config: a parallel size has one spelling; a patched scope declares its own (#36621)

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

python/sglang/srt/model_executor/cpu_graph_runner.py

@@ -609,9 +609,9 @@ class CPUGraphRunner:
         self.enable_profile_cuda_graph = (
             model_runner.server_args.enable_profile_cuda_graph
         )
-        self.tp_size = get_parallel().config.tp_size
+        self.tp_size = get_parallel().tp_size
         self.dp_size = get_parallel().dp_size
-        self.pp_size = get_parallel().config.pp_size
+        self.pp_size = get_parallel().pp_size
 
         self.capture_forward_mode = ForwardMode.DECODE
         self.capture_hidden_mode = self.return_hidden_states_mode

python/sglang/srt/model_executor/model_runner_components/spec_aux_hidden_state.py

@@ -240,7 +240,7 @@ def _resolve_dflash_draft_cell_size(
             draft_model_config=draft_model_config,
             draft_num_layers=draft_num_layers,
             draft_kv_cache_dtype=draft_kv_cache_dtype,
-            tp_size=get_parallel().config.tp_size,
+            tp_size=get_parallel().tp_size,
         )
     except Exception as e:  # noqa: BLE001
         logger.warning(

python/sglang/srt/model_executor/model_runner_components/startup_weight_load.py

@@ -119,10 +119,10 @@ class StartupWeightLoadOptions:
             prefill_cuda_graph_backend=cuda_graph_config.prefill.backend,
             is_draft_worker=is_draft_worker,
             speculative_algorithm=get_spec().speculative_algorithm,
-            tp_size=get_parallel().config.tp_size,
-            attn_cp_size=get_parallel().config.attn_cp_size,
-            dcp_size=get_parallel().config.dcp_size,
-            pp_size=get_parallel().config.pp_size,
+            tp_size=get_parallel().tp_size,
+            attn_cp_size=get_parallel().attn_cp_size,
+            dcp_size=get_parallel().dcp_size,
+            pp_size=get_parallel().pp_size,
             dp_size=get_parallel().dp_size,
             ep_size=get_parallel().ep_size,
             cpu_offload_gb=get_exec().offload.cpu_offload_gb,

python/sglang/srt/model_executor/runner/base_runner.py

@@ -216,10 +216,10 @@ class BaseRunner(ABC):
         self.model_runner = model_runner
         self.device = model_runner.device
         self.device_module = torch.get_device_module(self.device)
-        self.tp_size = get_parallel().config.tp_size
+        self.tp_size = get_parallel().tp_size
         # elastic-EP scale-up rewrites dp_size on the published config
         self.dp_size = get_parallel().dp_size
-        self.pp_size = get_parallel().config.pp_size
+        self.pp_size = get_parallel().pp_size
         self.enable_pdmux = model_runner.server_args.enable_pdmux
         self.return_hidden_states_mode = (
             CaptureHiddenMode.NULL
@@ -349,7 +349,7 @@ class BaseRunner(ABC):
             vocab_size=mr.model_config.vocab_size,
             dtype=mr.model_config.dtype,
             dp_size=get_parallel().dp_size,
-            pp_size=get_parallel().config.pp_size,
+            pp_size=get_parallel().pp_size,
             is_encoder_decoder=mr.model_config.is_encoder_decoder,
             require_mlp_tp_gather=require_mlp_tp_gather(),
             seq_len_fill_value=mr.attn_backend.get_cuda_graph_seq_len_fill_value(),
@@ -521,7 +521,7 @@ class BaseRunner(ABC):
             extend_prefix_lens = None
             extend_start_loc = None
 
-        if get_parallel().config.pp_size > 1:
+        if get_parallel().pp_size > 1:
             # PP0 already cp-split hidden_states before send.
             pp_hidden_tokens = num_tokens
             if (
@@ -645,7 +645,7 @@ class BaseRunner(ABC):
 
             kwargs = {}
             if (
-                get_parallel().config.pp_size > 1
+                get_parallel().pp_size > 1
                 and "pp_proxy_tensors" in inspect.signature(mr.model.forward).parameters
             ):
                 kwargs["pp_proxy_tensors"] = PPProxyTensors(