Support sequence parallel and trivial dims in dump comparator (#19460)

python/sglang/srt/debug_utils/comparator/aligner/unsharder/planner.py

@@ -31,9 +31,15 @@ def compute_unsharder_plan(
     sharded_axis_infos: dict[ParallelAxis, DimSpec] = {
         spec.parallel: spec for spec in dim_specs if spec.parallel is not None
     }
-    sharded_axes: set[ParallelAxis] = set(sharded_axis_infos)
+    sharded_axes_raw: set[ParallelAxis] = set(sharded_axis_infos)
 
     all_axes: set[ParallelAxis] = {axis for info in parallel_infos for axis in info}
+
+    # axis annotated in dims but absent from all parallel_infos -> axis_size=1, skip
+    sharded_axes: set[ParallelAxis] = sharded_axes_raw & all_axes
+    sharded_axis_infos = {
+        k: v for k, v in sharded_axis_infos.items() if k in sharded_axes
+    }
     replicated_axes: set[ParallelAxis] = all_axes - sharded_axes
 
     if not sharded_axes and not replicated_axes:

python/sglang/srt/debug_utils/dumper.py

@@ -1238,6 +1238,18 @@ class _MegatronPlugin(_FrameworkPlugin):
         except (AttributeError, AssertionError):
             info["megatron_error"] = True
 
+        # Megatron sequence parallel reuses the TP group (no dedicated parallel state API).
+        # When sequence_parallel=True, inject sp_rank/sp_size for the comparator unsharder.
+        try:
+            from megatron.training.global_vars import get_args
+
+            args = get_args()
+            if getattr(args, "sequence_parallel", False) and "tp_rank" in info:
+                info["sp_rank"] = info["tp_rank"]
+                info["sp_size"] = info["tp_size"]
+        except (ImportError, AssertionError, AttributeError):
+            pass
+
         return info
 
     def convert_value(