[Spec] Emit step trace span for multi-layer draft-extend graph replays (#32850)

python/sglang/srt/speculative/multi_layer_eagle_worker_v2.py

@@ -82,6 +82,8 @@ from sglang.srt.utils.async_probe import (
     maybe_detect_oob,
 )
 from sglang.srt.utils.common import empty_context, fast_topk
+from sglang.srt.utils.nvtx_utils import profile_range
+from sglang.srt.utils.profile_utils import build_step_span_name
 
 _is_npu = is_npu()
 _is_cpu = is_cpu()
@@ -738,6 +740,9 @@ class MultiLayerEagleDraftWorker(EagleDraftWorkerBase):
         next_token_ids_backup = batch_result.next_token_ids.clone()
 
         if can_cuda_graph:
+            # Graph replay bypasses ModelRunner.forward, which emits the
+            # step[...] trace span for every other phase; emit it here.
+            with profile_range(build_step_span_name(forward_batch)):
                 cgr = self.cuda_graph_runner_for_draft_extend
                 # Populate the single shared buffer set once; each step replays
                 # against it and the chain is advanced in place between steps.