VllmServerConfigΒΆ
class VllmServerConfig ( BaseServerConfig )
vLLM Omni Docker server config.
Polymorphic Type:
type: vllmAll
BaseServerConfigtypes:
vajra: VajraServerConfig
vllm: VllmServerConfig
sglang: SglangServerConfig
Fields:
env_pathOptional [ str ] =NonePath to a Python environment directory (virtualenv/conda).
modelstr ="meta-llama/Meta-Llama-3-8B-Instruct"Model name exposed to the benchmark client.
hoststr ="localhost"Host address for the server
portint =8000Port number for the server
api_keystr ="token-abc123"API key for server authentication
gpu_idsOptional [ list [ int ] ] =NoneList of GPU IDs to use (None means auto-assign)
startup_timeoutfloat =300.0Timeout in seconds for server startup
health_check_intervalfloat =2.0Interval in seconds between health checks
require_contiguous_gpusbool =TrueRequire contiguous GPU allocation (e.g., GPUs 0,1,2 instead of 0,2,5)
tensor_parallel_sizeint =1Number of GPUs for tensor parallelism
dtypestr ="auto"Data type for model weights (auto, float16, bfloat16, etc.)
max_model_lenOptional [ int ] =NoneMaximum model context length
additional_argsOptional [ str ] ="{}"Additional engine-specific arguments as JSON string, dict, or None.
api_baseOptional [ str ] =NoneExternal API base URL for the server.
health_urlOptional [ str ] =NoneHealth endpoint URL for the server.
setup_dirOptional [ str ] =NoneSource checkout used by subprocess engines.
image : str = "vllm-omni:0.21-local"
container_name : Optional [ str ] = None
container_port : Optional [ int ] = None
engine_args : list [ str ] = []
volumes : list [ str ] = []
docker_gpus : Optional [ str ] = None
docker_runtime : Optional [ str ] = "nvidia"
ipc_mode : Optional [ str ] = "host"
hf_model : str = ""
deploy_config : str = ""
container_deploy_config : Optional [ str ] = None
env : Optional [ typing.Dict[str, str] ] = None
pass_env : list [ str ] = []
bootstrap : Optional [ str ] = None