[Elastic EP] Reuse CUDA graphs across reconfiguration (#54985)

This commit is contained in:
Itay Alroy
2026-09-12 10:26:39 -04:00
committed by GitHub
parent 13e221f830
commit 658c8131c7
22 changed files with 372 additions and 270 deletions
+17 -1
View File
@@ -216,6 +216,8 @@ class ParallelConfig:
enable_elastic_ep: bool = False
"""Enable elastic expert parallelism with stateless NCCL groups for DP/EP."""
elastic_ep_max_dp_size: int = Field(default=None, ge=1) # type: ignore[assignment]
"""Maximum data parallel size supported by elastic expert parallelism."""
enable_dbo: bool = False
"""Enable dual batch overlap for the model executor."""
@@ -435,7 +437,9 @@ class ParallelConfig:
)
"""The configurations for fault tolerance."""
@field_validator("disable_nccl_for_dp_synchronization", mode="wrap")
@field_validator(
"disable_nccl_for_dp_synchronization", "elastic_ep_max_dp_size", mode="wrap"
)
@classmethod
def _skip_none_validation(cls, value: Any, handler: Callable) -> Any:
"""Skip validation if the value is `None` when initialisation is delayed."""
@@ -941,6 +945,18 @@ class ParallelConfig:
" for dense models."
)
max_dp_size = self.elastic_ep_max_dp_size
self.elastic_ep_max_dp_size = (
max_dp_size
if self.enable_elastic_ep and max_dp_size is not None
else self.data_parallel_size
)
if self.elastic_ep_max_dp_size < self.data_parallel_size:
raise ValueError(
"--elastic-ep-max-dp-size must be greater than or equal to "
f"the initial data_parallel_size ({self.data_parallel_size})."
)
self.data_parallel_index = self.data_parallel_rank
if self.distributed_executor_backend == "external_launcher":