Check the topology identities where the layout is written, and build at the published widths (#40340)

benchmark/kernels/fused_moe_triton/benchmark_sglang_fused_moe_triton.py

@@ -26,6 +26,7 @@ from sglang.srt.layers.moe.topk import (
     select_experts,
 )
 from sglang.srt.server_args import ServerArgs, set_global_server_args_for_scheduler
+from sglang.test.test_utils import publish_build_topology
 
 
 def fused_moe_triton_api(
@@ -227,10 +228,8 @@ def main():
             backend="nccl" if torch.cuda.is_available() else "gloo",
         )
 
-        initialize_model_parallel(
-            tensor_model_parallel_size=1,
-            expert_model_parallel_size=1,
-        )
+        publish_build_topology()
+        initialize_model_parallel()
 
         model_config = get_model_config(args.model, args.tp_size, args.ep_size)
         benchmark.run(

benchmark/kernels/fused_moe_triton/benchmark_vllm_vs_sglang_fused_moe_triton.py

@@ -15,6 +15,7 @@ from sglang.srt.distributed.parallel_state import (
 from sglang.srt.layers.moe.moe_runner.triton_utils.fused_moe import (
     fused_moe as fused_moe_sglang,
 )
+from sglang.test.test_utils import publish_build_topology
 
 from .common_utils import get_model_config
 
@@ -243,10 +244,8 @@ def main():
             backend="nccl" if torch.cuda.is_available() else "gloo",
         )
 
-        initialize_model_parallel(
-            tensor_model_parallel_size=1,
-            pipeline_model_parallel_size=1,
-        )
+        publish_build_topology()
+        initialize_model_parallel()
 
         shape_configs = get_model_config(args.model, args.tp_size, args.ep_size)
         benchmark.run(