[Elastic EP] Reuse CUDA graphs across reconfiguration (#54985)
This commit is contained in:
@@ -507,6 +507,7 @@ class EngineArgs:
|
||||
linear_backend: LinearBackend = KernelConfig.linear_backend
|
||||
all2all_backend: All2AllBackend = ParallelConfig.all2all_backend
|
||||
enable_elastic_ep: bool = ParallelConfig.enable_elastic_ep
|
||||
elastic_ep_max_dp_size: int = ParallelConfig.elastic_ep_max_dp_size
|
||||
enable_dbo: bool = ParallelConfig.enable_dbo
|
||||
ubatch_size: int = ParallelConfig.ubatch_size
|
||||
dbo_decode_token_threshold: int = ParallelConfig.dbo_decode_token_threshold
|
||||
@@ -1201,6 +1202,10 @@ class EngineArgs:
|
||||
parallel_group.add_argument(
|
||||
"--enable-elastic-ep", **parallel_kwargs["enable_elastic_ep"]
|
||||
)
|
||||
parallel_group.add_argument(
|
||||
"--elastic-ep-max-dp-size",
|
||||
**parallel_kwargs["elastic_ep_max_dp_size"],
|
||||
)
|
||||
parallel_group.add_argument(
|
||||
"--dbo-decode-token-threshold",
|
||||
**parallel_kwargs["dbo_decode_token_threshold"],
|
||||
@@ -2345,6 +2350,7 @@ class EngineArgs:
|
||||
enable_ep_weight_filter=self.enable_ep_weight_filter,
|
||||
all2all_backend=self.all2all_backend,
|
||||
enable_elastic_ep=self.enable_elastic_ep,
|
||||
elastic_ep_max_dp_size=self.elastic_ep_max_dp_size,
|
||||
enable_dbo=self.enable_dbo,
|
||||
ubatch_size=self.ubatch_size,
|
||||
dbo_decode_token_threshold=self.dbo_decode_token_threshold,
|
||||
|
||||
Reference in New Issue
Block a user