From 770aed629163f0db22139b92f68a615eb251ab47 Mon Sep 17 00:00:00 2001
From: Mukesh Garg <mukeshgarg81@gmail.com>
Date: Tue, 28 Jul 2026 22:24:38 -0700
Subject: [PATCH] =?UTF-8?q?analytical-viz:=20Auto=20Suggest=20tab=20?=
 =?UTF-8?q?=E2=80=94=20add=20Top-10=20by=20memory=20footprint=20table?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Adds a second table beneath the Pareto configurations table, sorted by
(hbm_utilization ↑, pes_used ↑, latency ↑) — top 10 smallest-memory
Pareto configs first.

Columns highlight the memory story:
  HBM % | weights (GB) | KV (GB) | PEs | SIPs | lat (ms) | CP TP PP DP | kv | ffn | tp_place | cp_place

Useful when memory pressure is the real binding constraint of a
deployment — e.g., picking a config for a per-PE HBM-limited SIP,
or spotting configs that trade small latency headroom for large
memory savings.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---
 tests/analytical_visualization/app.py | 29 +++++++++++++++++++++++++++
 1 file changed, 29 insertions(+)

diff --git a/tests/analytical_visualization/app.py b/tests/analytical_visualization/app.py
index df07fcb..b151e3c 100644
--- a/tests/analytical_visualization/app.py
+++ b/tests/analytical_visualization/app.py
@@ -1685,6 +1685,35 @@ def _render_auto_explore_tab():
     _df = pd.DataFrame(_rows)
     st.dataframe(_df, width='stretch', hide_index=True)
 
+    # ── Top-N by memory (smallest → largest) ─────────────────────
+    st.markdown(
+        "**Top 10 configurations by memory footprint** "
+        "(smallest HBM% first). Useful when memory pressure is the "
+        "real binding constraint of your deployment."
+    )
+    _pareto_by_mem = sorted(
+        _res.pareto_scores,
+        key=lambda s: (s.hbm_utilization, s.pes_used, s.total_latency_ns),
+    )[:10]
+    _mem_rows = []
+    for i, s in enumerate(_pareto_by_mem):
+        _mem_rows.append({
+            "#": i,
+            "HBM %": round(s.hbm_utilization * 100, 1),
+            "weights (GB)": round(s.weights_gb, 2),
+            "KV (GB)": round(s.kv_gb, 2),
+            "PEs": s.pes_used,
+            "SIPs": s.sips_used,
+            "lat (ms)": round(s.latency_ms, 3),
+            "CP": s.cp, "TP": s.tp, "PP": s.pp, "DP": s.dp,
+            "kv": s.kv_shard_mode,
+            "ffn": s.ffn_shard_scope,
+            "tp_place": s.tp_placement,
+            "cp_place": s.cp_placement,
+        })
+    _df_mem = pd.DataFrame(_mem_rows)
+    st.dataframe(_df_mem, width='stretch', hide_index=True)
+
     # ── Parallelism sensitivity subplots ──────────────────────
     st.markdown(
         "**Parallelism sensitivity** — for the *baseline* config picked "