Fix dp-attention token alignment in the dumper comparator e2e test (#26996)

test/registered/debug_utils/test_engine_dumper_comparator_e2e.py

@@ -195,6 +195,14 @@ class TestSourcePatcherE2ESGLang:
         step has tokens on both DP ranks, which breaks the dp:=attn_dp
         single-rank assumption and causes comparator errors.
 
+        The ``concat_steps`` token aligner is required because dp-attention
+        gathers tokens across DP ranks before the dump point, so the
+        non-empty rank's buffer holds the real tokens plus padding tokens
+        contributed by the empty DP rank (with a single request one DP
+        rank is always empty).  The aligner reconstructs the real per-step
+        token sequence from the dumped seq-lens, trimming that padding so
+        the target lines up with the un-padded TP baseline.
+
         mlp_output is allowed to fail because the FusedMoE dispatcher
         combine path may include an implicit all-reduce that makes the
         dumped value differ from the raw partial expert output.  All
@@ -208,6 +216,8 @@ class TestSourcePatcherE2ESGLang:
             extra_target_server_args=["--dp", "2", "--enable-dp-attention"],
             target_patch_config_yaml=PATCH_CONFIG_DP_ATTENTION_YAML,
             extra_comparator_args=[
+                "--token-aligner",
+                "concat_steps",
                 "--end-step",
                 "0",
                 "--allow-failed-pattern",