From 861eca8e2519e64a7d56ac1a947392e83873ee1c Mon Sep 17 00:00:00 2001
From: Yuhao Yang <47235274+yhyang201@users.noreply.github.com>
Date: Tue, 18 Aug 2026 02:01:47 +0800
Subject: [PATCH] docs: add NVFP4 quantization option to Kimi-K3 deploy panel
 (#35168)

---
 .../autoregressive/Moonshotai/Kimi-K3.mdx     |  2 +-
 .../snippets/configs/moonshotai/kimi-k3.jsx   | 40 +++++++++++++++++++
 2 files changed, 41 insertions(+), 1 deletion(-)

diff --git a/docs/cookbook/autoregressive/Moonshotai/Kimi-K3.mdx b/docs/cookbook/autoregressive/Moonshotai/Kimi-K3.mdx
index 6297569bf..c236bf5f7 100644
--- a/docs/cookbook/autoregressive/Moonshotai/Kimi-K3.mdx
+++ b/docs/cookbook/autoregressive/Moonshotai/Kimi-K3.mdx
@@ -40,7 +40,7 @@ For how to launch the image, see [Install → Method 3: Using Docker](../../../d
 
 </Accordion>
 
-Pick your hardware, then the deployment shape and operating point. Node count follows the hardware recipe (B200 2×8, GB200 4×4, H100 4×8, B300 1×8, H200 2×8 — 4×8 on Unified High-Throughput, GB300 2×4, MI350X/MI355X 1×8), so it is not a separate choice.
+Pick your hardware, then the deployment shape and operating point. Node count follows the hardware recipe (B200 2×8, GB200 4×4, H100 4×8, B300 1×8, H200 2×8 — 4×8 on Unified High-Throughput, GB300 2×4, MI350X/MI355X 1×8), so it is not a separate choice. If you serve the NVFP4 checkpoint (`nvidia/Kimi-K3-NVFP4`, the **Quantization** row in the panel below), use the `lmsysorg/sglang:dev-dev-kimi-k3-nvfp4` image.
 
 **PD Mode** — `Unified` serves prefill and decode together. `Prefill` / `Decode` split them into dedicated pools (see [PD disaggregation](#3-4-pd-disaggregation)); `Prefill` ships two strategies, both chunked at 16k. On the 8-GPU platforms (B300 1×8, GB300 2×4), `Default` is TP8 and `Long-Context` is `--pp-size 8 --tp-size 1`. On the 16-GPU platforms (B200 2×8, GB200 4×4), both are `--pp-size 16 --tp-size 1` and differ only in `--mem-fraction-static` (0.85 vs 0.90) — deep PP is the throughput shape there, not just the long-context one (see [Deep PP](#deep-pp-for-prefill)).
 
diff --git a/docs/src/snippets/configs/moonshotai/kimi-k3.jsx b/docs/src/snippets/configs/moonshotai/kimi-k3.jsx
index 309c86988..72783e08d 100644
--- a/docs/src/snippets/configs/moonshotai/kimi-k3.jsx
+++ b/docs/src/snippets/configs/moonshotai/kimi-k3.jsx
@@ -160,6 +160,39 @@ export const config = {
   // Orthogonal to the cell grid: the picked option layers flags onto whichever
   // cell is showing, so turning speculation on does not triple the cell count.
   overlayDims: [
+    {
+      // Checkpoint choice, orthogonal to the cell grid: MXFP4 is the shipping
+      // default, NVFP4 is NVIDIA's ModelOpt mixed checkpoint (NVFP4 SiTU routed
+      // experts + FP8_PB_WO 128x128 block-FP8 attention). NVFP4 swaps the model
+      // slug (modelNames) and pins the TRT-LLM MoE runner — the auto resolution
+      // never engages TRT-LLM deferred finalize and the NVFP4 MoE raises
+      // NotImplementedError at CUDA-graph capture, while flashinfer_cutlass has
+      // no SiTU kernel. The DSPARK overlay needs no change: the same draft
+      // checkpoint serves on top of the NVFP4 base.
+      id: "quant",
+      title: "Quantization",
+      default: "mxfp4",
+      options: [
+        { id: "mxfp4", label: "MXFP4", subtitle: "Moonshot AI checkpoint" },
+        {
+          id: "nvfp4",
+          label: "NVFP4",
+          subtitle: "NVIDIA checkpoint",
+          // The NVFP4 MoE kernels (FlashInfer TRT-LLM) are Blackwell-only.
+          disabled: (s) => !["b200", "gb200", "b300", "gb300"].includes(s.hw),
+          disableReason:
+            "The nvidia/Kimi-K3-NVFP4 checkpoint needs Blackwell: its routed experts run on FlashInfer TRT-LLM NVFP4 kernels (SiTU), which do not exist for Hopper or AMD.",
+          // B200's Balanced/High-Throughput cells pin flashinfer_mxfp4; Hopper
+          // pins marlin (unreachable here — NVFP4 is Blackwell-gated). Replace
+          // whatever the cell pins with the one working NVFP4 runner.
+          stripPrefixes: ["--moe-runner-backend"],
+          flags: ["--moe-runner-backend flashinfer_trtllm"],
+          hints: [
+            "Use docker image lmsysorg/sglang:dev-dev-kimi-k3-nvfp4 (CUDA 13).",
+          ],
+        },
+      ],
+    },
     {
       id: "mmTransport",
       title: "VLM Transport",
@@ -333,6 +366,7 @@ export const config = {
 
   modelNames: {
     default: "moonshotai/Kimi-K3",
+    nvfp4: "nvidia/Kimi-K3-NVFP4",
   },
 
   placeholders: {
@@ -379,6 +413,12 @@ export const config = {
     gb200:  "lmsysorg/sglang:kimi-k3",
     mi350x: "lmsysorg/sglang-rocm:rocm720-mi35x-k3-20260727",
     mi355x: "lmsysorg/sglang-rocm:rocm720-mi35x-k3-20260727",
+    // NVFP4 needs a build with sgl-project/sglang#35077; the purpose-built dev
+    // image is cut from that PR's head (CUDA 13).
+    "b300|nvfp4":  "lmsysorg/sglang:dev-dev-kimi-k3-nvfp4",
+    "gb300|nvfp4": "lmsysorg/sglang:dev-dev-kimi-k3-nvfp4",
+    "b200|nvfp4":  "lmsysorg/sglang:dev-dev-kimi-k3-nvfp4",
+    "gb200|nvfp4": "lmsysorg/sglang:dev-dev-kimi-k3-nvfp4",
   },
   // Pre-selects the issue template's `model` field on "Submit verified cell".
   github: {