[docs][cookbook] Laguna-M.1 playground: add HiCache; refresh EP / DP-Attention notes (#28774)

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Zijie Xia <zijie.xia@radixark.ai>

docs_new/src/snippets/configs/poolside/laguna-m1.jsx

@@ -118,9 +118,13 @@ sgl-eval run gsm8k \\
 
   playgroundFeatures: {
 
-    // M.1 is global-attention (no SWA); expose TP + DP-Attention. No CP.
-    // DP-Attention is a Playground experiment only — ~15% slower than plain TP on this GQA
-    // model (8 KV heads), so it is NOT in the shipped Balanced recipe.
+    // M.1 is global-attention (no SWA); expose TP + DP-Attention here. No CP: the default
+    // trtllm_mha backend has no CP-aware KV-store (crashes), and the engine's built-in attention CP
+    // knob emits NSA flags (--enable-nsa-prefill-context-parallel) that apply to DeepSeek-family
+    // models, not M.1. (CP works only via the fa3 backend, which is Hopper SM90 — left out here.)
+    // DP-Attention: VERIFIED functionally correct on 8×B200 BF16 (GSM8K 0.94, identical to the TP
+    // baseline) but ~15–28% slower on this GQA model (8 KV heads). Playground experiment only —
+    // deliberately NOT in the shipped Balanced recipe.
     attention: {
       knobs: [
         { id: "tp",     label: "TP",           values: [null, 1, 2, 4, 8] },
@@ -129,14 +133,12 @@ sgl-eval run gsm8k \\
       ],
     },
 
-    // 256-expert top-16 MoE. DeepEP all-to-all + EP degree.
+    // 256-expert top-16 MoE — EP degree only.
+    // EP: VERIFIED on 8×B200 BF16 (--ep-size 8, GSM8K 0.94, identical to the TP baseline).
+    // DeepEP intentionally NOT exposed — it does not work on M.1: top-16 routing exceeds DeepEP's
+    // low-latency internode kernel cap of 11 (internode_ll.cu kNumMaxTopK=11) → assert at decode
+    // CUDA-graph capture, and `--deepep-mode normal` is NotImplemented for unquantized weights. Use EP.
     moe: {
-      backend: {
-        options: [
-          { id: null,     label: "Inherited" },
-          { id: "deepep", label: "DeepEP", flags: ["--moe-a2a-backend deepep"] },
-        ],
-      },
       ep: { label: "EP", values: [null, 1, 2, 4, 8] },
     },
 
@@ -149,6 +151,20 @@ sgl-eval run gsm8k \\
       ],
     },
 
+    // HiCache (hierarchical KV cache). VERIFIED on 8×B200 BF16: enabling the host L2 tier on a
+    // zipfian shared-prefix workload cut mean TTFT ~36% (median ~43%) and lifted throughput ~19% vs
+    // GPU-only, with ~1.14M tokens served from the host tier (TPOT unchanged — the win is on prefill
+    // / prefix reuse). Biggest gains on reuse-heavy traffic: shared system prompts, multi-turn
+    // agentic coding, repeated long contexts. "Enable" emits --enable-hierarchical-cache (+host L2);
+    // Write policy is optional. (L3 storage backends exist but were not validated, so none exposed.)
+    hicache: {
+      writePolicies: [
+        { id: "auto",          label: "Auto" },
+        { id: "write_through", label: "Write-through" },
+        { id: "write_back",    label: "Write-back" },
+      ],
+    },
+
     // Prefill-Decode disaggregation (§3.3). M.1 is standard-KV (global attention, no sparse
     // index buffer), so it disaggregates with just the --disaggregation-* flags — no model-specific
     // backend pinning. Verified on 2×8×H200 (TP8+TP8, BF16) over InfiniBand.