[V1] Support DP with Ray (#18779)

This commit is contained in:
Rui Qiao
2025-06-02 21:15:13 -07:00
committed by GitHub
parent 9e6f61e8c3
commit bdce64f236
10 changed files with 551 additions and 120 deletions
+9 -4
View File
@@ -27,7 +27,8 @@ from vllm.transformers_utils.tokenizer_group import init_tokenizer_from_configs
from vllm.usage.usage_lib import UsageContext
from vllm.utils import Device, cdiv
from vllm.v1.engine import EngineCoreRequest
from vllm.v1.engine.core_client import AsyncMPClient, DPAsyncMPClient
from vllm.v1.engine.core_client import (AsyncMPClient, DPAsyncMPClient,
RayDPClient)
from vllm.v1.engine.exceptions import EngineDeadError, EngineGenerateError
from vllm.v1.engine.output_processor import (OutputProcessor,
RequestOutputCollector)
@@ -119,9 +120,13 @@ class AsyncLLM(EngineClient):
log_stats=self.log_stats)
# EngineCore (starts the engine in background process).
core_client_class = AsyncMPClient if (
vllm_config.parallel_config.data_parallel_size
== 1) else DPAsyncMPClient
core_client_class: type[AsyncMPClient]
if vllm_config.parallel_config.data_parallel_size == 1:
core_client_class = AsyncMPClient
elif vllm_config.parallel_config.data_parallel_backend == "ray":
core_client_class = RayDPClient
else:
core_client_class = DPAsyncMPClient
self.engine_core = core_client_class(
vllm_config=vllm_config,