docs: add DSPARK speculative decoding option to Ling-3.0-flash cookbook (#35861)

docs/cookbook/autoregressive/InclusionAI/Ling-3.0-flash.mdx

@@ -20,7 +20,7 @@ For how to launch the image, see [Install → Method 3: Using Docker](../../../d
 
 Pick your hardware + recipe to generate the launch command. Three serving strategies are covered:
 
-- **Low-Latency** — fastest reply for a single user. Pick for chat. These recipes run NEXTN speculative decoding.
+- **Low-Latency** — fastest reply for a single user. Pick for chat. These recipes run speculative decoding: pick **NEXTN** (built-in MTP layer, no extra checkpoint) or **DSPARK** (external draft model) in the **Spec Decode** selector of the Deploy panel.
 - **High-Throughput** — most tokens per second across many users. Best for batch jobs. These recipes turn speculative decoding off, since at saturation the draft/verify overhead outweighs the speedup.
 - **HiCache + Mooncake** — writes reusable prefixes to Mooncake L3 storage. Start the Mooncake services in §3.3 before launching the generated server command.
 
@@ -50,6 +50,7 @@ It is a hybrid-reasoning model with thinking enabled by default, and it supports
 - **FP8** (blockwise E4M3): [inclusionAI/Ling-3.0-flash-fp8](https://huggingface.co/inclusionAI/Ling-3.0-flash-fp8)
 - **INT4** (compressed-tensors W4A16): [inclusionAI/Ling-3.0-flash-int4](https://huggingface.co/inclusionAI/Ling-3.0-flash-int4)
 - **MXFP4**: [inclusionAI/Ling-3.0-flash-fp4](https://huggingface.co/inclusionAI/Ling-3.0-flash-fp4)
+- **DSPARK draft**: [inclusionAI/Ling-3.0-flash-dspark](https://huggingface.co/inclusionAI/Ling-3.0-flash-dspark) — 5-layer DSpark draft checkpoint for DSPARK speculative decoding
 
 **License:** MIT
 
@@ -64,7 +65,8 @@ It is a hybrid-reasoning model with thinking enabled by default, and it supports
 - `--reasoning-parser ling3` and `--tool-call-parser ling3` enable Ling-3.0-specific reasoning and structured tool-call parsing; toggle them in the **Parsers** card of the [Playground](#playground).
 - Both the chat template and the `ling3` reasoning parser default to thinking on. A single request can turn it off with `"chat_template_kwargs": {"enable_thinking": false}` (see §3.1).
 - The BF16/FP8 recipes use `--mem-fraction-static 0.8`; INT4/MXFP4 use `0.85`. These values reserve the headroom used by the validated graph-enabled runs.
-- The checkpoint ships a built-in MTP layer (`num_nextn_predict_layers: 1`); enable it with `--speculative-algorithm NEXTN` — no separate draft model is needed. The Low-Latency recipes have it on; toggle it in the **Speculative Decoding** card of the [Playground](#playground).
+- The checkpoint ships a built-in MTP layer (`num_nextn_predict_layers: 1`); enable it with `--speculative-algorithm NEXTN` — no separate draft model is needed. The Low-Latency recipes default to it via the **Spec Decode** selector; toggle it in the **Speculative Decoding** card of the [Playground](#playground).
+- DSPARK is the alternative speculative path: it drafts with the external [inclusionAI/Ling-3.0-flash-dspark](https://huggingface.co/inclusionAI/Ling-3.0-flash-dspark) checkpoint instead of the built-in MTP layer. Pick **DSPARK** in the Deploy panel's **Spec Decode** selector (also available in the [Playground](#playground)); it composes `--speculative-algorithm DSPARK`, the draft path, and `--enable-linear-replayssm-spec`, the KDA verify path that folds per-draft intermediate states onto a fixed ring instead of snapshotting full states. The draft block size auto-infers from the checkpoint; tune it with `--speculative-dspark-block-size`. The draft's block size of 8 makes the verify window 9 tokens, and the KDA ReplaySSM ring must be a power of two at least twice the window — the recipe therefore pins `--linear-replayssm-cache-len 32`, since the 16 default fails startup validation. Validated on BF16, 4×B200 TP4, thinking on: full GSM8K 96.66% with a 99.77% stop rate (NEXTN on the same setup: 96.44% / 99.62%); per-cell speed and accuracy numbers sit under the B200 low-latency DSPARK cell in the Deploy panel.
 - Native context is 256K; SGLang reads it from the checkpoint's `max_position_embeddings`, so no `--context-length` flag is needed.
 - The **HiCache** card in the [Playground](#playground) exposes the validated Mooncake L3 path. It adds the hybrid-KDA scheduler and prefix-key settings together; see §3.3 for the required services.
 

docs/src/snippets/_playground.jsx

@@ -825,6 +825,8 @@ export const Playground = ({ config }) => {
               || head === "--speculative-eagle-topk"
               || head === "--speculative-num-draft-tokens"
               || head === "--speculative-dspark-block-size"
+              || head === "--enable-linear-replayssm-spec"
+              || head === "--linear-replayssm-cache-len"
               || head === "--speculative-ngram-max-bfs-breadth";
         });
         if (baseSpec.length === 0) return "off";
@@ -850,7 +852,8 @@ export const Playground = ({ config }) => {
         flags = h.stripFlagsByFirstToken(flags, [
           "--speculative-algorithm", "--speculative-num-steps",
           "--speculative-eagle-topk", "--speculative-num-draft-tokens",
-          "--speculative-dspark-block-size",
+          "--speculative-dspark-block-size", "--enable-linear-replayssm-spec",
+          "--linear-replayssm-cache-len",
           "--speculative-ngram-max-bfs-breadth",
         ]);
         const preset = (fc.options || []).find((p) => p.id === value);

docs/src/snippets/configs/inclusionAI/ling-3.0-flash-benchmarks.jsx

@@ -12,14 +12,14 @@ export const benchmarks = [
   // H200 + BF16 (TP4)
   // ====================================================================
   {
-    match: { hw: "h200", variant: "default", quant: "bf16", strategy: "low-latency", nodes: "single" },
+    match: { hw: "h200", variant: "default", quant: "bf16", strategy: "low-latency", spec: "nextn", nodes: "single" },
     sglang_version: "PR #33561 @ c5071ded",
     accuracy: { gsm8k_pct: 96.59 },
   },
   {
     // Rejected by the full GSM8K gate at request 1319: a no-EOS runaway generated
     // >33k tokens. Recipe stays `verified: false` in ling-3.0-flash.jsx.
-    match: { hw: "h200", variant: "default", quant: "bf16", strategy: "high-throughput", nodes: "single" },
+    match: { hw: "h200", variant: "default", quant: "bf16", strategy: "high-throughput", spec: "off", nodes: "single" },
     accuracy: { gsm8k_pct: null },
     notes: "Full GSM8K gate did not complete: one request ran away without emitting EOS (>33k generated tokens).",
   },
@@ -28,17 +28,17 @@ export const benchmarks = [
   // H200 + FP8 (TP4 + EP4)
   // ====================================================================
   {
-    match: { hw: "h200", variant: "default", quant: "fp8", strategy: "low-latency", nodes: "single" },
+    match: { hw: "h200", variant: "default", quant: "fp8", strategy: "low-latency", spec: "nextn", nodes: "single" },
     sglang_version: "PR #33561 @ e57e030b",
     accuracy: { gsm8k_pct: 95.83 },
   },
   {
-    match: { hw: "h200", variant: "default", quant: "fp8", strategy: "high-throughput", nodes: "single" },
+    match: { hw: "h200", variant: "default", quant: "fp8", strategy: "high-throughput", spec: "off", nodes: "single" },
     sglang_version: "PR #33561 @ e57e030b",
     accuracy: { gsm8k_pct: 96.51 },
   },
   {
-    match: { hw: "h200", variant: "default", quant: "int4", strategy: "high-throughput", nodes: "single" },
+    match: { hw: "h200", variant: "default", quant: "int4", strategy: "high-throughput", spec: "off", nodes: "single" },
     sglang_version: "PR #33561 @ e1a24a18",
     speed: [
       { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 1 },
@@ -50,7 +50,7 @@ export const benchmarks = [
     notes: "Full GSM8K stop rate 100%; default decode CUDA Graph captured 32 shapes through batch 218.",
   },
   {
-    match: { hw: "h200", variant: "default", quant: "mxfp4", strategy: "high-throughput", nodes: "single" },
+    match: { hw: "h200", variant: "default", quant: "mxfp4", strategy: "high-throughput", spec: "off", nodes: "single" },
     sglang_version: "PR #33561 @ e1a24a18",
     speed: [
       { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 1 },
@@ -66,7 +66,7 @@ export const benchmarks = [
   // H200 + HiCache (Mooncake tiered cache)
   // ====================================================================
   {
-    match: { hw: "h200", variant: "default", quant: "bf16", strategy: "hicache", nodes: "single" },
+    match: { hw: "h200", variant: "default", quant: "bf16", strategy: "hicache", spec: "nextn", nodes: "single" },
     sglang_version: "PR #33561 @ 51bcd89c",
     accuracy: { gsm8k_pct: 96.44 },
   },
@@ -75,12 +75,30 @@ export const benchmarks = [
   // B200 + BF16 (TP4)
   // ====================================================================
   {
-    match: { hw: "b200", variant: "default", quant: "bf16", strategy: "low-latency", nodes: "single" },
+    match: { hw: "b200", variant: "default", quant: "bf16", strategy: "low-latency", spec: "nextn", nodes: "single" },
-    sglang_version: "PR #33561 @ c5071ded",
+    sglang_version: "PR #33561 @ 0e5e40d8f (dev-Ling-3.0-flash image)",
-    accuracy: { gsm8k_pct: 96.44 },
+    speed: [
+      { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 1 },
+        ttft_ms: 171.57, tpot_ms: 2.34, tokens_per_sec_per_gpu: 886 },
+      { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 16 },
+        ttft_ms: 248.30, tpot_ms: 5.29, tokens_per_sec_per_gpu: 5936 },
+    ],
+    accuracy: { gsm8k_pct: 96.66 },
   },
   {
-    match: { hw: "b200", variant: "default", quant: "bf16", strategy: "high-throughput", nodes: "single" },
+    match: { hw: "b200", variant: "default", quant: "bf16", strategy: "low-latency", spec: "dspark", nodes: "single" },
+    sglang_version: "PR #33561 @ 0e5e40d8f (dev-Ling-3.0-flash image)",
+    speed: [
+      { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 1 },
+        ttft_ms: 156.44, tpot_ms: 0.78, tokens_per_sec_per_gpu: 1902 },
+      { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 16 },
+        ttft_ms: 247.27, tpot_ms: 2.87, tokens_per_sec_per_gpu: 8025 },
+    ],
+    accuracy: { gsm8k_pct: 96.59 },
+    notes: "Average accept length 6.22 (200q GSM8K window); deterministic 20-request stop rate 100%.",
+  },
+  {
+    match: { hw: "b200", variant: "default", quant: "bf16", strategy: "high-throughput", spec: "off", nodes: "single" },
     sglang_version: "PR #33561 @ c5071ded",
     accuracy: { gsm8k_pct: 96.51 },
   },
@@ -89,17 +107,17 @@ export const benchmarks = [
   // B200 + FP8 (TP4 + EP4)
   // ====================================================================
   {
-    match: { hw: "b200", variant: "default", quant: "fp8", strategy: "low-latency", nodes: "single" },
+    match: { hw: "b200", variant: "default", quant: "fp8", strategy: "low-latency", spec: "nextn", nodes: "single" },
     sglang_version: "PR #33561 @ e57e030b",
     accuracy: { gsm8k_pct: 96.59 },
   },
   {
-    match: { hw: "b200", variant: "default", quant: "fp8", strategy: "high-throughput", nodes: "single" },
+    match: { hw: "b200", variant: "default", quant: "fp8", strategy: "high-throughput", spec: "off", nodes: "single" },
     sglang_version: "PR #33561 @ e57e030b",
     accuracy: { gsm8k_pct: 97.04 },
   },
   {
-    match: { hw: "b200", variant: "default", quant: "int4", strategy: "high-throughput", nodes: "single" },
+    match: { hw: "b200", variant: "default", quant: "int4", strategy: "high-throughput", spec: "off", nodes: "single" },
     sglang_version: "PR #33561 @ e1a24a18",
     speed: [
       { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 1 },
@@ -111,7 +129,7 @@ export const benchmarks = [
     notes: "Full GSM8K stop rate 99.70%; default decode CUDA Graph captured 40 shapes through batch 305.",
   },
   {
-    match: { hw: "b200", variant: "default", quant: "mxfp4", strategy: "high-throughput", nodes: "single" },
+    match: { hw: "b200", variant: "default", quant: "mxfp4", strategy: "high-throughput", spec: "off", nodes: "single" },
     sglang_version: "PR #33561 @ e1a24a18",
     speed: [
       { workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 1 },
@@ -129,14 +147,14 @@ export const benchmarks = [
   // TODO: both cells are `verified: true` (gated on the final head) but the GSM8K
   // percentages were not recorded in the PR body or in the verifying commits —
   // fill from the run logs.
-  { match: { hw: "gb300", variant: "default", quant: "bf16", strategy: "low-latency", nodes: "single" } },
+  { match: { hw: "gb300", variant: "default", quant: "bf16", strategy: "low-latency", spec: "nextn", nodes: "single" } },
-  { match: { hw: "gb300", variant: "default", quant: "bf16", strategy: "high-throughput", nodes: "single" } },
+  { match: { hw: "gb300", variant: "default", quant: "bf16", strategy: "high-throughput", spec: "off", nodes: "single" } },
 
   // ====================================================================
   // GB300 + FP8 (TP4 + EP4)
   // ====================================================================
   // TODO: both cells are `verified: true` on the final head; the 96.66% / 96.44%
   // pair in the PR body predates the TP+EP change — fill with the re-measured values.
-  { match: { hw: "gb300", variant: "default", quant: "fp8", strategy: "low-latency", nodes: "single" } },
+  { match: { hw: "gb300", variant: "default", quant: "fp8", strategy: "low-latency", spec: "nextn", nodes: "single" } },
-  { match: { hw: "gb300", variant: "default", quant: "fp8", strategy: "high-throughput", nodes: "single" } },
+  { match: { hw: "gb300", variant: "default", quant: "fp8", strategy: "high-throughput", spec: "off", nodes: "single" } },
 ];

docs/src/snippets/configs/inclusionAI/ling-3.0-flash.jsx

@@ -4,9 +4,7 @@ export const config = {
   supportedHardware: ["h20-3e", "h200", "h800", "h100", "b200", "gb300"],
   groupHardware: false,
 
-  variants: [
+  variants: [{ id: "default", label: "Ling-3.0-flash" }],
-    { id: "default", label: "Ling-3.0-flash" },
-  ],
   quantizations: [
     { id: "bf16", label: "BF16" },
     { id: "fp8", label: "FP8" },
@@ -18,9 +16,7 @@ export const config = {
     { id: "high-throughput", label: "High-Throughput" },
     { id: "hicache", label: "HiCache + Mooncake" },
   ],
-  nodesOptions: [
+  nodesOptions: [{ id: "single", label: "Single Node" }],
-    { id: "single", label: "Single Node" },
-  ],
 
   modelNames: {
     "default|bf16": "inclusionAI/Ling-3.0-flash",
@@ -98,6 +94,20 @@ sgl-eval run gsm8k \\
         { id: "current", label: "Inherited from base" },
         { id: "off", label: "Off (greedy)" },
         { id: "nextn", label: "NEXTN (built-in MTP)", flags: ["--speculative-algorithm NEXTN"] },
+        {
+          id: "dspark",
+          label: "DSPARK (draft model)",
+          // --linear-replayssm-cache-len 32: the draft's block size 8 makes the
+          // verify window 9 tokens, and the KDA ReplaySSM ring must be a power
+          // of two >= 2x the window — the 16 default is too small and the
+          // server refuses to start.
+          flags: [
+            "--speculative-algorithm DSPARK",
+            "--speculative-draft-model-path inclusionAI/Ling-3.0-flash-dspark",
+            "--enable-linear-replayssm-spec",
+            "--linear-replayssm-cache-len 32",
+          ],
+        },
       ],
     },
     hicache: {
@@ -126,9 +136,60 @@ sgl-eval run gsm8k \\
     },
   },
 
-  cells: [
+  matchDims: [
+    { id: "variant", title: "Model Variant", options: [{ id: "default", label: "Ling-3.0-flash" }] },
     {
-      match: { hw: "h20-3e", variant: "default", quant: "bf16", strategy: "low-latency", nodes: "single" },
+      id: "quant",
+      title: "Quantization",
+      options: [
+        { id: "bf16", label: "BF16" },
+        { id: "fp8", label: "FP8" },
+        { id: "int4", label: "INT4" },
+        { id: "mxfp4", label: "MXFP4" },
+      ],
+    },
+    {
+      id: "strategy",
+      title: "Strategy",
+      options: [
+        { id: "low-latency", label: "Low-Latency" },
+        { id: "high-throughput", label: "High-Throughput" },
+        { id: "hicache", label: "HiCache + Mooncake" },
+      ],
+    },
+    {
+      id: "spec",
+      title: "Spec Decode",
+      options: [
+        { id: "nextn", label: "NEXTN (built-in MTP)" },
+        { id: "dspark", label: "DSPARK (draft model)" },
+        { id: "off", label: "Off (greedy)" },
+      ],
+    },
+    { id: "nodes", title: "Nodes", options: [{ id: "single", label: "Single Node" }] },
+  ],
+
+  // DSPARK twins of the low-latency cells: same shape, NEXTN swapped for the
+  // external-draft path. --linear-replayssm-cache-len 32 because the draft's
+  // block size 8 needs a power-of-two ring >= 2x the 9-token verify window.
+  // Everything lives inside this IIFE because Mintlify's snippet compiler only
+  // evaluates the exported expression — top-level module code is dropped.
+  cells: (() => {
+    const DSPARK_FLAGS = [
+      "--speculative-algorithm DSPARK",
+      "--speculative-draft-model-path inclusionAI/Ling-3.0-flash-dspark",
+      "--enable-linear-replayssm-spec",
+      "--linear-replayssm-cache-len 32",
+    ];
+    const dsparkTwin = (cell, verified) => ({
+      ...cell,
+      verified,
+      match: { ...cell.match, spec: "dspark" },
+      flags: cell.flags.flatMap((f) => (f === "--speculative-algorithm NEXTN" ? DSPARK_FLAGS : [f])),
+    });
+    const lowLatencyCells = [
+    {
+      match: { hw: "h20-3e", variant: "default", quant: "bf16", strategy: "low-latency", spec: "nextn", nodes: "single" },
       verified: false,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -140,7 +201,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "h200", variant: "default", quant: "bf16", strategy: "low-latency", nodes: "single" },
+      match: { hw: "h200", variant: "default", quant: "bf16", strategy: "low-latency", spec: "nextn", nodes: "single" },
       verified: true,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -152,7 +213,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "h800", variant: "default", quant: "bf16", strategy: "low-latency", nodes: "single" },
+      match: { hw: "h800", variant: "default", quant: "bf16", strategy: "low-latency", spec: "nextn", nodes: "single" },
       verified: false,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -164,7 +225,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "h100", variant: "default", quant: "bf16", strategy: "low-latency", nodes: "single" },
+      match: { hw: "h100", variant: "default", quant: "bf16", strategy: "low-latency", spec: "nextn", nodes: "single" },
       verified: false,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -176,7 +237,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "b200", variant: "default", quant: "bf16", strategy: "low-latency", nodes: "single" },
+      match: { hw: "b200", variant: "default", quant: "bf16", strategy: "low-latency", spec: "nextn", nodes: "single" },
       verified: true,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -188,7 +249,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "gb300", variant: "default", quant: "bf16", strategy: "low-latency", nodes: "single" },
+      match: { hw: "gb300", variant: "default", quant: "bf16", strategy: "low-latency", spec: "nextn", nodes: "single" },
       verified: true,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -200,7 +261,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "h20-3e", variant: "default", quant: "fp8", strategy: "low-latency", nodes: "single" },
+      match: { hw: "h20-3e", variant: "default", quant: "fp8", strategy: "low-latency", spec: "nextn", nodes: "single" },
       verified: false,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -213,7 +274,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "h200", variant: "default", quant: "fp8", strategy: "low-latency", nodes: "single" },
+      match: { hw: "h200", variant: "default", quant: "fp8", strategy: "low-latency", spec: "nextn", nodes: "single" },
       verified: true,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -226,7 +287,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "h800", variant: "default", quant: "fp8", strategy: "low-latency", nodes: "single" },
+      match: { hw: "h800", variant: "default", quant: "fp8", strategy: "low-latency", spec: "nextn", nodes: "single" },
       verified: false,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -239,7 +300,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "h100", variant: "default", quant: "fp8", strategy: "low-latency", nodes: "single" },
+      match: { hw: "h100", variant: "default", quant: "fp8", strategy: "low-latency", spec: "nextn", nodes: "single" },
       verified: false,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -252,7 +313,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "b200", variant: "default", quant: "fp8", strategy: "low-latency", nodes: "single" },
+      match: { hw: "b200", variant: "default", quant: "fp8", strategy: "low-latency", spec: "nextn", nodes: "single" },
       verified: true,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -265,7 +326,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "gb300", variant: "default", quant: "fp8", strategy: "low-latency", nodes: "single" },
+      match: { hw: "gb300", variant: "default", quant: "fp8", strategy: "low-latency", spec: "nextn", nodes: "single" },
       verified: true,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -277,8 +338,14 @@ sgl-eval run gsm8k \\
         "--port {{PORT}}",
       ],
     },
+];
+    return [
+      ...lowLatencyCells.flatMap((c) => [
+        c,
+        dsparkTwin(c, c.match.hw === "b200" && c.match.quant === "bf16"),
+      ]),
     {
-      match: { hw: "h20-3e", variant: "default", quant: "bf16", strategy: "high-throughput", nodes: "single" },
+      match: { hw: "h20-3e", variant: "default", quant: "bf16", strategy: "high-throughput", spec: "off", nodes: "single" },
       verified: false,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -289,7 +356,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "h200", variant: "default", quant: "bf16", strategy: "high-throughput", nodes: "single" },
+      match: { hw: "h200", variant: "default", quant: "bf16", strategy: "high-throughput", spec: "off", nodes: "single" },
       verified: true,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -300,7 +367,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "h800", variant: "default", quant: "bf16", strategy: "high-throughput", nodes: "single" },
+      match: { hw: "h800", variant: "default", quant: "bf16", strategy: "high-throughput", spec: "off", nodes: "single" },
       verified: false,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -311,7 +378,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "h100", variant: "default", quant: "bf16", strategy: "high-throughput", nodes: "single" },
+      match: { hw: "h100", variant: "default", quant: "bf16", strategy: "high-throughput", spec: "off", nodes: "single" },
       verified: false,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -322,7 +389,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "b200", variant: "default", quant: "bf16", strategy: "high-throughput", nodes: "single" },
+      match: { hw: "b200", variant: "default", quant: "bf16", strategy: "high-throughput", spec: "off", nodes: "single" },
       verified: true,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -333,7 +400,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "gb300", variant: "default", quant: "bf16", strategy: "high-throughput", nodes: "single" },
+      match: { hw: "gb300", variant: "default", quant: "bf16", strategy: "high-throughput", spec: "off", nodes: "single" },
       verified: true,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -344,7 +411,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "h20-3e", variant: "default", quant: "fp8", strategy: "high-throughput", nodes: "single" },
+      match: { hw: "h20-3e", variant: "default", quant: "fp8", strategy: "high-throughput", spec: "off", nodes: "single" },
       verified: false,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -356,7 +423,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "h200", variant: "default", quant: "fp8", strategy: "high-throughput", nodes: "single" },
+      match: { hw: "h200", variant: "default", quant: "fp8", strategy: "high-throughput", spec: "off", nodes: "single" },
       verified: true,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -368,7 +435,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "h800", variant: "default", quant: "fp8", strategy: "high-throughput", nodes: "single" },
+      match: { hw: "h800", variant: "default", quant: "fp8", strategy: "high-throughput", spec: "off", nodes: "single" },
       verified: false,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -380,7 +447,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "h100", variant: "default", quant: "fp8", strategy: "high-throughput", nodes: "single" },
+      match: { hw: "h100", variant: "default", quant: "fp8", strategy: "high-throughput", spec: "off", nodes: "single" },
       verified: false,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -392,7 +459,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "b200", variant: "default", quant: "fp8", strategy: "high-throughput", nodes: "single" },
+      match: { hw: "b200", variant: "default", quant: "fp8", strategy: "high-throughput", spec: "off", nodes: "single" },
       verified: true,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -404,7 +471,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "gb300", variant: "default", quant: "fp8", strategy: "high-throughput", nodes: "single" },
+      match: { hw: "gb300", variant: "default", quant: "fp8", strategy: "high-throughput", spec: "off", nodes: "single" },
       verified: true,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -416,7 +483,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "h200", variant: "default", quant: "int4", strategy: "high-throughput", nodes: "single" },
+      match: { hw: "h200", variant: "default", quant: "int4", strategy: "high-throughput", spec: "off", nodes: "single" },
       verified: true,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -429,7 +496,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "b200", variant: "default", quant: "int4", strategy: "high-throughput", nodes: "single" },
+      match: { hw: "b200", variant: "default", quant: "int4", strategy: "high-throughput", spec: "off", nodes: "single" },
       verified: true,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -442,7 +509,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "h200", variant: "default", quant: "mxfp4", strategy: "high-throughput", nodes: "single" },
+      match: { hw: "h200", variant: "default", quant: "mxfp4", strategy: "high-throughput", spec: "off", nodes: "single" },
       verified: true,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -456,7 +523,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "b200", variant: "default", quant: "mxfp4", strategy: "high-throughput", nodes: "single" },
+      match: { hw: "b200", variant: "default", quant: "mxfp4", strategy: "high-throughput", spec: "off", nodes: "single" },
       verified: true,
       flags: [
         "--model-path {{MODEL_NAME}}",
@@ -474,7 +541,7 @@ sgl-eval run gsm8k \\
     // Hybrid KDA must pass prefix keys to Mooncake; otherwise storage writes are empty.
     // Cold uncached extends above chunked_prefill_size skip write-through for that influx.
     {
-      match: { hw: "h200", variant: "default", quant: "bf16", strategy: "hicache", nodes: "single" },
+      match: { hw: "h200", variant: "default", quant: "bf16", strategy: "hicache", spec: "nextn", nodes: "single" },
       verified: true,
       env: [
         "MOONCAKE_MASTER={{MOONCAKE_MASTER}}",
@@ -502,7 +569,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "h200", variant: "default", quant: "fp8", strategy: "hicache", nodes: "single" },
+      match: { hw: "h200", variant: "default", quant: "fp8", strategy: "hicache", spec: "nextn", nodes: "single" },
       verified: false,
       env: [
         "MOONCAKE_MASTER={{MOONCAKE_MASTER}}",
@@ -531,7 +598,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "gb300", variant: "default", quant: "bf16", strategy: "hicache", nodes: "single" },
+      match: { hw: "gb300", variant: "default", quant: "bf16", strategy: "hicache", spec: "nextn", nodes: "single" },
       verified: false,
       env: [
         "MOONCAKE_MASTER={{MOONCAKE_MASTER}}",
@@ -559,7 +626,7 @@ sgl-eval run gsm8k \\
       ],
     },
     {
-      match: { hw: "gb300", variant: "default", quant: "fp8", strategy: "hicache", nodes: "single" },
+      match: { hw: "gb300", variant: "default", quant: "fp8", strategy: "hicache", spec: "nextn", nodes: "single" },
       verified: false,
       env: [
         "MOONCAKE_MASTER={{MOONCAKE_MASTER}}",
@@ -587,5 +654,6 @@ sgl-eval run gsm8k \\
         "--port {{PORT}}",
       ],
     },
-  ],
+    ];
+  })(),
 };