diff --git a/python/sglang/srt/layers/attention/nsa_backend.py b/python/sglang/srt/layers/attention/nsa_backend.py
index f4022eb99..807bbab08 100644
--- a/python/sglang/srt/layers/attention/nsa_backend.py
+++ b/python/sglang/srt/layers/attention/nsa_backend.py
@@ -2725,18 +2725,7 @@ class NativeSparseAttnBackend(
                         forward_batch,
                         layer.layer_id,
                         "nsa",
-                        {
-                            "q_all": q_all,
-                            "attn_out": attn_output,
-                            "topk_indices": page_table_1,
-                        },
-                    )
-                    _cp_dump(
-                        forward_batch,
-                        layer.layer_id,
-                        "nsa_kv",
-                        {"kv_cache": kv_cache},
-                        big_layers=(0, 19, 39, 59, 77),
+                        {"q_all": q_all},
                     )
                 except Exception:
                     pass
diff --git a/python/sglang/srt/mem_cache/cp_hicache_trace.py b/python/sglang/srt/mem_cache/cp_hicache_trace.py
index 644177317..dac1ca7f2 100644
--- a/python/sglang/srt/mem_cache/cp_hicache_trace.py
+++ b/python/sglang/srt/mem_cache/cp_hicache_trace.py
@@ -132,12 +132,15 @@ def fwd_hash(forward_batch, layer_id, stage, t, *, level: int = 3) -> None:
     level-2 KV run is unaffected). Guards: only the EAGER EXTEND path (the reload
     case) -- never under cuda-graph decode (.item() sync would corrupt capture);
     handles topk_indices=None and tuple/quant hidden_states."""
-    # Complete-dump accumulation runs regardless of trace level (gated internally
-    # by SGLANG_NSA_DUMP_DIR + rid 'dump-'): captures every stage at every layer.
-    try:
-        dump_tensors(forward_batch, layer_id, stage, {"v": t})
-    except Exception:
-        pass
+    # LEAN dump (gated internally by SGLANG_NSA_DUMP_DIR + rid 'dump-'): only the
+    # residual-stream trajectory (attn_in/attn_out per layer) — enough to localize
+    # the first divergent layer. (The full-stage/kv dump was too heavy and hung the
+    # server.) Stage 2 drills into the localized layer.
+    if stage in ("attn_in", "attn_out"):
+        try:
+            dump_tensors(forward_batch, layer_id, stage, {"v": t})
+        except Exception:
+            pass
     if trace_level() < level:
         return
     try: