biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[core] platform agnostic executor via collective_rpc (#11256)

Browse Source 
Signed-off-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
youkaichao
2025-01-15 13:45:21 +08:00
committed by GitHub
parent f218f9c24d
commit ad34c0df0f
43 changed files with 851 additions and 2641 deletions
Download Patch File Download Diff File Expand all files Collapse all files

94
vllm/engine/llm_engine.py
View File

@@ -28,8 +28,6 @@ from vllm.engine.output_processor.util import create_output_by_sequence_group
from vllm.entrypoints.openai.logits_processors import (
get_logits_processors as get_openai_logits_processors)
from vllm.executor.executor_base import ExecutorBase
from vllm.executor.gpu_executor import GPUExecutor
from vllm.executor.ray_utils import initialize_ray_cluster
from vllm.inputs import (INPUT_REGISTRY, InputRegistry, ProcessorInputs,
PromptType, SingletonInputsAdapter)
from vllm.inputs.parse import is_encoder_decoder_inputs, is_token_prompt
@@ -442,64 +440,26 @@ class LLMEngine:
raise TypeError(
"distributed_executor_backend must be a subclass of "
f"ExecutorBase. Got {distributed_executor_backend}.")
if distributed_executor_backend.uses_ray: # type: ignore
initialize_ray_cluster(engine_config.parallel_config)
executor_class = distributed_executor_backend
elif engine_config.device_config.device_type == "neuron":
from vllm.executor.neuron_executor import NeuronExecutor
executor_class = NeuronExecutor
elif engine_config.device_config.device_type == "tpu":
elif engine_config.parallel_config.world_size > 1:
if distributed_executor_backend == "ray":
initialize_ray_cluster(engine_config.parallel_config)
from vllm.executor.ray_tpu_executor import RayTPUExecutor
executor_class = RayTPUExecutor
else:
assert distributed_executor_backend is None
from vllm.executor.tpu_executor import TPUExecutor
executor_class = TPUExecutor
elif engine_config.device_config.device_type == "cpu":
from vllm.executor.cpu_executor import CPUExecutor
executor_class = CPUExecutor
elif engine_config.device_config.device_type == "hpu":
if distributed_executor_backend == "ray":
initialize_ray_cluster(engine_config.parallel_config)
from vllm.executor.ray_hpu_executor import RayHPUExecutor
executor_class = RayHPUExecutor
else:
from vllm.executor.hpu_executor import HPUExecutor
executor_class = HPUExecutor
elif engine_config.device_config.device_type == "openvino":
from vllm.executor.openvino_executor import OpenVINOExecutor
executor_class = OpenVINOExecutor
elif engine_config.device_config.device_type == "xpu":
if distributed_executor_backend == "ray":
initialize_ray_cluster(engine_config.parallel_config)
from vllm.executor.ray_xpu_executor import RayXPUExecutor
executor_class = RayXPUExecutor
from vllm.executor.ray_distributed_executor import (
RayDistributedExecutor)
executor_class = RayDistributedExecutor
elif distributed_executor_backend == "mp":
# FIXME(kunshang):
# spawn needs calling `if __name__ == '__main__':``
# fork is not supported for xpu start new process.
logger.error(
"Both start methods (spawn and fork) have issue "
"on XPU if you use mp backend, Please try ray instead.")
else:
from vllm.executor.xpu_executor import XPUExecutor
executor_class = XPUExecutor
elif distributed_executor_backend == "ray":
initialize_ray_cluster(engine_config.parallel_config)
from vllm.executor.ray_gpu_executor import RayGPUExecutor
executor_class = RayGPUExecutor
elif distributed_executor_backend == "mp":
from vllm.executor.multiproc_gpu_executor import (
MultiprocessingGPUExecutor)
assert not envs.VLLM_USE_RAY_SPMD_WORKER, (
"multiprocessing distributed executor backend does not "
"support VLLM_USE_RAY_SPMD_WORKER=1")
executor_class = MultiprocessingGPUExecutor
from vllm.executor.mp_distributed_executor import (
MultiprocessingDistributedExecutor)
assert not envs.VLLM_USE_RAY_SPMD_WORKER, (
"multiprocessing distributed executor backend does not "
"support VLLM_USE_RAY_SPMD_WORKER=1")
executor_class = MultiprocessingDistributedExecutor
elif distributed_executor_backend == "uni":
# JAX-style, single-process, multi-device executor.
from vllm.executor.uniproc_executor import UniProcExecutor
executor_class = UniProcExecutor
else:
from vllm.executor.gpu_executor import GPUExecutor
executor_class = GPUExecutor
from vllm.executor.uniproc_executor import UniProcExecutor
executor_class = UniProcExecutor
return executor_class
@classmethod
@@ -1845,27 +1805,17 @@ class LLMEngine:
def list_prompt_adapters(self) -> List[int]:
return self.model_executor.list_prompt_adapters()
def start_profile(self) -> None:
self.model_executor.start_profile()
def stop_profile(self) -> None:
self.model_executor.stop_profile()
def check_health(self) -> None:
if self.tokenizer:
self.tokenizer.check_health()
self.model_executor.check_health()
def start_profile(self) -> None:
# using type instead of isinstance to check to avoid capturing
# inherited classes (MultiprocessingGPUExecutor)
if type(self.model_executor) == GPUExecutor: # noqa: E721
self.model_executor.start_profile()
else:
self.model_executor._run_workers("start_profile")
def stop_profile(self) -> None:
# using type instead of isinstance to check to avoid capturing
# inherited classes (MultiprocessingGPUExecutor)
if type(self.model_executor) == GPUExecutor: # noqa: E721
self.model_executor.stop_profile()
else:
self.model_executor._run_workers("stop_profile")
def is_tracing_enabled(self) -> bool:
return self.tracer is not None