 @@ class KVTransferConfig:

79 # no factors to consider.

80 # this config will not affect the computation graph.

81 factors: list[Any] = []

82 - hash_str = hashlib.md5(str(factors).encode(), usedforsecurity=False).hexdigest()

82 + hash_str = safe_hash(str(factors).encode(), usedforsecurity=False).hexdigest()

83 return hash_str

84

85 def __post_init__(self) -> None:

vllm/config/load.py

+2 -2

@@ -1,7 +1,6 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4 -import hashlib

4 from typing import TYPE_CHECKING, Any

5

6 from pydantic import Field, field_validator

@@ -9,6 +8,7 @@ from pydantic.dataclasses import dataclass

8

9 from vllm.config.utils import config

10 from vllm.logger import init_logger

11 +from vllm.utils.hashing import safe_hash

12

13 if TYPE_CHECKING:

14 from vllm.model_executor.model_loader import LoadFormats

@@ -104,7 +104,7 @@ class LoadConfig:

104 # no factors to consider.

105 # this config will not affect the computation graph.

106 factors: list[Any] = []

107 - hash_str = hashlib.md5(str(factors).encode(), usedforsecurity=False).hexdigest()

107 + hash_str = safe_hash(str(factors).encode(), usedforsecurity=False).hexdigest()

108 return hash_str

109

110 @field_validator("load_format", mode="after")

vllm/config/lora.py

+2 -2

@@ -1,7 +1,6 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4 -import hashlib

4 from typing import TYPE_CHECKING, Any, Literal

5

6 import torch

@@ -11,6 +10,7 @@ from typing_extensions import Self

10

11 from vllm.config.utils import config

12 from vllm.logger import init_logger

13 +from vllm.utils.hashing import safe_hash

14

15 if TYPE_CHECKING:

16 from vllm.config import ModelConfig

@@ -74,7 +74,7 @@ class LoRAConfig:

74 factors.append(self.fully_sharded_loras)

75 factors.append(self.lora_dtype)

76

77 - hash_str = hashlib.md5(str(factors).encode(), usedforsecurity=False).hexdigest()

77 + hash_str = safe_hash(str(factors).encode(), usedforsecurity=False).hexdigest()

78 return hash_str

79

80 @model_validator(mode="after")

vllm/config/multimodal.py

+2 -2

@@ -1,7 +1,6 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4 -import hashlib

4 from collections.abc import Mapping

5 from typing import TYPE_CHECKING, Any, Literal, TypeAlias

6

@@ -9,6 +8,7 @@ from pydantic import ConfigDict, Field, field_validator, model_validator

8 from pydantic.dataclasses import dataclass

9

10 from vllm.config.utils import config

11 +from vllm.utils.hashing import safe_hash

12

13 if TYPE_CHECKING:

14 from vllm.attention.backends.registry import AttentionBackendEnum

@@ -216,7 +216,7 @@ class MultiModalConfig:

216 if self.mm_encoder_attn_backend is not None

217 else None

218 ]

219 - hash_str = hashlib.md5(str(factors).encode(), usedforsecurity=False).hexdigest()

219 + hash_str = safe_hash(str(factors).encode(), usedforsecurity=False).hexdigest()

220 return hash_str

221

222 def get_limit_per_prompt(self, modality: str) -> int:

vllm/config/observability.py

+2 -2

@@ -1,7 +1,6 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4 -import hashlib

4 from functools import cached_property

5 from typing import Any, Literal, cast

6

@@ -11,6 +10,7 @@ from pydantic.dataclasses import dataclass

10

11 from vllm import version

12 from vllm.config.utils import config

13 +from vllm.utils.hashing import safe_hash

14

15 DetailedTraceModules = Literal["model", "worker", "all"]

16

@@ -78,7 +78,7 @@ class ObservabilityConfig:

78 # no factors to consider.

79 # this config will not affect the computation graph.

80 factors: list[Any] = []

81 - hash_str = hashlib.md5(str(factors).encode(), usedforsecurity=False).hexdigest()

81 + hash_str = safe_hash(str(factors).encode(), usedforsecurity=False).hexdigest()

82 return hash_str

83

84 @field_validator("show_hidden_metrics_for_version")

vllm/config/pooler.py

+2 -2

@@ -1,13 +1,13 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4 -import hashlib

4 from typing import Any

5

6 from pydantic.dataclasses import dataclass

7

8 from vllm.config.utils import config

9 from vllm.logger import init_logger

10 +from vllm.utils.hashing import safe_hash

11

12 logger = init_logger(__name__)

13

@@ -102,7 +102,7 @@ class PoolerConfig:

102 # no factors to consider.

103 # this config will not affect the computation graph.

104 factors: list[Any] = []

105 - hash_str = hashlib.md5(str(factors).encode(), usedforsecurity=False).hexdigest()

105 + hash_str = safe_hash(str(factors).encode(), usedforsecurity=False).hexdigest()

106 return hash_str

107

108

vllm/config/scheduler.py

+2 -2

@@ -1,7 +1,6 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4 -import hashlib

4 from collections.abc import Callable

5 from dataclasses import InitVar

6 from typing import TYPE_CHECKING, Any, ClassVar, Literal, cast

@@ -12,6 +11,7 @@ from typing_extensions import Self, deprecated

11

12 from vllm.config.utils import config

13 from vllm.logger import init_logger

14 +from vllm.utils.hashing import safe_hash

15 from vllm.utils.import_utils import resolve_obj_by_qualname

16

17 if TYPE_CHECKING:

@@ -178,7 +178,7 @@ class SchedulerConfig:

178 # no factors to consider.

179 # this config will not affect the computation graph.

180 factors: list[Any] = []

181 - hash_str = hashlib.md5(str(factors).encode(), usedforsecurity=False).hexdigest()

181 + hash_str = safe_hash(str(factors).encode(), usedforsecurity=False).hexdigest()

182 return hash_str

183

184 @field_validator("scheduler_cls", "async_scheduling", mode="wrap")

vllm/config/speculative.py

+2 -2

@@ -2,7 +2,6 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4 import ast

5 -import hashlib

5 from typing import TYPE_CHECKING, Any, Literal, get_args

6

7 from pydantic import Field, SkipValidation, model_validator

@@ -13,6 +12,7 @@ from vllm.config.model import ModelConfig

12 from vllm.config.parallel import ParallelConfig

13 from vllm.config.utils import config

14 from vllm.logger import init_logger

15 +from vllm.utils.hashing import safe_hash

16 from vllm.utils.import_utils import LazyLoader, has_arctic_inference

17

18 if TYPE_CHECKING:

@@ -162,7 +162,7 @@ class SpeculativeConfig:

162 # Eagle3 affects the computation graph because it returns intermediate

163 # hidden states in addition to the final hidden state.

164 factors.append(self.method == "eagle3")

165 - hash_str = hashlib.md5(str(factors).encode(), usedforsecurity=False).hexdigest()

165 + hash_str = safe_hash(str(factors).encode(), usedforsecurity=False).hexdigest()

166 return hash_str

167

168 @staticmethod

vllm/config/structured_outputs.py

+2 -2

@@ -1,7 +1,6 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4 -import hashlib

4 from typing import Any, Literal

5

6 from pydantic import model_validator

@@ -9,6 +8,7 @@ from pydantic.dataclasses import dataclass

8 from typing_extensions import Self

9

10 from vllm.config.utils import config

11 +from vllm.utils.hashing import safe_hash

12

13 StructuredOutputsBackend = Literal[

14 "auto", "xgrammar", "guidance", "outlines", "lm-format-enforcer"

@@ -58,7 +58,7 @@ class StructuredOutputsConfig:

58 # no factors to consider.

59 # this config will not affect the computation graph.

60 factors: list[Any] = []

61 - hash_str = hashlib.md5(str(factors).encode(), usedforsecurity=False).hexdigest()

61 + hash_str = safe_hash(str(factors).encode(), usedforsecurity=False).hexdigest()

62 return hash_str

63

64 @model_validator(mode="after")

vllm/config/vllm.py

+5 -5

@@ -3,7 +3,6 @@

3

4 import copy

5 import getpass

6 -import hashlib

6 import json

7 import os

8 import tempfile

@@ -25,6 +24,7 @@ from vllm.config.speculative import EagleModelTypes

24 from vllm.logger import enable_trace_function_call, init_logger

25 from vllm.transformers_utils.runai_utils import is_runai_obj_uri

26 from vllm.utils import random_uuid

27 +from vllm.utils.hashing import safe_hash

28

29 from .cache import CacheConfig

30 from .compilation import CompilationConfig, CompilationMode, CUDAGraphMode

@@ -193,7 +193,7 @@ class VllmConfig:

193 vllm_factors.append("None")

194 if self.additional_config:

195 if isinstance(additional_config := self.additional_config, dict):

196 - additional_config_hash = hashlib.md5(

196 + additional_config_hash = safe_hash(

197 json.dumps(additional_config, sort_keys=True).encode(),

198 usedforsecurity=False,

199 ).hexdigest()

@@ -204,9 +204,9 @@ class VllmConfig:

204 vllm_factors.append("None")

205 factors.append(vllm_factors)

206

207 - hash_str = hashlib.md5(

208 - str(factors).encode(), usedforsecurity=False

209 - ).hexdigest()[:10]

207 + hash_str = safe_hash(str(factors).encode(), usedforsecurity=False).hexdigest()[

208 + :10

209 + ]

210 return hash_str

211

212 def pad_for_cudagraph(self, batch_size: int) -> int:

vllm/distributed/kv_transfer/kv_connector/v1/shared_storage_connector.py

+2 -2

@@ -1,6 +1,5 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 -import hashlib

3 import os

4 from dataclasses import dataclass, field

5 from typing import TYPE_CHECKING, Any, Optional

@@ -15,6 +14,7 @@ from vllm.distributed.kv_transfer.kv_connector.v1.base import (

14 KVConnectorRole,

15 )

16 from vllm.logger import init_logger

17 +from vllm.utils.hashing import safe_hash

18 from vllm.v1.attention.backends.mla.common import MLACommonMetadata

19 from vllm.v1.core.sched.output import SchedulerOutput

20

@@ -423,7 +423,7 @@ class SharedStorageConnector(KVConnectorBase_V1):

423 if mm_hashes:

424 mm_str = "-".join(mm_hashes)

425 token_bytes += mm_str.encode("utf-8")

426 - input_ids_hash = hashlib.md5(token_bytes, usedforsecurity=False).hexdigest()

426 + input_ids_hash = safe_hash(token_bytes, usedforsecurity=False).hexdigest()

427

428 foldername = os.path.join(self._storage_path, input_ids_hash)

429 if create_folder:

vllm/model_executor/models/registry.py

+2 -2

@@ -5,7 +5,6 @@ Whenever you add an architecture to this page, please also update

5 `tests/models/registry.py` with example HuggingFace models for it.

6 """

7

8 -import hashlib

8 import importlib

9 import json

10 import os

@@ -32,6 +31,7 @@ from vllm.config import (

31 from vllm.logger import init_logger

32 from vllm.logging_utils import logtime

33 from vllm.transformers_utils.dynamic_module import try_get_class_from_dynamic_module

34 +from vllm.utils.hashing import safe_hash

35

36 from .interfaces import (

37 has_inner_state,

@@ -654,7 +654,7 @@ class _LazyRegisteredModel(_BaseRegisteredModel):

654

655 if model_path.exists():

656 with open(model_path, "rb") as f:

657 - module_hash = hashlib.md5(f.read(), usedforsecurity=False).hexdigest()

657 + module_hash = safe_hash(f.read(), usedforsecurity=False).hexdigest()

658

659 mi = self._load_modelinfo_from_cache(module_hash)

660 if mi is not None:

vllm/utils/hashing.py

+18

@@ -5,6 +5,7 @@ from __future__ import annotations

5

6 import hashlib

7 import pickle

8 +from _hashlib import HASH, UnsupportedDigestmodError

9 from collections.abc import Callable

10 from typing import Any

11

@@ -61,3 +62,20 @@ def get_hash_fn_by_name(hash_fn_name: str) -> Callable[[Any], bytes]:

62 return sha256_cbor

63

64 raise ValueError(f"Unsupported hash function: {hash_fn_name}")

65 +

66 +

67 +def safe_hash(data: bytes, usedforsecurity: bool = True) -> HASH:

68 + """Hash for configs, defaulting to md5 but falling back to sha256

69 + in FIPS constrained environments.

70 +

71 + Args:

72 + data: bytes

73 + usedforsecurity: Whether the hash is used for security purposes

74 +

75 + Returns:

76 + Hash object

77 + """

78 + try:

79 + return hashlib.md5(data, usedforsecurity=usedforsecurity)

80 + except (UnsupportedDigestmodError, ValueError):

81 + return hashlib.sha256(data)

[Misc] Add backup hash algorithm for FIPS constrained environments (#28795)