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, 23:30 UTC
Simplify logic of locating CUDART so file path (#13203)
Signed-off-by: Lu Fang <lufang@fb.com>
Lu Fang committed
Feb 12, 2025 at 21:52 UTC
009439caeb3ae27d1d6c94e550eee13bbd0520af
1 file changed
+1
-25
vllm/distributed/device_communicators/cuda_wrapper.py
+1
-25
@@ -5,7 +5,6 @@ convenient for use when we just need to call a few functions.
5
"""
6
7
import ctypes
8
-import glob
8
from dataclasses import dataclass
9
from typing import Any, Dict, List, Optional
10
@@ -62,29 +61,6 @@ def find_loaded_library(lib_name) -> Optional[str]:
61
return path
62
63
65
-def get_cudart_lib_path_from_env() -> Optional[str]:
66
- """
67
- In some system, find_loaded_library() may not work. So we allow users to
68
- specify the path through environment variable VLLM_CUDART_SO_PATH.
69
- """
70
- cudart_so_env = envs.VLLM_CUDART_SO_PATH
71
- if cudart_so_env is not None:
72
- cudart_paths = [
73
- cudart_so_env,
74
- ]
75
- for path in cudart_paths:
76
- file_paths = glob.glob(path)
77
- if len(file_paths) > 0:
78
- logger.info(
79
- "Found cudart library at %s through env var"
80
- "VLLM_CUDART_SO_PATH=%s",
81
- file_paths[0],
82
- cudart_so_env,
83
- )
84
- return file_paths[0]
85
- return None
86
-
87
-
64
class CudaRTLibrary:
65
exported_functions = [
66
# cudaError_t cudaSetDevice ( int device )
@@ -131,7 +107,7 @@ class CudaRTLibrary:
107
if so_file is None:
108
so_file = find_loaded_library("libcudart")
109
if so_file is None:
134
- so_file = get_cudart_lib_path_from_env()
110
+ so_file = envs.VLLM_CUDART_SO_PATH # fallback to env var
111
assert so_file is not None, \
112
(
113
"libcudart is not loaded in the current process, "