30 lines
862 B
Python
30 lines
862 B
Python
"""vLLM: a high-throughput and memory-efficient inference engine for LLMs"""
|
|
import os
|
|
|
|
# By default, to avoid memory fragmentation, disable UMD mempool
|
|
if os.getenv("UMD_ENABLEMEMPOOL") is None:
|
|
os.environ["UMD_ENABLEMEMPOOL"] = "0"
|
|
os.environ["NCCL_FORCESYNC_DISABLE"] = "1"
|
|
|
|
from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
|
|
from vllm.engine.async_llm_engine import AsyncLLMEngine
|
|
from vllm.engine.llm_engine import LLMEngine
|
|
from vllm.engine.ray_utils import initialize_cluster
|
|
from vllm.entrypoints.llm import LLM
|
|
from vllm.outputs import CompletionOutput, RequestOutput
|
|
from vllm.sampling_params import SamplingParams
|
|
|
|
__version__ = "0.3.3"
|
|
|
|
__all__ = [
|
|
"LLM",
|
|
"SamplingParams",
|
|
"RequestOutput",
|
|
"CompletionOutput",
|
|
"LLMEngine",
|
|
"EngineArgs",
|
|
"AsyncLLMEngine",
|
|
"AsyncEngineArgs",
|
|
"initialize_cluster",
|
|
]
|