diff --git a/tests/analytical_visualization/app.py b/tests/analytical_visualization/app.py
index c7b4c50..4aca6bc 100644
--- a/tests/analytical_visualization/app.py
+++ b/tests/analytical_visualization/app.py
@@ -176,14 +176,68 @@ if st.session_state.get("_last_preset") != preset_name:
 
 with st.sidebar.expander("Parallelism", expanded=True):
     st.caption(
-        f"Auto-suggest: CP={_default_suggestion.cp}, "
+        f"Auto-suggest (memory-min): CP={_default_suggestion.cp}, "
         f"TP={_default_suggestion.tp}, PP={_default_suggestion.pp}"
     )
-    if st.button("Apply auto-suggest", width='stretch'):
-        st.session_state["cp"] = _snap(_default_suggestion.cp, _CP_OPTS)
-        st.session_state["tp"] = _snap(_default_suggestion.tp, _TP_OPTS)
-        st.session_state["pp"] = _snap(_default_suggestion.pp, _PP_OPTS)
-        st.rerun()
+    # Three latency-optimal apply buttons — pick a scope, load the
+    # latency-min Pareto config into the sidebar sliders.
+    from tests.analytical_visualization.auto_explore import (
+        run_auto_explore as _run_auto_explore_sb,
+    )
+    _sb_scope_meta = {
+        "attn": (True,  False, "Attention"),
+        "ffn":  (False, True,  "FFN/MoE"),
+        "full": (True,  True,  "Attn+FFN"),
+    }
+    _sb_cols = st.columns(3)
+    _sb_clicks = {}
+    for _sb_scope, _sb_col in zip(("attn", "ffn", "full"), _sb_cols):
+        with _sb_col:
+            _sb_clicks[_sb_scope] = st.button(
+                _sb_scope_meta[_sb_scope][2], width='stretch',
+                key=f"_sb_apply_{_sb_scope}",
+                help=f"Apply latency-optimal parallelism for "
+                     f"{_sb_scope_meta[_sb_scope][2]} scope.",
+            )
+    for _sb_scope, _sb_click in _sb_clicks.items():
+        if not _sb_click:
+            continue
+        _sb_ia, _sb_ff, _sb_lbl = _sb_scope_meta[_sb_scope]
+        with st.spinner(f"Finding latency-optimal parallelism ({_sb_lbl})..."):
+            _sb_res = _run_auto_explore_sb(
+                model, _default_machine, s_kv=s_kv, mode=mode,
+                include_attention=_sb_ia, include_ffn=_sb_ff,
+            )
+        if not _sb_res.pareto_scores:
+            st.error(
+                f"No feasible parallelism for {_sb_lbl}. Try raising "
+                "`pe_hbm_gb` or reducing `s_kv`."
+            )
+        else:
+            _sb_best = min(_sb_res.pareto_scores,
+                            key=lambda s: s.total_latency_ns)
+            st.session_state["cp"] = _snap(_sb_best.cp, _CP_OPTS)
+            st.session_state["tp"] = _snap(_sb_best.tp, _TP_OPTS)
+            st.session_state["pp"] = _snap(_sb_best.pp, _PP_OPTS)
+            st.session_state["dp"] = _snap(_sb_best.dp, _DP_OPTS)
+            st.session_state["tp_placement"] = _sb_best.tp_placement
+            st.session_state["cp_placement"] = _sb_best.cp_placement
+            st.session_state["cp_ring_variant"] = _sb_best.cp_ring_variant
+            st.session_state["kv_mode"] = _sb_best.kv_shard_mode
+            _sb_tp2 = _snap(_sb_best.tp, _TP_OPTS)
+            _sb_cp2 = _snap(_sb_best.cp, _CP_OPTS)
+            _sb_dp2 = _snap(_sb_best.dp, _DP_OPTS)
+            _sb_ffn_map = {
+                "TP":       f"TP only (div={max(1, _sb_tp2)})",
+                "TP+CP":    f"TP*CP (div={max(1, _sb_tp2 * _sb_cp2)})",
+                "TP+CP+DP": f"TP*CP*DP (div={max(1, _sb_tp2 * _sb_cp2 * _sb_dp2)})",
+            }
+            st.session_state["ffn_scope_label"] = \
+                _sb_ffn_map[_sb_best.ffn_shard_scope]
+            # Track for the Physical Layout scope filter, so the tables
+            # below also filter to this scope.
+            st.session_state["_pl_active_scope"] = _sb_scope
+            st.rerun()
 
     # Sharding dims — arrange in 2x3 grid to save vertical space
     p_row1 = st.columns(3)