chore: move ray utils to executor dir

aphrodite/__init__.py

@@ -1,7 +1,7 @@
 from aphrodite.engine.args_tools import AsyncEngineArgs, EngineArgs
 from aphrodite.engine.async_aphrodite import AsyncAphrodite
 from aphrodite.engine.aphrodite_engine import AphroditeEngine
-from aphrodite.engine.ray_tools import initialize_ray_cluster
+from aphrodite.executor.ray_utils import initialize_ray_cluster
 from aphrodite.endpoints.llm import LLM
 from aphrodite.modeling.models import ModelRegistry
 from aphrodite.common.outputs import CompletionOutput, RequestOutput

aphrodite/engine/aphrodite_engine.py

@@ -22,7 +22,7 @@ from aphrodite.engine.output_processor.interfaces import \
 from aphrodite.engine.output_processor.stop_checker import StopChecker
 from aphrodite.engine.output_processor.util import \
     create_output_by_sequence_group
-from aphrodite.engine.ray_tools import initialize_ray_cluster
+from aphrodite.executor.ray_utils import initialize_ray_cluster
 from aphrodite.executor.executor_base import ExecutorBase
 from aphrodite.lora.request import LoRARequest
 from aphrodite.processing.scheduler import Scheduler, SchedulerOutputs

aphrodite/engine/async_aphrodite.py

@@ -14,7 +14,7 @@ from aphrodite.common.sampling_params import SamplingParams
 from aphrodite.common.sequence import MultiModalData
 from aphrodite.engine.args_tools import AsyncEngineArgs
 from aphrodite.engine.aphrodite_engine import AphroditeEngine
-from aphrodite.engine.ray_tools import initialize_ray_cluster, ray
+from aphrodite.executor.ray_utils import initialize_ray_cluster, ray
 from aphrodite.lora.request import LoRARequest
 
 ENGINE_ITERATION_TIMEOUT_S = int(

aphrodite/executor/ray_gpu_executor.py

@@ -11,7 +11,7 @@ from aphrodite.common.sequence import SamplerOutput, SequenceGroupMetadata
 from aphrodite.common.utils import (get_aphrodite_instance_id,
                                     get_distributed_init_method, get_ip,
                                     get_open_port, make_async)
-from aphrodite.engine.ray_tools import RayWorkerWrapper, ray
+from aphrodite.executor.ray_utils import RayWorkerWrapper, ray
 from aphrodite.executor.executor_base import ExecutorAsyncBase, ExecutorBase
 from aphrodite.lora.request import LoRARequest
 

aphrodite/transformers_utils/tokenizer_group/__init__.py

@@ -5,7 +5,7 @@ from aphrodite.transformers_utils.tokenizer_group.base_tokenizer_group import (
     BaseTokenizerGroup)
 from aphrodite.transformers_utils.tokenizer_group.tokenizer_group import (
     TokenizerGroup)
-from aphrodite.engine.ray_tools import ray
+from aphrodite.executor.ray_utils import ray
 
 if ray:
     from aphrodite.transformers_utils.tokenizer_group.ray_tokenizer_group import (  # noqa: E501

aphrodite/transformers_utils/tokenizer_group/ray_tokenizer_group.py

@@ -6,7 +6,7 @@ from transformers import PreTrainedTokenizer
 
 from aphrodite.common.config import TokenizerPoolConfig
 from aphrodite.lora.request import LoRARequest
-from aphrodite.engine.ray_tools import ray
+from aphrodite.executor.ray_utils import ray
 from aphrodite.transformers_utils.tokenizer_group.base_tokenizer_group import (
     BaseTokenizerGroup)
 from aphrodite.transformers_utils.tokenizer_group.tokenizer_group import (