[BugFix] fix gpt-oss accuracy issue when enabling piecewise cuda graph (#18013)

python/sglang/srt/managers/schedule_batch.py

@@ -2283,6 +2283,15 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
     def maybe_evict_swa(self):
         if self.tree_cache.supports_swa():
             sliding_window_size = self.tree_cache.sliding_window_size
+            server_args = get_global_server_args()
+
+            if (
+                self.forward_mode.is_decode()
+                and server_args.enable_piecewise_cuda_graph
+                and not self.tree_cache.is_chunk_cache()
+            ):
+                return
+
             for idx, req in enumerate(self.reqs):
                 if self.forward_mode.is_decode():
                     # We set evict_swa condition here with two reasons:
@@ -2297,7 +2306,6 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
                         if req.extend_batch_idx < 2:
                             continue
                         else:
-                            server_args = get_global_server_args()
                             pre_len = (
                                 pre_len - server_args.chunked_prefill_size
                                 if server_args.chunked_prefill_size > 0