[Docs] Fill GLM-5.2 H200 FP8 speed cells (low-latency, balanced); fix MTP notation (#31554)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

docs/cookbook/autoregressive/GLM/GLM-5.2.mdx

@@ -57,7 +57,7 @@ import { benchmarks } from "/src/snippets/configs/zai-org/glm-5.2-benchmarks.jsx
 </Warning>
 
 <Note>
-  Speed numbers are measured with `--random-range-ratio 1.0`, `--flush-cache`, on `main @ 09ca4fc`. Spec cells pin the EAGLE acceptance length via the serve env `SGLANG_SIMULATE_ACC_LEN` (low-latency 5-1-6 = 3.5, balanced 2-1-3 = 2); high-throughput has no spec.
+  Speed numbers are measured with `--random-range-ratio 1.0`, `--flush-cache`, on `main @ 09ca4fc` (H200 FP8 cells: `v0.5.14 @ 49e384ce`). Spec cells pin the EAGLE acceptance length via the serve env `SGLANG_SIMULATE_ACC_LEN` (low-latency 5-1-6 = 3.5; FP8 balanced 1-1-2 = 2; NVFP4 balanced 2-1-3 = 2); high-throughput has no spec.
 </Note>
 
 ## Playground

docs/src/snippets/configs/zai-org/glm-5.2-benchmarks.jsx

@@ -3,10 +3,41 @@
 // Numbers pending: each entry is a bare `match` stub (renders "pending") until measured
 // end-to-end on the corresponding hardware, then filled with sglang_version + speed/accuracy.
 export const benchmarks = [
-  // ---- H200 + FP8 ----  (serve recipe in glm-5.2.jsx; benchmark pending re-measurement)
+  // ---- H200 + FP8 ----  (8-GPU single node; serve recipe in glm-5.2.jsx; real weights,
-  { match: { hw: "h200", variant: "default", quant: "fp8", strategy: "low-latency", nodes: "single" } },
+  // --random-range-ratio 1.0, flush-cache every run)
-  { match: { hw: "h200", variant: "default", quant: "fp8", strategy: "balanced", nodes: "single" } },
+  {
-  { match: { hw: "h200", variant: "default", quant: "fp8", strategy: "high-throughput", nodes: "single" } },
+    // EAGLE MTP 5-1-6, mfs 0.8. env SGLANG_SIMULATE_ACC_LEN=3.5
+    // (match-expected: 50% accept 3 / 50% accept 4) fixes the acceptance length.
+    match: { hw: "h200", variant: "default", quant: "fp8", strategy: "low-latency", nodes: "single" },
+    sglang_version: "v0.5.14 @ 49e384ce",
+    speed: [
+      { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 1 },
+        ttft_ms: 668, tpot_ms: 5.05, tokens_per_sec_per_gpu: 197 },
+      { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 16 },
+        ttft_ms: 6148, tpot_ms: 16.11, tokens_per_sec_per_gpu: 813 },
+    ],
+  },
+  {
+    // Balanced: DP8 + deepep + mfs 0.85 + chunked-prefill 32768 (÷dp8 = 4096) + max-running 256,
+    // 1-1-2 EAGLE. env SGLANG_SIMULATE_ACC_LEN=2 (match-expected: accept 2 of 2 draft tokens).
+    match: { hw: "h200", variant: "default", quant: "fp8", strategy: "balanced", nodes: "single" },
+    sglang_version: "v0.5.14 @ 49e384ce",
+    speed: [
+      { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 64 },
+        ttft_ms: 7473, tpot_ms: 23.49, tokens_per_sec_per_gpu: 2343 },
+      { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 256 },
+        ttft_ms: 80562, tpot_ms: 28.08, tokens_per_sec_per_gpu: 2391 },
+    ],
+  },
+  {
+    // HT: DP8 + deepep + mfs 0.85 + max-running 256, no spec (so no SIMULATE_ACC_LEN).
+    match: { hw: "h200", variant: "default", quant: "fp8", strategy: "high-throughput", nodes: "single" },
+    sglang_version: "v0.5.14 @ 49e384ce",
+    speed: [
+      { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 1024 },
+        ttft_ms: 553480, tpot_ms: 61.71, tokens_per_sec_per_gpu: 1656 },
+    ],
+  },
   // ---- B200 + FP8 ----  (8-GPU single node, TP8; real weights, --random-range-ratio 1.0, flush-cache every run)
   {
     // EAGLE MTP 5-1-6, mfs 0.8, no cuda-graph-max-bs. env SGLANG_SIMULATE_ACC_LEN=3.5