@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 7a6d45bc8a

Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 15, 17:15 UTC

Support FIPS enabled machines with MD5 hashing (#15299)

Signed-off-by: Matthew Vine <32849887+MattTheCuber@users.noreply.github.com>

Matthew Vine committed Mar 26, 2025 at 20:19 UTC 7a6d45bc8a201623c646627becd837afd6b35bc7

4 files changed +36 -19

tests/compile/piecewise/test_toy_llama.py

+2 -1

@@ -63,7 +63,8 @@ class LlamaConfig:

63 factors.append((k, v))

64 factors.sort()

65 import hashlib

66 - return hashlib.md5(str(factors).encode()).hexdigest()

66 + return hashlib.md5(str(factors).encode(),

67 + usedforsecurity=False).hexdigest()

69 def __post_init__(self):

70 assert self.mlp_size >= self.hidden_size

vllm/compilation/backends.py

+4 -3

@@ -381,8 +381,8 @@ class VllmBackend:

381 with open(filepath) as f:

382 hash_content.append(f.read())

383 import hashlib

384 - code_hash = hashlib.md5(

385 - "\n".join(hash_content).encode()).hexdigest()

384 + code_hash = hashlib.md5("\n".join(hash_content).encode(),

385 + usedforsecurity=False).hexdigest()

386 factors.append(code_hash)

387

388 # 3. compiler hash

@@ -390,7 +390,8 @@ class VllmBackend:

390 factors.append(compiler_hash)

391

392 # combine all factors to generate the cache dir

393 - hash_key = hashlib.md5(str(factors).encode()).hexdigest()[:10]

393 + hash_key = hashlib.md5(str(factors).encode(),

394 + usedforsecurity=False).hexdigest()[:10]

395

396 cache_dir = os.path.join(

397 envs.VLLM_CACHE_ROOT,

vllm/compilation/compiler_interface.py

+2 -1

@@ -139,7 +139,8 @@ class InductorAdaptor(CompilerInterface):

139 from torch._inductor.codecache import torch_key

140 torch_factors = torch_key()

141 factors.append(torch_factors)

142 - hash_str = hashlib.md5(str(factors).encode()).hexdigest()[:10]

142 + hash_str = hashlib.md5(str(factors).encode(),

143 + usedforsecurity=False).hexdigest()[:10]

144 return hash_str

145

146 def initialize_cache(self, cache_dir: str, disable_cache: bool = False):

vllm/config.py

+28 -14

@@ -1111,7 +1111,8 @@ class CacheConfig:

1111 factors: list[Any] = []

1112 factors.append(self.cache_dtype)

1113 # `cpu_offload_gb` does not use `torch.compile` yet.

1114 - hash_str = hashlib.md5(str(factors).encode()).hexdigest()

1114 + hash_str = hashlib.md5(str(factors).encode(),

1115 + usedforsecurity=False).hexdigest()

1116 return hash_str

1117

1118 def __init__(

@@ -1243,7 +1244,8 @@ class TokenizerPoolConfig:

1244 # no factors to consider.

1245 # this config will not affect the computation graph.

1246 factors: list[Any] = []

1246 - hash_str = hashlib.md5(str(factors).encode()).hexdigest()

1247 + hash_str = hashlib.md5(str(factors).encode(),

1248 + usedforsecurity=False).hexdigest()

1249 return hash_str

1250

1251 def __post_init__(self):

@@ -1354,7 +1356,8 @@ class LoadConfig:

1356 # no factors to consider.

1357 # this config will not affect the computation graph.

1358 factors: list[Any] = []

1357 - hash_str = hashlib.md5(str(factors).encode()).hexdigest()

1359 + hash_str = hashlib.md5(str(factors).encode(),

1360 + usedforsecurity=False).hexdigest()

1361 return hash_str

1362

1363 def __post_init__(self):

@@ -1674,7 +1677,8 @@ class SchedulerConfig:

1677 # no factors to consider.

1678 # this config will not affect the computation graph.

1679 factors: list[Any] = []

1677 - hash_str = hashlib.md5(str(factors).encode()).hexdigest()

1680 + hash_str = hashlib.md5(str(factors).encode(),

1681 + usedforsecurity=False).hexdigest()

1682 return hash_str

1683

1684 def __post_init__(self) -> None:

@@ -1810,7 +1814,8 @@ class DeviceConfig:

1814 # the device/platform information will be summarized

1815 # by torch/vllm automatically.

1816 factors: list[Any] = []

1813 - hash_str = hashlib.md5(str(factors).encode()).hexdigest()

1817 + hash_str = hashlib.md5(str(factors).encode(),

1818 + usedforsecurity=False).hexdigest()

1819 return hash_str

1820

1821 def __init__(self, device: str = "auto") -> None:

@@ -1983,7 +1988,8 @@ class SpeculativeConfig:

1988 # no factors to consider.

1989 # spec decode does not use `torch.compile` yet.

1990 factors: list[Any] = []

1986 - hash_str = hashlib.md5(str(factors).encode()).hexdigest()

1991 + hash_str = hashlib.md5(str(factors).encode(),

1992 + usedforsecurity=False).hexdigest()

1993 return hash_str

1994

1995 @classmethod

@@ -2358,7 +2364,8 @@ class LoRAConfig:

2364 factors.append(self.lora_extra_vocab_size)

2365 factors.append(self.long_lora_scaling_factors)

2366 factors.append(self.bias_enabled)

2361 - hash_str = hashlib.md5(str(factors).encode()).hexdigest()

2367 + hash_str = hashlib.md5(str(factors).encode(),

2368 + usedforsecurity=False).hexdigest()

2369 return hash_str

2370

2371 def __post_init__(self):

@@ -2424,7 +2431,8 @@ class PromptAdapterConfig:

2431 # no factors to consider.

2432 # this config will not affect the computation graph.

2433 factors: list[Any] = []

2427 - hash_str = hashlib.md5(str(factors).encode()).hexdigest()

2434 + hash_str = hashlib.md5(str(factors).encode(),

2435 + usedforsecurity=False).hexdigest()

2436 return hash_str

2437

2438 def __post_init__(self):

@@ -2469,7 +2477,8 @@ class MultiModalConfig:

2477 # no factors to consider.

2478 # this config will not affect the computation graph.

2479 factors: list[Any] = []

2472 - hash_str = hashlib.md5(str(factors).encode()).hexdigest()

2480 + hash_str = hashlib.md5(str(factors).encode(),

2481 + usedforsecurity=False).hexdigest()

2482 return hash_str

2483

2484 def get_limit_per_prompt(self, modality: str) -> int:

@@ -2535,7 +2544,8 @@ class PoolerConfig:

2544 # no factors to consider.

2545 # this config will not affect the computation graph.

2546 factors: list[Any] = []

2538 - hash_str = hashlib.md5(str(factors).encode()).hexdigest()

2547 + hash_str = hashlib.md5(str(factors).encode(),

2548 + usedforsecurity=False).hexdigest()

2549 return hash_str

2550

2551 @staticmethod

@@ -2816,7 +2826,8 @@ class DecodingConfig:

2826 # no factors to consider.

2827 # this config will not affect the computation graph.

2828 factors: list[Any] = []

2819 - hash_str = hashlib.md5(str(factors).encode()).hexdigest()

2829 + hash_str = hashlib.md5(str(factors).encode(),

2830 + usedforsecurity=False).hexdigest()

2831 return hash_str

2832

2833 def __post_init__(self):

@@ -2866,7 +2877,8 @@ class ObservabilityConfig:

2877 # no factors to consider.

2878 # this config will not affect the computation graph.

2879 factors: list[Any] = []

2869 - hash_str = hashlib.md5(str(factors).encode()).hexdigest()

2880 + hash_str = hashlib.md5(str(factors).encode(),

2881 + usedforsecurity=False).hexdigest()

2882 return hash_str

2883

2884 def __post_init__(self):

@@ -2928,7 +2940,8 @@ class KVTransferConfig(BaseModel):

2940 # no factors to consider.

2941 # this config will not affect the computation graph.

2942 factors: list[Any] = []

2931 - hash_str = hashlib.md5(str(factors).encode()).hexdigest()

2943 + hash_str = hashlib.md5(str(factors).encode(),

2944 + usedforsecurity=False).hexdigest()

2945 return hash_str

2946

2947 @classmethod

@@ -3425,7 +3438,8 @@ class VllmConfig:

3438 vllm_factors.append("None")

3439 factors.append(vllm_factors)

3440

3428 - hash_str = hashlib.md5(str(factors).encode()).hexdigest()[:10]

3441 + hash_str = hashlib.md5(str(factors).encode(),

3442 + usedforsecurity=False).hexdigest()[:10]

3443 return hash_str

3444

3445 def pad_for_cudagraph(self, batch_size: int) -> int: