[Core] Use a faster hash function (LMCache#1020)

* perf: replace sha256 by xxhash

Signed-off-by: Zhou Fang <[email protected]>

* build: add xxhash to requirements

Signed-off-by: Zhou Fang <[email protected]>

* style: import

Signed-off-by: Zhou Fang <[email protected]>

* build: add xxhash to common requirements

Signed-off-by: Zhou Fang <[email protected]>

---------

Signed-off-by: Zhou Fang <[email protected]>

Author: zhouwfang

lmcache/cache_engine.py

@@ -14,12 +14,12 @@
 
 # Standard
 from typing import Dict, Iterable, List, Optional, Tuple, Union
-import hashlib
 import logging
 import time
 
 # Third Party
 import torch
+import xxhash
 
 # First Party
 from lmcache.config import LMCacheEngineConfig, LMCacheEngineMetadata
@@ -85,8 +85,7 @@ def _hash(
         tokens: torch.Tensor,
         prefix_hash: str,
     ) -> str:
-        # TODO: change it to a more efficient hash function
-        hasher = hashlib.sha256()
+        hasher = xxhash.xxh64()
         hasher.update(prefix_hash.encode("ascii"))
         hasher.update(tokens.numpy().tobytes())
         return hasher.hexdigest()

pyproject.toml

@@ -7,6 +7,7 @@ requires = [
     "setuptools_scm>=8",
     "torch==2.7.0",
     "wheel",
+    "xxhash==3.5.0",
 ]
 build-backend = "setuptools.build_meta"
 

requirements/build.txt

@@ -6,3 +6,4 @@ setuptools>=77.0.3,<81.0.0
 setuptools_scm>=8
 torch==2.7.0 # Corresponds to the version used by vLLM main branch
 wheel
+xxhash==3.5.0

requirements/common.txt

@@ -17,3 +17,4 @@ setuptools_scm>=8
 sortedcontainers
 torch==2.7.0  # Should correspond to the version used by vLLM main branch
 transformers >= 4.51.1
+xxhash==3.5.0