[1/N] Elastic EP Milestone 2 (#34861)

Signed-off-by: Yongji Wu <wuyongji317@gmail.com>
Signed-off-by: Itay Alroy <ialroy@nvidia.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Signed-off-by: Ron Tourgeman <rtourgeman@nvidia.com>
Co-authored-by: Yongji Wu <wuyongji317@gmail.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Ron Tourgeman <rtourgeman@nvidia.com>
This commit is contained in:
Itay Alroy
2026-02-28 04:46:42 +00:00
committed by GitHub
co-authored by Yongji Wu Tyler Michael Smith Ron Tourgeman
parent 90805ff464
commit dea268336f
53 changed files with 3612 additions and 1015 deletions
+16 -5
View File
@@ -38,6 +38,7 @@ from vllm.distributed.parallel_state import (
get_pcp_group,
get_pp_group,
get_tp_group,
model_parallel_is_initialized,
)
from vllm.envs import enable_envs_cache
from vllm.logger import init_logger
@@ -580,17 +581,20 @@ class WorkerProc:
)
self.async_output_copy_thread.start()
# Initialize device
self.worker.init_device()
# Set process title and log prefix
self.setup_proc_title_and_log_prefix(
enable_ep=vllm_config.parallel_config.enable_expert_parallel
)
# Load model
self._init_message_queues(input_shm_handle, vllm_config)
self.worker.load_model()
is_eep_new_worker = envs.VLLM_ELASTIC_EP_SCALE_UP_LAUNCH
if not is_eep_new_worker:
self.worker.init_device()
# Update process title now that parallel groups are initialized
self.setup_proc_title_and_log_prefix(
enable_ep=vllm_config.parallel_config.enable_expert_parallel
)
self.worker.load_model()
# Enable environment variable cache (e.g. assume no more
# environment variable overrides after this point)
@@ -885,6 +889,13 @@ class WorkerProc:
@staticmethod
def setup_proc_title_and_log_prefix(enable_ep: bool) -> None:
# Check if parallel groups are initialized first
if not model_parallel_is_initialized():
# Parallel groups not yet initialized, use default process name
set_process_title(name="Worker")
decorate_logs("Worker")
return
dp_size = get_dp_group().world_size
dp_rank = get_dp_group().rank_in_group
pp_size = get_pp_group().world_size
+4 -2
View File
@@ -382,8 +382,10 @@ class RayDistributedExecutor(Executor):
all_kwargs.append(kwargs)
self.collective_rpc("init_worker", args=(all_kwargs,))
self.collective_rpc("init_device")
self.collective_rpc("load_model")
is_eep_new_worker = envs.VLLM_ELASTIC_EP_SCALE_UP_LAUNCH
if not is_eep_new_worker:
self.collective_rpc("init_device")
self.collective_rpc("load_model")
for pp_rank in range(self.parallel_config.pipeline_parallel_size):
self.pp_tp_workers.append([])
+4 -13
View File
@@ -14,7 +14,6 @@ import vllm.envs as envs
from vllm.logger import init_logger
from vllm.utils.network_utils import get_distributed_init_method, get_ip, get_open_port
from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
from vllm.v1.engine import ReconfigureDistributedRequest, ReconfigureRankType
from vllm.v1.executor.abstract import Executor
from vllm.v1.outputs import AsyncModelRunnerOutput, DraftTokenIds, ModelRunnerOutput
from vllm.v1.serial_utils import run_method
@@ -43,9 +42,11 @@ class UniProcExecutor(Executor):
max_workers=1, thread_name_prefix="WorkerAsyncOutput"
)
is_eep_new_worker = envs.VLLM_ELASTIC_EP_SCALE_UP_LAUNCH
self.driver_worker.init_worker(all_kwargs=[kwargs])
self.driver_worker.init_device()
self.driver_worker.load_model()
if not is_eep_new_worker:
self.driver_worker.init_device()
self.driver_worker.load_model()
def _distributed_args(self) -> tuple[str, int, int]:
"""Return (distributed_init_method, rank, local_rank)."""
@@ -122,16 +123,6 @@ class UniProcExecutor(Executor):
# it's running.
return
def reinitialize_distributed(
self, reconfig_request: ReconfigureDistributedRequest
) -> None:
self.driver_worker.reinitialize_distributed(reconfig_request)
if (
reconfig_request.new_data_parallel_rank
== ReconfigureRankType.SHUTDOWN_CURRENT_RANK
):
self.shutdown()
def shutdown(self) -> None:
if worker := self.driver_worker:
worker.shutdown()