From 77abc95d7816ee866fe8eddc6c2b303f58583945 Mon Sep 17 00:00:00 2001
From: Mukesh Garg <mukeshgarg81@gmail.com>
Date: Wed, 29 Jul 2026 15:13:08 -0700
Subject: [PATCH] =?UTF-8?q?analytical-viz:=20capacity=20planning=20tab=20?=
 =?UTF-8?q?=E2=80=94=20add=20rules=20of=20thumb=20+=20deployment=20templat?=
 =?UTF-8?q?es?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Two more table sections on the Capacity planning tab:

**Section 4 — Practical rules of thumb (by context regime)**
- Short (S_kv < L*): pack HBM, B ~ 2·B*, high util, cheap tier
- Long (S_kv > L*): fewer users/replica, low B, util drops as
  1/(1 + S_kv/L*), pricier
- Extreme (S_kv >> L*): dedicated pool, heavy CP, disaggregated
  prefill, very low util without sparse attention

Columns: Regime, Batch strategy, Utilization, Cost/token, Deployment.

**Section 5 — Sample deployment templates**
Same base model, three different sharding recipes routed to by the
API gateway based on request context length:
- Config_small : CP=1,  TP=8, PP=1 → 8 GPUs, up to 32k, B=64
- Config_medium: CP=4,  TP=8, PP=1 → 32 GPUs, up to 128k, B=32
- Config_large : CP=32, TP=8, PP=1 → 256 GPUs, up to 1M, B=4

Columns: Tier, CP, TP, PP, Total GPUs/replica, Max context, Typical
B, Best for.

Playbook section renumbered to #6.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---
 tests/analytical_visualization/app.py | 64 ++++++++++++++++++++++++++-
 1 file changed, 62 insertions(+), 2 deletions(-)

diff --git a/tests/analytical_visualization/app.py b/tests/analytical_visualization/app.py
index c84702a..e1828fe 100644
--- a/tests/analytical_visualization/app.py
+++ b/tests/analytical_visualization/app.py
@@ -2828,8 +2828,68 @@ with tab_planning:
 
     st.divider()
 
-    # ── Section 4: Binding-axis playbook ──────────────────────────
-    st.markdown("### 4. What to do when each axis binds")
+    # ── Section 4: Practical rules of thumb by context regime ─────
+    st.markdown("### 4. Practical rules of thumb (by context regime)")
+    _rules_rows = [
+        {"Regime": "Short context (S_kv < L*)",
+         "Batch strategy": "Pack as many users as HBM allows; aim for "
+                           "B ≈ 2·B*",
+         "Utilization": "High (near compute floor)",
+         "Cost / token": "Low — cheap tier",
+         "Deployment": "Standard elastic pool"},
+        {"Regime": "Long context (S_kv > L*)",
+         "Batch strategy": "Fewer users per replica; more HBM per user; "
+                           "lower B",
+         "Utilization": "Drops as 1 / (1 + S_kv/L*)",
+         "Cost / token": "Higher — accept lower utilization",
+         "Deployment": "Length-tier replica with more CP sharding"},
+        {"Regime": "Extreme long context (S_kv ≫ L*)",
+         "Batch strategy": "Dedicated pool, heavy CP; often "
+                           "disaggregated prefill",
+         "Utilization": "Very low without sparse attention / MLA",
+         "Cost / token": "Much higher — priced accordingly",
+         "Deployment": "Long-context tier + disaggregated prefill/decode"},
+    ]
+    st.dataframe(pd.DataFrame(_rules_rows), width='stretch',
+                  hide_index=True)
+
+    st.divider()
+
+    # ── Section 5: Sample deployment templates ────────────────────
+    st.markdown("### 5. Sample deployment templates (same model, different sharding)")
+    st.caption(
+        "One base model runs on all three configs. What changes is the "
+        "sharding (CP/TP/PP) that shapes the effective 'small' vs "
+        "'large' setup. The API gateway routes each request to the "
+        "right tier based on its context length."
+    )
+    _tmpl_rows = [
+        {"Tier": "Config_small",
+         "CP": 1, "TP": 8, "PP": 1,
+         "Total GPUs / replica": 8,
+         "Max context": "up to 32k",
+         "Typical B": 64,
+         "Best for": "Chat, RAG, short prompts — high utilization"},
+        {"Tier": "Config_medium",
+         "CP": 4, "TP": 8, "PP": 1,
+         "Total GPUs / replica": 32,
+         "Max context": "up to 128k",
+         "Typical B": 32,
+         "Best for": "Long-doc analysis, code review — standard tier"},
+        {"Tier": "Config_large",
+         "CP": 32, "TP": 8, "PP": 1,
+         "Total GPUs / replica": 256,
+         "Max context": "up to 1M",
+         "Typical B": 4,
+         "Best for": "Whole-repo / long-video / massive-doc — premium tier"},
+    ]
+    st.dataframe(pd.DataFrame(_tmpl_rows), width='stretch',
+                  hide_index=True)
+
+    st.divider()
+
+    # ── Section 6: Binding-axis playbook ──────────────────────────
+    st.markdown("### 6. What to do when each axis binds")
     _playbook_rows = [
         {"Binding axis": "A. Capacity (weights)",
          "Symptom": "Weights alone are close to per-PE HBM budget",