[Model Runner V2] batch-sharded sample (#50465)

Signed-off-by: Giancarlo Delfin <gdelfin@inferact.ai>
This commit is contained in:
Giancarlo Delfin
2026-08-24 15:49:57 -07:00
committed by GitHub
parent 0d7d5ed0b2
commit d154d90d6c
19 changed files with 1781 additions and 22 deletions
+8
View File
@@ -496,6 +496,9 @@ class EngineArgs:
data_parallel_multi_port_external_lb: bool = False
data_parallel_backend: DataParallelBackend = ParallelConfig.data_parallel_backend
enable_expert_parallel: bool = ParallelConfig.enable_expert_parallel
enable_batch_sharded_sampling: bool | None = (
ParallelConfig.enable_batch_sharded_sampling
)
enable_ep_weight_filter: bool = ParallelConfig.enable_ep_weight_filter
moe_backend: MoEBackend = KernelConfig.moe_backend
linear_backend: LinearBackend = KernelConfig.linear_backend
@@ -1156,6 +1159,10 @@ class EngineArgs:
"-ep",
**parallel_kwargs["enable_expert_parallel"],
)
parallel_group.add_argument(
"--enable-batch-sharded-sampling",
**parallel_kwargs["enable_batch_sharded_sampling"],
)
parallel_group.add_argument(
"--enable-ep-weight-filter",
**parallel_kwargs["enable_ep_weight_filter"],
@@ -2279,6 +2286,7 @@ class EngineArgs:
data_parallel_hybrid_lb=self.data_parallel_hybrid_lb,
is_moe_model=model_config.is_moe,
enable_expert_parallel=self.enable_expert_parallel,
enable_batch_sharded_sampling=self.enable_batch_sharded_sampling,
enable_ep_weight_filter=self.enable_ep_weight_filter,
all2all_backend=self.all2all_backend,
enable_elastic_ep=self.enable_elastic_ep,