Update GLM-5.2 NVFP4 cookbook (#30826)

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

@@ -56,10 +56,6 @@ import { benchmarks } from "/src/snippets/configs/zai-org/glm-5.2-benchmarks.jsx
   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 1-1-2 = 2); high-throughput has no spec.
 </Note>
 
-<Note>
-  **NVFP4 (B300 / GB300)** deploys on the dev image **`lmsysorg/sglang:dev-glm52-nvfp4`** — the command panel's **Docker** toggle selects it automatically. The FP8 / BF16 recipes use the release **`lmsysorg/sglang:latest`** (the `docker pull` in Install above).
-</Note>
-
 ## Playground
 
 The Playground is where you experiment with **SGLang features beyond the verified matrix**. The Deploy panel above only emits combinations the SGLang team has signed off on; the Playground lets you turn on additional knobs on top of whichever cell the Deploy panel is currently showing.
@@ -107,7 +103,7 @@ import { Playground } from "/src/snippets/_playground.jsx";
 
 - **DeepSeek Sparse Attention (DSA).** GLM-5.2 uses the `glm_moe_dsa` architecture; SGLang auto-selects the DSA attention backends (`flashmla_sparse` prefill, `fa3` decode, `sgl-kernel` indexer topk). No attention-backend flag is needed on the supported hardware. SGLang also auto-selects the KV-cache dtype for DSA models — `fp8_e4m3` on Blackwell (B200/GB300/B300, which then routes DSA through the TensorRT-LLM backend) and `bf16` on Hopper (H200) — so no `--kv-cache-dtype` flag is required.
 - **MTP / speculative decoding.** The checkpoint ships one nextn layer. Enable EAGLE MTP for lower latency (`--speculative-algorithm EAGLE --speculative-num-steps 5 --speculative-eagle-topk 1 --speculative-num-draft-tokens 6` for low-latency; `1-1-2` for balanced). The config's `index_share_for_mtp_iteration` reuses the DSA indexer's topk across draft steps (effective only at `--speculative-eagle-topk 1`). **Tune the draft length to the accept length.** GLM-5.2's MTP head is strong — accept length runs high (4+ in many workloads, near-saturating at 5–6 in low-latency runs). Watch the server's reported **accept length** and adjust `--speculative-num-steps` / `--speculative-num-draft-tokens` accordingly: while accept length stays close to the draft-token count there is headroom to push them higher (more accepted tokens per step); if it falls well below, lower them — every rejected draft token is wasted verification compute.
-- **Context Parallelism (CP) for long prefill.** DSA prefill CP splits the long-prefill attention across `--attn-cp-size` ranks. On **Hopper (H200)** this gives a large prefill-latency win at long context — e.g. round-robin CP (`--tp 8 --attn-cp-size 8 --enable-dsa-prefill-context-parallel --dsa-prefill-cp-mode round-robin-split`) cut 64K-token prefill TTFT roughly **2.5–2.8×** vs. plain TP8 in our testing. Trade-offs: CP partitions the KV pool (lower max context at the same `--mem-fraction-static`) and adds some decode-side overhead, so it pays off only for long sequences. **CP is currently verified on Hopper only** — the Blackwell (sm100) DSA-CP FP8 rope kernel is not yet adapted, so leave CP off on B200/B300/GB300.
+- **Context Parallelism (CP) for long prefill.** DSA prefill CP splits the long-prefill attention across `--attn-cp-size` ranks. This gives a large prefill-latency win at long context — e.g. interleaved CP (`--tp 8 --attn-cp-size 8 --enable-prefill-cp --cp-strategy interleave`) cut 64K-token prefill TTFT roughly **2.5–2.8×** vs. plain TP8 in our testing. Trade-offs: CP partitions the KV pool (lower max context at the same `--mem-fraction-static`) and adds some decode-side overhead, so it pays off only for long sequences.
 - **Memory.** The FP8 weights are large (MoE total, not active params). Start around `--mem-fraction-static 0.8` on H200 (TP8) and tune up; raise it for the 4-GPU GB300 single-node layout (TP4).
 - **DP-Attention + DeepEP** for the balanced/high-throughput strategies spreads attention across data-parallel ranks and routes MoE through DeepEP.
 - **BF16 weights need more GPUs.** The full-precision build (`zai-org/GLM-5.2`, ~1.5 TB) does not fit a single 8×H200 / 8×B200 / 4×GB300 node. It fits single-node on **8×B300** (TP8, ~2.1 TB HBM) — **verified**; on the smaller GPUs it needs a **multi-node** layout (e.g. 2×8×H200 or 2×8×B200 at TP16, 2×4×GB300 at TP8), and those **multi-node BF16 recipes are still proposed/inferred** (`verified: false`). FP8 is the recommended deployment. Use the same DSA / MTP / chunked-prefill guidance as FP8. On B300, BF16 low-latency matches FP8 (the sm103 FP8 path is not yet optimized), but FP8 wins at the balanced/high-throughput points.

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

@@ -103,12 +103,11 @@ export const benchmarks = [
   { match: { hw: "gb300", variant: "default", quant: "bf16", strategy: "balanced",        nodes: "multi-2" } },
   { match: { hw: "gb300", variant: "default", quant: "bf16", strategy: "high-throughput", nodes: "multi-2" } },
   // ---- B200 + NVFP4 ----  (8-GPU single node, TP8; nvidia/GLM-5.2-NVFP4 via --quantization modelopt_fp4,
-  // measured on the lmsysorg/sglang:dev-glm52-nvfp4 preview image, flush-cache every run.
+  // flush-cache every run.
   // ttft_ms/tpot_ms are P50; tokens_per_sec_per_gpu = total (in+out) tok/s/GPU (output/GPU × (isl+osl)/osl).
   // balanced & high-throughput add DP-Attention (dp8); low-latency uses MTP 5-1-6, balanced MTP 2-1-3.)
   {
     match: { hw: "b200", variant: "default", quant: "nvfp4", strategy: "low-latency", nodes: "single" },
-    sglang_version: "dev-glm52-nvfp4",
     speed: [
       { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 1 },
         ttft_ms: 295, tpot_ms: 1.85, tokens_per_sec_per_gpu: 527 },
@@ -118,7 +117,6 @@ export const benchmarks = [
   },
   {
     match: { hw: "b200", variant: "default", quant: "nvfp4", strategy: "balanced", nodes: "single" },
-    sglang_version: "dev-glm52-nvfp4",
     speed: [
       { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 64 },
         ttft_ms: 5837, tpot_ms: 12.70, tokens_per_sec_per_gpu: 3770 },
@@ -128,19 +126,17 @@ export const benchmarks = [
   },
   {
     match: { hw: "b200", variant: "default", quant: "nvfp4", strategy: "high-throughput", nodes: "single" },
-    sglang_version: "dev-glm52-nvfp4",
     speed: [
       { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 1024 },
         ttft_ms: 130174, tpot_ms: 67.12, tokens_per_sec_per_gpu: 5305 },
     ],
   },
   // ---- B300 + NVFP4 ----  (8-GPU single node, TP8; nvidia/GLM-5.2-NVFP4 via --quantization modelopt_fp4,
-  // measured on the lmsysorg/sglang:dev-glm52-nvfp4 preview image, flush-cache every run.
+  // flush-cache every run.
   // tokens_per_sec_per_gpu = total (in+out) tok/s/GPU (measured output/GPU 51/224/153/205/430 × (isl+osl)/osl).
   // aime25 overrides the variant default (87.7 → 89.58, measured on this NVFP4 build); gsm8k inherits the default.)
   {
     match: { hw: "b300", variant: "default", quant: "nvfp4", strategy: "low-latency", nodes: "single" },
-    sglang_version: "dev-glm52-nvfp4",
     accuracy: { aime25_pct: 89.58 },
     speed: [
       { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 1 },
@@ -151,7 +147,6 @@ export const benchmarks = [
   },
   {
     match: { hw: "b300", variant: "default", quant: "nvfp4", strategy: "balanced", nodes: "single" },
-    sglang_version: "dev-glm52-nvfp4",
     accuracy: { aime25_pct: 89.58 },
     speed: [
       { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 64 },
@@ -162,7 +157,6 @@ export const benchmarks = [
   },
   {
     match: { hw: "b300", variant: "default", quant: "nvfp4", strategy: "high-throughput", nodes: "single" },
-    sglang_version: "dev-glm52-nvfp4",
     accuracy: { aime25_pct: 89.58 },
     speed: [
       { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 1024 },

docs_new/src/snippets/configs/zai-org/glm-5.2.jsx

@@ -97,10 +97,6 @@ sgl-eval run aime25 \\
     mi355x: "lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260618",
     mi325x: "lmsysorg/sglang-rocm:v0.5.13.post1-rocm700-mi30x-20260616",
     mi300x: "lmsysorg/sglang-rocm:v0.5.13.post1-rocm700-mi30x-20260616",
-    // NVFP4 needs the dev image with modelopt_fp4 support (per-quant override).
-    "b200|nvfp4":  "lmsysorg/sglang:dev-glm52-nvfp4",
-    "b300|nvfp4":  "lmsysorg/sglang:dev-glm52-nvfp4",
-    "gb300|nvfp4": "lmsysorg/sglang:dev-glm52-nvfp4",
   },
 
   github: {
@@ -668,7 +664,6 @@ sgl-eval run aime25 \\
     // high-throughput add DP-Attention (dp8). low-latency uses MTP 5-1-6, balanced MTP 2-1-3.
     // GB300: 4-GPU single node, TP4 (the node fits the ~381 GB build); GB300 adds dp4 on
     // balanced & high-throughput; low-latency uses MTP 5-1-6.
-    // Blackwell NVFP4 measured on the dev-glm52-nvfp4 preview image.
     // ====================================================================
     {
       match: { hw: "b200", variant: "default", quant: "nvfp4", strategy: "low-latency", nodes: "single" },
@@ -743,6 +738,11 @@ sgl-eval run aime25 \\
         "--speculative-num-draft-tokens 6",
         "--chunked-prefill-size 8192",
         "--mem-fraction-static 0.85",
+        "--kv-cache-dtype fp8_e4m3",
+        "--bf16-gemm-backend cutedsl",
+        "--max-running-requests 16",
+        "--cuda-graph-max-bs 16",
+        "--max-prefill-tokens 8192",
         "--host {{HOST_IP}}",
         "--port {{PORT}}",
       ],
@@ -805,6 +805,11 @@ sgl-eval run aime25 \\
         "--speculative-num-draft-tokens 6",
         "--chunked-prefill-size 8192",
         "--mem-fraction-static 0.85",
+        "--kv-cache-dtype fp8_e4m3",
+        "--bf16-gemm-backend cutedsl",
+        "--max-running-requests 16",
+        "--cuda-graph-max-bs 16",
+        "--max-prefill-tokens 8192",
         "--host {{HOST_IP}}",
         "--port {{PORT}}",
       ],