config: constructing a config no longer resolves it (#35907)

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

python/sglang/benchmark/endpoint.py

@@ -48,6 +48,11 @@ def _launch_server_target(launch_server_func: Callable, server_args: ServerArgs)
 
 
 def launch_or_reuse_server(launch_server_func: Callable, server_args: ServerArgs):
+    # Resolve in the parent, before the fork. The pipeline probes the device
+    # (the default attention backend reads the CUDA capability), and a forked
+    # child cannot re-initialize CUDA once this process has.
+    server_args.resolve_once()
+
     base_url = resolve_base_url("", server_args.host, server_args.port)
 
     # Reuse an already-running server instead of forking a second one onto the

python/sglang/benchmark/offline_throughput.py

@@ -398,7 +398,7 @@ def _create_ray_engine_backend(server_args: ServerArgs):
             placement_group=pg,
             placement_group_bundle_index=0,
         ),
-    ).remote(**dataclasses.asdict(server_args))
+    ).remote(**dict(server_args._raw_input))
 
     class _Proxy:
         """Forwards method calls to the remote RayEngine actor."""
@@ -432,15 +432,18 @@ def throughput_test(
     server_args: ServerArgs,
     bench_args: BenchArgs,
 ):
+    # A programmatic caller may hand over a freshly constructed record, and
+    # the backends below read the resolved paths and the raw snapshot.
+    server_args.resolve_once()
     if bench_args.backend == "engine":
         if server_args.use_ray:
             backend = _create_ray_engine_backend(server_args)
         else:
-            backend = Engine(**dataclasses.asdict(server_args))
+            backend = Engine(server_args=server_args)
         if not backend:
             raise ValueError("Please provide valid engine arguments")
     elif bench_args.backend == "runtime":
-        backend = Runtime(**dataclasses.asdict(server_args))
+        backend = Runtime(**dict(server_args._raw_input))
     else:
         raise ValueError('Please set backend to either "engine" or "runtime"')
 
@@ -569,6 +572,7 @@ def cli_main():
                 raise e
 
     server_args = ServerArgs.from_cli_args(args)
+    server_args.resolve_once()
     bench_args = BenchArgs.from_cli_args(args)
 
     logging.basicConfig(

python/sglang/benchmark/one_batch.py

@@ -989,8 +989,10 @@ def latency_test(
 
 
 def main(server_args, bench_args):
-    # Post-init write to the legacy cuda_graph_max_bs_decode field would
-    # not propagate to cuda_graph_config; update the decode phase directly.
+    server_args.resolve_once()
+
+    # The legacy cuda_graph_max_bs_decode field does not propagate; set the
+    # decode phase.
     if server_args.cuda_graph_config is not None:
         server_args.cuda_graph_config[Phase.DECODE].max_bs = max(bench_args.batch_size)
 

python/sglang/benchmark/one_batch_server.py

@@ -1255,6 +1255,7 @@ def cli_main():
     args = parser.parse_args()
 
     server_args = ServerArgs.from_cli_args(args)
+    server_args.resolve_once()
     bench_args = BenchArgs.from_cli_args(args)
 
     run_benchmark(server_args, bench_args)