elastic_ep: Fix issues with repeated scale up/down cycles (#37131)

Signed-off-by: Itay Alroy <ialroy@nvidia.com>
Co-authored-by: Ron Tourgeman <rtourgeman@nvidia.com>
This commit is contained in:
Itay Alroy
2026-03-20 23:13:02 +00:00
committed by GitHub
co-authored by Ron Tourgeman
parent e5ed6c6c13
commit c57d38d603
10 changed files with 129 additions and 90 deletions
+8 -7
View File
@@ -602,13 +602,14 @@ class WorkerProc:
)
# 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.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
)
if envs.VLLM_ELASTIC_EP_SCALE_UP_LAUNCH:
self.worker.elastic_ep_execute("load_model")
else:
self.worker.load_model()
scheduler_config = vllm_config.scheduler_config
+4 -3
View File
@@ -382,9 +382,10 @@ class RayDistributedExecutor(Executor):
all_kwargs.append(kwargs)
self.collective_rpc("init_worker", args=(all_kwargs,))
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("init_device")
if envs.VLLM_ELASTIC_EP_SCALE_UP_LAUNCH:
self.collective_rpc("elastic_ep_execute", args=("load_model",))
else:
self.collective_rpc("load_model")
def _update_block_size(worker):
+6 -4
View File
@@ -43,12 +43,14 @@ 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])
if not is_eep_new_worker:
self.driver_worker.init_device()
self.driver_worker.init_device()
if envs.VLLM_ELASTIC_EP_SCALE_UP_LAUNCH:
self.driver_worker.elastic_ep_execute("load_model")
else:
self.driver_worker.load_model()
current_platform.update_block_size_for_backend(self.vllm_config)
current_platform.update_block_size_for_backend(self.vllm_config)
def _distributed_args(self) -> tuple[str, int, int]:
"""Return (distributed_init_method, rank, local_rank)."""