[NPU] [DOC] Add kimi k3 cookbook for 950PR/DT Series (#40575)

2026-09-21 20:18:03 +08:00
parent 0f6761b54f
commit b410010087
5 changed files with 311 additions and 84 deletions
@@ -40,14 +40,16 @@ For how to launch the image, see [Install → Method 3: Using Docker](../../../d
 <Tab title="NPU">

 ```bash Command
-docker pull quay.io/ascend/sglang:main-cann9.0.0-a3
+docker pull quay.io/ascend/sglang:main-cann9.0.0-a3   # Ascend A3 Series
+docker pull swr.cn-southwest-2.myhuaweicloud.com/base_image/dockerhub/lmsysorg/sglang:cann9.1.0-950-B070   # Ascend 950PR/DT Series
 ```

 For host and platform setup, see the
 [NPU installation guide](../../../docs/hardware-platforms/ascend-npus/getting-started/installation) and the
 [quick start guide](../../../docs/hardware-platforms/ascend-npus/getting-started/quick_start).

-**Weights (NPU):** [Kimi-K3-W4A8](https://www.modelscope.cn/models/sgl-npu/Kimi-K3-W4A8) (W4A8, 1.49 TB) ·
+**Weights (NPU):** [Kimi-K3](https://modelscope.cn/models/moonshotai/Kimi-K3) (950PR/DT Series, MXFP4) ·
+[Kimi-K3-W4A8](https://www.modelscope.cn/models/sgl-npu/Kimi-K3-W4A8) (A3 Series, W4A8, 1.49 TB) ·
 [Kimi-K3-DSpark](https://www.modelscope.cn/models/RadixArk/Kimi-K3-DSpark) (DSPARK draft, 4.5 GB)

 </Tab>
@@ -56,7 +58,7 @@ For host and platform setup, see the

 </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, Ascend A3 Series 4×8 — 32 cards / 64 dies), 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.
+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, Ascend A3 Series 4×8 — 32 cards / 64 dies, Ascend 950PR/DT Series 4×8), 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)).

@@ -71,7 +73,7 @@ Pick your hardware, then the deployment shape and operating point. Node count fo
 **Spec Decode** — layers onto the strategy without changing it, on every platform except B200. DSPARK proposes 7 draft tokens per step (tune in the Playground) and requires `pp_size == 1`, so on B200 it also drops the pipeline and re-lays the same 16 GPUs flat: PP2 × TP8 → TP16, PP2 × DCPEP8 → DCPEP16. DFLASH has no published draft checkpoint. The win is largest on short interactive traffic and fades as the prompt grows.

 <Note>
-`--mamba-full-memory-ratio` is the one sizing flag, computed live: set your average request length in the [Mamba ratio calculator](#mamba-ratio-calculator); everything else follows the panels, and the result is pinned into the command. (The Ascend A3 Series uses `--max-mamba-cache-size` instead.)
+`--mamba-full-memory-ratio` is the one sizing flag, computed live: set your average request length in the [Mamba ratio calculator](#mamba-ratio-calculator); everything else follows the panels, and the result is pinned into the command. (The Ascend NPU recipes don't use the calculator and set no ratio at all — the Ascend path sizes both pools itself: the A3 Series pins `--max-mamba-cache-size` instead, and the A5 recipe serves with the radix cache off.)
 </Note>

 import { Deployment } from "/src/snippets/_deployment.jsx";
@@ -146,7 +148,7 @@ not been re-measured on any cell — re-measure before you rely on one.

 ## 2. Configuration Tips

-**Memory: two pools, one flag.** K3 splits static memory into a worst-case-reserved **KDA state pool** (it sets the concurrency ceiling) and a paged **MLA KV pool**, divided by `--mamba-full-memory-ratio`. The command panel pins that flag to the [calculator](#mamba-ratio-calculator)'s output — set your average request length there; every other calculator input follows the panels. (On the Ascend A3 Series: `--max-mamba-cache-size`, no calculator.) After boot, read back `max_total_num_tokens` (the KV side) and the admitted-request cap (the state side).
+**Memory: two pools, one flag.** K3 splits static memory into a worst-case-reserved **KDA state pool** (it sets the concurrency ceiling) and a paged **MLA KV pool**, divided by `--mamba-full-memory-ratio`. The command panel pins that flag to the [calculator](#mamba-ratio-calculator)'s output — set your average request length there; every other calculator input follows the panels. (On the Ascend NPU recipes there is no calculator and no ratio flag: the Ascend path sizes both pools itself — the A3 Series pins `--max-mamba-cache-size`, and the A5 recipe runs the radix cache off.) After boot, read back `max_total_num_tokens` (the KV side) and the admitted-request cap (the state side).

 Capacity levers, all in the Playground. Each trades precision or cache behavior for capacity — re-verify accuracy on your workload:

@@ -180,6 +182,7 @@ Speculation: DSPARK holds block size + 1 (= 8) intermediate states per request
 | H100 4×8 | TP32/EP32, Marlin + FlashMLA | SM90a build of the K3 image; pin NCCL/Gloo to the same NIC on all nodes; least post-weight headroom (80 GB) |
 | MI350X/MI355X 1×8 | TP8 ROCm/AITER | AITER A8W4 FlyDSL MoE, Triton attention (`SGLANG_MLA_DECODE_TUNE=1` for gfx950 MLA decode geometry), graph bs up to 256, fp8 kvcache; DSPARK supported. Activation-quant and fused-KDA-decode knobs: [AMD ROCm/AITER environment](#amd-env) |
 | Ascend A3 Series 4×8 (32 cards / 64 dies) | TP64/DP4 + DeepEP | PD-mixed `Unified` only; DSPARK baked in; pin `GLOO`/`HCCL_SOCKET_IFNAME` on every node |
+| Ascend 950PR/DT Series 4×8 | TP32/dp1 + DeepEP | PD-mixed `Unified` only; DSPARK baked in; shared experts / dense MLP shard over attention-TP (`--shared-experts-tp-size 4`); radix cache off; pin `GLOO`/`HCCL_SOCKET_IFNAME` on every node |

 **DCP notes** — the DCP cells are Balanced and High-Throughput on every Blackwell platform, in both the `Unified` and `Decode` roles:

@@ -239,7 +242,7 @@ Pending update...

 ### 3.2 Tool Calling

-Enable the `kimi_k3` tool-call parser (toggle **Tool Call Parser** in the **Parsers** card of the [Playground above](#playground)) to surface structured tool calls via `message.tool_calls`. Because K3 is a thinking model, the follow-up turn may put text in `reasoning_content` as well as `content` — print both. (Not yet supported on the Ascend A3 Series.)
+Enable the `kimi_k3` tool-call parser (toggle **Tool Call Parser** in the **Parsers** card of the [Playground above](#playground)) to surface structured tool calls via `message.tool_calls`. Because K3 is a thinking model, the follow-up turn may put text in `reasoning_content` as well as `content` — print both. (Not yet supported on the Ascend NPU recipes — A3 and A5.)

 <Accordion title="Tool Calling Example (Python)">

@@ -147,10 +147,14 @@ export const Deployment = ({ config, benchmarks }) => {
      { id: "mi355x", label: "MI355X", vram: "288GB",
        multiNodeDockerFlags: [...AMD_RDMA_DOCKER_FLAGS] },
    ],
-    // Ascend A3 Series: 1 card = 2 dies, so --tp-size is 2× the card
-    // count (32 cards -> --tp-size 64).
+    // Ascend device layout: one /dev/davinciN per core. An A3 Series card is
+    // the exception — 2 dies per card, so an 8-card node exposes 16 devices
+    // and --tp-size is twice the card count. A 950PR/DT Series card is a
+    // single core, so the device count and --tp-size follow the cards. Both
+    // counts feed the docker `--device` list (`npuDevices`).
    npu: [
-      { id: "a3", label: "Ascend A3 Series", vram: "64GB/die" },
+      { id: "a3", label: "A3 Series",        vram: "64GB/die", npuDevices: 16 },
+      { id: "a5", label: "950PR/DT Series",  vram: "128GB",    npuDevices: 8  },
    ],
  };

@@ -819,14 +823,31 @@ export const Deployment = ({ config, benchmarks }) => {
        return (extra && extra.vendor) || "nvidia";
      };
      // `config.hardware` overrides by id, as in buildHardwareGroups.
-      const fabricFlagsOf = (hwId) => {
+      const catalogEntryOf = (hwId) => {
        const extra = (config.hardware || []).find((h) => h.id === hwId);
-        if (extra) return extra.multiNodeDockerFlags || [];
+        if (extra) return extra;
        for (const list of Object.values(HARDWARE_CATALOG)) {
          const hit = list.find((h) => h.id === hwId);
-          if (hit) return hit.multiNodeDockerFlags || [];
+          if (hit) return hit;
        }
-        return [];
+        return null;
+      };
+      const fabricFlagsOf = (hwId) =>
+        (catalogEntryOf(hwId) || {}).multiNodeDockerFlags || [];
+      // NPU cards are reached with --device, one per /dev/davinciN core;
+      // `npuDevices` carries the per-product-line count (16 on an A3 Series
+      // node, 8 on a 950PR/DT Series node), four devices per line as the host
+      // docs show.
+      const davinciLines = (devices) => {
+        const lines = [];
+        for (let i = 0; i < devices; i += 4) {
+          const group = [];
+          for (let k = i; k < Math.min(i + 4, devices); k++) {
+            group.push(`--device=/dev/davinci${k}`);
+          }
+          lines.push("  " + group.join(" "));
+        }
+        return lines;
      };
      const gpuAccessLines = vendorOf(sel.hw) === "amd"
        ? [
@@ -838,14 +859,10 @@ export const Deployment = ({ config, benchmarks }) => {
          ]
        : vendorOf(sel.hw) === "npu"
        ? [
-            // NPU: --privileged grants the davinci devices (16 dies on an
-            // 8-card Ascend A3 Series node); the host CANN driver/firmware/state
-            // must be mounted in.
+            // NPU: --privileged grants the davinci devices; the host CANN
+            // driver/firmware/state must be mounted in.
            "docker run --privileged --shm-size=16g",
-            "  --device=/dev/davinci0 --device=/dev/davinci1 --device=/dev/davinci2 --device=/dev/davinci3",
-            "  --device=/dev/davinci4 --device=/dev/davinci5 --device=/dev/davinci6 --device=/dev/davinci7",
-            "  --device=/dev/davinci8 --device=/dev/davinci9 --device=/dev/davinci10 --device=/dev/davinci11",
-            "  --device=/dev/davinci12 --device=/dev/davinci13 --device=/dev/davinci14 --device=/dev/davinci15",
+            ...davinciLines((catalogEntryOf(sel.hw) || {}).npuDevices || 16),
            "  --device=/dev/davinci_manager",
            "  --device=/dev/hisi_hdc",
            "  -v /usr/local/sbin:/usr/local/sbin",
@@ -135,10 +135,18 @@ export const KimiK3MambaRatioCalculator = () => {
  const eff = derive(cfg.flags, cfg.env);
  const bs = derive(cfg.baseFlags.length ? cfg.baseFlags : cfg.flags,
                    cfg.baseFlags.length ? cfg.baseEnv : cfg.env);
-  // A --max-mamba-cache-size cell sizes the pool explicitly — no ratio to
-  // compute or broadcast.
-  const explicitSizing = (cfg.baseFlags.length ? cfg.baseFlags : cfg.flags)
-    .some((f) => f.startsWith("--max-mamba-cache-size"));
+  // Recipes that size the dual pool without the ratio neither render one nor
+  // broadcast one:
+  //   - a --max-mamba-cache-size cell pins the KDA slot count explicitly;
+  //   - the Ascend NPU recipes size both pools internally and never set
+  //     --mamba-full-memory-ratio (every NPU recipe carries --device npu).
+  const baseFlagList = cfg.baseFlags.length ? cfg.baseFlags : cfg.flags;
+  const explicitSizing = baseFlagList.some((f) => f.startsWith("--max-mamba-cache-size"));
+  const npuRecipe = baseFlagList.some((f) => {
+    const [head, ...rest] = f.trim().split(/[\s=]+/);
+    return head === "--device" && rest[0] === "npu";
+  });
+  const ratioNotApplicable = explicitSizing || npuRecipe;
  const { ratio, tp, dp, attnTp, dcp, kvDtype, ssmDtype, radixOff, strategy, skipLock, slots, specOn, replaySpec, block, pdRole } = eff;
  const valid = Number.isFinite(ratio) && ratio > 0 && length > 0 && 96 % attnTp === 0;
  const baseValid = Number.isFinite(bs.ratio) && bs.ratio > 0 && length > 0;
@@ -155,18 +163,22 @@ export const KimiK3MambaRatioCalculator = () => {
  const cliFlag = valid ? `--mamba-full-memory-ratio ${result}` : "";

  // Broadcast both results: the Deploy command takes the base-config value,
-  // the Playground's composed command takes the effective one.
+  // the Playground's composed command takes the effective one. On a recipe
+  // that does not use the ratio, broadcast nulls instead of skipping the
+  // dispatch — the panels must drop a ratio pinned for a recipe the reader
+  // has since left rather than keep injecting it.
  useEffect(() => {
-    if (explicitSizing) return;
    window.dispatchEvent(
      new CustomEvent("sglang-k3-mamba-ratio", {
-        detail: {
-          ratio: valid ? result : null,
-          baseRatio: baseValid ? baseResult : null,
-        },
+        detail: ratioNotApplicable
+          ? { ratio: null, baseRatio: null }
+          : {
+              ratio: valid ? result : null,
+              baseRatio: baseValid ? baseResult : null,
+            },
      })
    );
-  }, [result, valid, baseResult, baseValid, explicitSizing]);
+  }, [result, valid, baseResult, baseValid, ratioNotApplicable]);

  const copyFlag = () => {
    if (!cliFlag || typeof navigator === "undefined" || !navigator.clipboard) return;
@@ -235,10 +247,12 @@ export const KimiK3MambaRatioCalculator = () => {
    specLabel,
  ].filter(Boolean);

-  if (explicitSizing) {
+  if (ratioNotApplicable) {
    return (
      <div className="not-prose" style={{ padding: "14px", border: `1px solid ${colors.border}`, borderRadius: "8px", background: colors.panel, color: colors.muted, fontSize: "13px" }}>
-        This recipe sizes the KDA state pool explicitly with <code style={{ color: colors.text }}>--max-mamba-cache-size</code>, so the ratio calculator does not apply.
+        {explicitSizing
+          ? <>This recipe sizes the KDA state pool explicitly with <code style={{ color: colors.text }}>--max-mamba-cache-size</code>, so the ratio calculator does not apply.</>
+          : <>This recipe runs on Ascend NPUs, which size the KDA state pool and the MLA KV pool internally and never set <code style={{ color: colors.text }}>--mamba-full-memory-ratio</code>, so the ratio calculator does not apply.</>}
      </div>
    );
  }
@@ -1723,9 +1723,25 @@ export const Playground = ({ config }) => {
        mi355x: AMD_RDMA_DOCKER_FLAGS,
      };
      const fabricFlags = HW_MULTINODE_DOCKER_FLAGS[sel.hw] || [];
-      // Mirrors the vendor branch in _deployment.jsx: ROCm reaches its GPUs
-      // through /dev/kfd + /dev/dri and the video group, not --gpus all.
+      // Mirrors the vendor branches in _deployment.jsx: ROCm reaches its GPUs
+      // through /dev/kfd + /dev/dri and the video group (not --gpus all), and
+      // Ascend NPUs are reached with --device, one per /dev/davinciN core (16
+      // on an A3 Series node, 8 on a 950PR/DT Series node — the catalog's
+      // `npuDevices`).
      const isAmdHw = /^mi\d/.test(sel.hw || "");
+      const HW_NPU_DEVICES = { a3: 16, a5: 8 };
+      const npuDevices = HW_NPU_DEVICES[sel.hw];
+      const davinciLines = (devices) => {
+        const lines = [];
+        for (let i = 0; i < devices; i += 4) {
+          const group = [];
+          for (let k = i; k < Math.min(i + 4, devices); k++) {
+            group.push(`--device=/dev/davinci${k}`);
+          }
+          lines.push("  " + group.join(" "));
+        }
+        return lines;
+      };
      const dockerLines = [
        ...(isAmdHw
          ? [
@@ -1735,6 +1751,21 @@ export const Playground = ({ config }) => {
              "  --cap-add=SYS_PTRACE --security-opt seccomp=unconfined",
              "  --shm-size 32g",
            ]
+          : npuDevices
+          ? [
+              // NPU: --privileged grants the davinci devices; the host CANN
+              // driver/firmware/state must be mounted in.
+              "docker run --privileged --shm-size=16g",
+              ...davinciLines(npuDevices),
+              "  --device=/dev/davinci_manager",
+              "  --device=/dev/hisi_hdc",
+              "  -v /usr/local/sbin:/usr/local/sbin",
+              "  -v /usr/local/Ascend/driver:/usr/local/Ascend/driver",
+              "  -v /usr/local/Ascend/firmware:/usr/local/Ascend/firmware",
+              "  -v /etc/ascend_install.info:/etc/ascend_install.info",
+              "  -v /var/queue_schedule:/var/queue_schedule",
+              "  -v ~/.cache/:/root/.cache/",
+            ]
          : [
              "docker run --gpus all",
              "  --shm-size 32g",
@@ -1743,7 +1774,8 @@ export const Playground = ({ config }) => {
        // A PD pair is cross-host even when each role is a single-node cell, so
        // the RDMA fabric flags are needed for `pdMode` too, not just multinode.
        ...((multinode || pdMode) ? fabricFlags.map((x) => "  " + x) : []),
-        "  -v ~/.cache/huggingface:/root/.cache/huggingface",
+        // The NPU device block already mounts ~/.cache/.
+        ...(npuDevices ? [] : ["  -v ~/.cache/huggingface:/root/.cache/huggingface"]),
        ...(config.dockerMounts || []).map((mount) => `  -v ${mount}`),
        `  --env "HF_TOKEN={{HF_TOKEN}}"`,
        ...cellEnv.map((e) => `  --env ${e}`),
@@ -14,8 +14,10 @@ export const config = {
  // Low-Latency, PP2 × DCPEP8 on Balanced and High-Throughput, while DSPARK
  // re-lays the same 16 as flat TP16 / DCPEP16), GB200 (4×4 TP16 MNNVL),
  // H200 (2×8 TP16/EP16, or 4×8 TP32/EP32 for High-Throughput), H100
-  // (4×8 TP32/EP32), and MI350X/MI355X (1×8 TP8) have serving recipes.
-  supportedHardware: ["b300", "gb300", "b200", "gb200", "h200", "h100", "mi350x", "mi355x", "a3"],
+  // (4×8 TP32/EP32), MI350X/MI355X (1×8 TP8), Ascend A3 Series (4×8, TP64
+  // over 2-die cards), and Ascend 950PR/DT Series (4×8, TP32, one rank per
+  // card) have serving recipes.
+  supportedHardware: ["b300", "gb300", "b200", "gb200", "h200", "h100", "mi350x", "mi355x", "a3", "a5"],

  // ---- Cell introspection (config-internal; the engines ignore these keys) ----
  //
@@ -47,6 +49,15 @@ export const config = {
    const f = config.flagOf(cell, name);
    return f ? Number(f.split(/[\s=]/)[1]) || 1 : 1;
  },
+  // The two NPU recipes (A3 and A5) each ship exactly one operating point —
+  // Unified PD, the Balanced strategy, DSPARK, no HiCache, no tool calling —
+  // so every panel gate below keys off this helper instead of enumerating
+  // platforms. Shape differences between the two (TP64/DP4 vs TP32/dp1,
+  // Modelslim W4A8 vs the MXFP4 checkpoint, mamba-cache sizing) are handled by
+  // the cells, the TP/DP knob rules and the Quantization axis, not here.
+  isNpuHw(s) {
+    return s.hw === "a3" || s.hw === "a5";
+  },
  // Does this recipe shard the TP-replicated MLA KV? Replaces the per-platform
  // "which cells carry DCP" tables the HiCache tiers used to hardcode.
  hasDcp(s) {
@@ -139,14 +150,14 @@ export const config = {
        {
          id: "prefill",
          label: "Prefill",
-          disabled: (s) => s.hw === "a3",
-          disableReason: (s) => (s.hw === "a3" ? "Only Unified PD is supported on this recipe." : ""),
+          disabled: (s) => config.isNpuHw(s),
+          disableReason: (s) => (config.isNpuHw(s) ? "Only Unified PD is supported on this recipe." : ""),
        },
        {
          id: "decode",
          label: "Decode",
-          disabled: (s) => s.hw === "a3",
-          disableReason: (s) => (s.hw === "a3" ? "Only Unified PD is supported on this recipe." : ""),
+          disabled: (s) => config.isNpuHw(s),
+          disableReason: (s) => (config.isNpuHw(s) ? "Only Unified PD is supported on this recipe." : ""),
        },
      ],
    },
@@ -160,16 +171,16 @@ export const config = {
          id: "low-latency",
          label: "Low-Latency",
          showWhen: (s) => s.pdMode !== "prefill",
-          disabled: (s) => s.hw === "a3",
-          disableReason: (s) => (s.hw === "a3" ? "Only the Balanced operating point is supported on this recipe." : ""),
+          disabled: (s) => config.isNpuHw(s),
+          disableReason: (s) => (config.isNpuHw(s) ? "Only the Balanced operating point is supported on this recipe." : ""),
        },
        { id: "balanced",        label: "Balanced",        showWhen: (s) => s.pdMode !== "prefill" },
        {
          id: "high-throughput",
          label: "High-Throughput",
          showWhen: (s) => s.pdMode !== "prefill",
-          disabled: (s) => s.hw === "a3",
-          disableReason: (s) => (s.hw === "a3" ? "Only the Balanced operating point is supported on this recipe." : ""),
+          disabled: (s) => config.isNpuHw(s),
+          disableReason: (s) => (config.isNpuHw(s) ? "Only the Balanced operating point is supported on this recipe." : ""),
        },
        { id: "default",         label: "Default",         showWhen: (s) => s.pdMode === "prefill" },
        { id: "long-context",    label: "Long-Context",    showWhen: (s) => s.pdMode === "prefill" },
@@ -194,11 +205,15 @@ export const config = {
      default: "mxfp4",
      options: [
        { id: "mxfp4", label: "MXFP4", subtitle: "Moonshot AI checkpoint",
+          // The 950PR/DT recipe serves this checkpoint; the A3 Series recipe
+          // serves the W4A8 build under it instead.
          disabled: (s) => s.hw === "a3",
-          disableReason: (s) => (s.hw === "a3" ? "Only Modelslim (W4A8) is supported on this recipe." : ""),
+          disableReason: (s) => (s.hw === "a3" ? "The A3 Series recipe serves the sgl-npu Modelslim (W4A8) checkpoint." : ""),
        },
        {
-          // A3 Series only (NPU W4A8 checkpoint); hidden on the GPU recipes.
+          // A3 Series only (ModelSlim W4A8 checkpoint); hidden elsewhere — the
+          // 950PR/DT recipe serves the MXFP4 checkpoint above with no
+          // --quantization flag.
          id: "modelslim",
          label: "Modelslim (W4A8)",
          subtitle: "ModelScope NPU checkpoint",
@@ -227,7 +242,7 @@ export const config = {
      id: "mmTransport",
      title: "VLM Transport",
      default: "auto",
-      showWhen: (s) => s.pdMode !== "decode" && s.hw !== "a3",
+      showWhen: (s) => s.pdMode !== "decode" && !config.isNpuHw(s),
      options: [
        {
          id: "auto",
@@ -266,8 +281,8 @@ export const config = {
      options: [
        { id: "none", label: "Non-Spec",
          env: (s) => (["mi350x", "mi355x"].includes(s.hw) ? ["SGLANG_MLA_DECODE_TUNE=1"] : []),
-          disabled: (s) => s.hw === "a3",
-          disableReason: (s) => (s.hw === "a3" ? "Only DSPARK is supported on this recipe." : ""),
+          disabled: (s) => config.isNpuHw(s),
+          disableReason: (s) => (config.isNpuHw(s) ? "Only DSPARK is supported on this recipe." : ""),
        },
        {
          id: "dspark",
@@ -296,10 +311,10 @@ export const config = {
            "--speculative-algorithm DSPARK",
            "--speculative-draft-model-path RadixArk/Kimi-K3-DSpark",
            "--speculative-dspark-block-size 7",
-            // The NPU recipe adds the NPU draft path's own knobs (draft
-            // attention backend, topk 1, unquantized draft weights) on top of
-            // the common trio.
-            ...(s.hw === "a3"
+            // The NPU recipes (A3 and A5) add the NPU draft path's own knobs
+            // (draft attention backend, topk 1, unquantized draft weights) on
+            // top of the common trio.
+            ...(config.isNpuHw(s)
              ? [
                  "--speculative-draft-attention-backend ascend",
                  "--speculative-eagle-topk 1",
@@ -311,7 +326,7 @@ export const config = {
            // the Triton decode kernel, the K3 default). It is CUDA-only, and
            // the PD prefill role opts out — it never runs verify and rejects
            // the flag at startup.
-            ...(s.hw !== "a3" && s.pdMode !== "prefill"
+            ...(!config.isNpuHw(s) && s.pdMode !== "prefill"
              ? ["--enable-linear-replayssm-spec"]
              : []),
          ],
@@ -346,8 +361,8 @@ export const config = {
        {
          id: "l2",
          label: "L1+L2 (host)",
-          disabled: (s) => s.hw === "a3",
-          disableReason: (s) => (s.hw === "a3"
+          disabled: (s) => config.isNpuHw(s),
+          disableReason: (s) => (config.isNpuHw(s)
            ? "NPU HiCache does not support mamba cache (K3's KDA state is one)."
            : ""),
          flags: [
@@ -370,8 +385,8 @@ export const config = {
        {
          id: "l3",
          label: "+ L3 (Mooncake)",
-          disabled: (s) => s.hw === "a3",
-          disableReason: (s) => (s.hw === "a3"
+          disabled: (s) => config.isNpuHw(s),
+          disableReason: (s) => (config.isNpuHw(s)
            ? "NPU HiCache does not support mamba cache (K3's KDA state is one)."
            : ""),
          flags: [
@@ -413,6 +428,10 @@ export const config = {
    default: "moonshotai/Kimi-K3",
    nvfp4: "nvidia/Kimi-K3-NVFP4",
    a3: "sgl-npu/Kimi-K3-W4A8",
+    // The 950PR/DT recipe serves the official Moonshot checkpoint (MXFP4) from
+    // ModelScope — the same id the NVIDIA recipes resolve to, fetched through
+    // SGLANG_USE_MODELSCOPE. The A3 Series recipe serves the W4A8 build.
+    a5: "moonshotai/Kimi-K3",
  },

  placeholders: {
@@ -468,6 +487,9 @@ export const config = {
    "b200|nvfp4":  "lmsysorg/sglang:dev-dev-kimi-k3-nvfp4",
    "gb200|nvfp4": "lmsysorg/sglang:dev-dev-kimi-k3-nvfp4",
    a3:     "quay.io/ascend/sglang:main-cann9.0.0-a3",
+    // 950PR/DT Series builds are CANN 9.1.0 and published from the Ascend SWR
+    // registry (the quay.io A2/A3 line carries no 950PR/DT tag).
+    a5:     "swr.cn-southwest-2.myhuaweicloud.com/base_image/dockerhub/lmsysorg/sglang:cann9.1.0-950-B070",
  },
  // Pre-selects the issue template's `model` field on "Submit verified cell".
  github: {
@@ -492,6 +514,10 @@ export const config = {
                when: { hw: ["a3"] },
                reason: "Only TP64 is supported on this recipe.",
              },
+              {
+                when: { hw: ["a5"] },
+                reason: "Only TP32 is supported on this recipe.",
+              },
            ]; },
          },
          {
@@ -501,6 +527,10 @@ export const config = {
                when: { hw: ["a3"] },
                reason: "Only TP64 is supported on this recipe.",
              },
+              {
+                when: { hw: ["a5"] },
+                reason: "Only TP32 is supported on this recipe.",
+              },
              {
                when: { hw: ["b300", "gb300"] },
                reason: "TP=16 needs 16 ranks; the B300 and GB300 recipes have 8 ranks.",
@@ -513,14 +543,27 @@ export const config = {
            ]; },
          },
          {
-            // A3 Series only: 64 ranks (4 nodes × 8 cards × 2 dies); hidden on the GPU recipes.
+            // 950PR/DT Series only: 32 ranks (4 nodes × 8 cards, one rank per
+            // card); hidden on the other recipes.
+            value: 32,
+            hide: { hw: ["b300", "gb300", "b200", "gb200", "h200", "h100", "mi350x", "mi355x", "a3"] },
+          },
+          {
+            // A3 Series only: 64 ranks (4 nodes × 8 cards × 2 dies); hidden on
+            // the other recipes.
            value: 64,
-            hide: { hw: ["b300", "gb300", "b200", "gb200", "h200", "h100", "mi350x", "mi355x"] },
+            hide: { hw: ["b300", "gb300", "b200", "gb200", "h200", "h100", "mi350x", "mi355x", "a5"] },
          },
        ]},
        { id: "dpAttn", label: "DP-Attention",
          values: [
            null,
+            {
+              // 950PR/DT Series only: the recipe enables DP-Attention at dp=1
+              // (attn-TP 32), unlike A3's dp=4.
+              value: 1,
+              hide: { hw: ["b300", "gb300", "b200", "gb200", "h200", "h100", "mi350x", "mi355x", "a3"] },
+            },
            {
              value: false,
              get disable() { return [
@@ -528,6 +571,10 @@ export const config = {
                  when: { hw: ["a3"] },
                  reason: "Only DP-Attention=4 is supported on this recipe.",
                },
+                {
+                  when: { hw: ["a5"] },
+                  reason: "Only DP-Attention=1 is supported on this recipe.",
+                },
              ]; },
            },
            {
@@ -537,9 +584,21 @@ export const config = {
                  when: { hw: ["a3"] },
                  reason: "Only DP-Attention=4 is supported on this recipe.",
                },
+                {
+                  when: { hw: ["a5"] },
+                  reason: "Only DP-Attention=1 is supported on this recipe.",
+                },
+              ]; },
+            },
+            {
+              value: 4,
+              get disable() { return [
+                {
+                  when: { hw: ["a5"] },
+                  reason: "Only DP-Attention=1 is supported on this recipe.",
+                },
              ]; },
            },
-            4,
            {
              value: 8,
              get disable() { return [
@@ -547,6 +606,10 @@ export const config = {
                  when: { hw: ["a3"] },
                  reason: "Only DP-Attention=4 is supported on this recipe.",
                },
+                {
+                  when: { hw: ["a5"] },
+                  reason: "Only DP-Attention=1 is supported on this recipe.",
+                },
                {
                  when: { hw: ["b300", "gb300"] },
                  reason: "On an 8-rank deployment (B300 1×8, GB300 2×4) dp=8 leaves attn_tp=1, so each rank holds the full unsharded MLA KV and OOMs — prefer dp=2/attn_tp=4.",
@@ -565,6 +628,10 @@ export const config = {
                  when: { hw: ["a3"] },
                  reason: "Only DP-Attention=4 is supported on this recipe.",
                },
+                {
+                  when: { hw: ["a5"] },
+                  reason: "Only DP-Attention=1 is supported on this recipe.",
+                },
                {
                  when: { hw: ["b300", "gb300"] },
                  reason: "DP-Attention=16 needs 16 TP ranks; the B300 and GB300 recipes have 8.",
@@ -595,10 +662,10 @@ export const config = {
          // Blackwell-only: runs FlashInfer's official trtllm-gen SiTU kernels.
          { id: "flashinfer_mxfp4", label: "FlashInfer (MXFP4)", flags: ["--moe-runner-backend flashinfer_mxfp4"],
            requiresHw: ["b200", "b300", "gb200", "gb300"],
-            disable: [{ when: { hw: ["a3"] },
+            disable: [{ when: { hw: ["a3", "a5"] },
              reason: "FlashInfer is a CUDA kernel and is not supported on NPU." }] },
          { id: "marlin",           label: "Marlin (W4A16)",    flags: ["--moe-runner-backend marlin"],
-            disable: [{ when: { hw: ["a3"] },
+            disable: [{ when: { hw: ["a3", "a5"] },
              reason: "Marlin is a CUDA kernel and is not supported on NPU." }] },
        ],
      },
@@ -614,7 +681,7 @@ export const config = {
        ],
      },
      ep: {
-        showWhen: (b) => b.hw !== "a3",
+        showWhen: (b) => !config.isNpuHw(b),
        label: "EP",
        values: [
          null, 1, 2, 4, 8,
@@ -640,9 +707,10 @@ export const config = {
    parsers: {
      items: [
        { id: "reasoning", label: "Reasoning Parser", flag: "--reasoning-parser kimi_k3" },
-        // Tool calling is not yet supported on the NPU, so the item hides on a3.
+        // Tool calling is not yet supported on the NPU recipes, so the item
+        // hides on a3 and a5.
        { id: "toolCall",  label: "Tool Call Parser", flag: "--tool-call-parser kimi_k3",
-          hide: { hw: ["a3"] } },
+          hide: { hw: ["a3", "a5"] } },
      ],
    },

@@ -719,8 +787,8 @@ export const config = {
        //   EAGLE   --speculative-num-steps N           (chain; topk>1 is a tree)
        // Only DSPARK is selectable today, so only its form is emitted.
        id: "proposedDraftTokens", title: "Proposed Draft Tokens",
-        // The A3 Series recipe pins the shipped block size (7).
-        showWhen: (b) => b.spec === "dspark" && b.hw !== "a3",
+        // The NPU recipes pin the shipped block size (7).
+        showWhen: (b) => b.spec === "dspark" && !config.isNpuHw(b),
        control: "slider",
        stripPrefixes: [
          "--speculative-dspark-block-size",
@@ -743,10 +811,10 @@ export const config = {
        // Spec-only, so gate the row on DSPARK; every DSPARK recipe (except the PD
        // prefill role) turns it on in the base, so this row derives to On and
        // exists mainly as the opt-out.
-        // Needs the Triton linear-attn decode backend (the K3 default); the A3 Series
-        // script never sets it.
+        // Needs the Triton linear-attn decode backend (the K3 default); the NPU
+        // recipes never enable it.
        id: "replaySsm", title: "ReplaySSM (spec)",
-        showWhen: (b) => b.spec === "dspark" && b.hw !== "a3",
+        showWhen: (b) => b.spec === "dspark" && !config.isNpuHw(b),
        stripPrefixes: ["--enable-linear-replayssm-spec"],
        options: [
          { id: "off", label: "Off" },
@@ -766,8 +834,8 @@ export const config = {
        // without --speculative-dspark-sps-table-path (every step still
        // verifies full width); fails fast with ReplaySSM or DCP > 1.
        id: "raggedVerify", title: "Ragged Verify Mode (spec)",
-        // The A3 Series recipe pins static.
-        showWhen: (b) => b.spec === "dspark" && b.hw !== "a3",
+        // The NPU recipes pin static.
+        showWhen: (b) => b.spec === "dspark" && !config.isNpuHw(b),
        stripEnv: ["SGLANG_RAGGED_VERIFY_MODE"],
        options: [
          { id: "static",  label: "Auto (static)" },
@@ -776,7 +844,7 @@ export const config = {
      },
      {
        id: "kvCacheDtype", title: "KV Cache Precision",
-        showWhen: (b) => b.hw !== "a3",
+        showWhen: (b) => !config.isNpuHw(b),
        stripPrefixes: ["--kv-cache-dtype"],
        options: [
          { id: "auto", label: "Auto (BF16)" },
@@ -785,7 +853,7 @@ export const config = {
      },
      {
        id: "mambaSsmDtype", title: "KDA State Precision",
-        showWhen: (b) => b.hw !== "a3",
+        showWhen: (b) => !config.isNpuHw(b),
        stripPrefixes: ["--mamba-ssm-dtype"],
        options: [
          { id: "auto", label: "Auto (FP32)" },
@@ -800,7 +868,7 @@ export const config = {
        // Off suits prefix-free traffic (offline batch, evals): 1 state slot
        // per request instead of 4-5.
        id: "prefixCache", title: "Prefix Cache",
-        showWhen: (b) => b.hw !== "a3",
+        showWhen: (b) => !config.isNpuHw(b),
        stripPrefixes: ["--disable-radix-cache"],
        options: [
          { id: "on",  label: "On" },
@@ -812,7 +880,7 @@ export const config = {
        // prefix cache off, so the row hides (and stops emitting) there.
        // Slot cost per request: extra_buffer 5, extra_buffer_lazy 4.
        id: "mambaRadix", title: "KDA Radix Cache Strategy",
-        showWhen: (b, v, d) => (b.hw !== "a3")
+        showWhen: (b, v, d) => (!config.isNpuHw(b))
          && ((((v && v.prefixCache) ?? (d && d.prefixCache)) !== "off")),
        stripPrefixes: ["--mamba-radix-cache-strategy"],
        options: [
@@ -827,7 +895,7 @@ export const config = {
        // (extra_buffer 5→4, extra_buffer_lazy 4→3; no_buffer stays 3). Off by
        // default. Env var, not a flag, so it emits via env/stripEnv.
        id: "mambaSlotSaving", title: "KDA Slot Saving (experimental)",
-        showWhen: (b) => b.hw !== "a3",
+        showWhen: (b) => !config.isNpuHw(b),
        stripEnv: ["SGLANG_OPT_MAMBA_SKIP_DECODE_LOCK"],
        options: [
          { id: "off", label: "Off" },
@@ -863,7 +931,7 @@ export const config = {
      {
        // Only meaningful with EP a2a on (MoE card or a large-scale preset).
        id: "eplb", title: "Expert Rebalancing (EPLB)",
-        showWhen: (b) => b.hw !== "a3",
+        showWhen: (b) => !config.isNpuHw(b),
        stripPrefixes: ["--enable-eplb"],
        options: [
          { id: "off", label: "Off" },
@@ -875,7 +943,7 @@ export const config = {
        // Validated on the no-a2a MXFP4 runner; untested against SBO (EP a2a)
        // and DP attention.
        id: "prefillGraph", title: "Prefill CUDA Graph",
-        showWhen: (b) => b.hw !== "a3",
+        showWhen: (b) => !config.isNpuHw(b),
        stripPrefixes: ["--cuda-graph-backend-prefill"],
        options: [
          { id: "auto", label: "Auto (off)" },
@@ -898,7 +966,7 @@ export const config = {
        // resolves it into the full parallelism shape (tp/ep/dp/dcp; attn-tp =
        // tp/dp); pool sizing rides the calculator-driven ratio.
        id: "lsGpus", title: "Cluster Size (large-scale)",
-        showWhen: (b) => (b.hw !== "a3") && (b.pdMode === undefined || b.pdMode === "unified"),
+        showWhen: (b) => (!config.isNpuHw(b)) && (b.pdMode === undefined || b.pdMode === "unified"),
        // Default follows the base cell's own GPU count (tp8 lanes -> 8,
        // tp16 lanes -> 16), so a preset starts from "same hardware, new shape".
        default: (b) =>
@@ -914,7 +982,7 @@ export const config = {
      },
      {
        id: "lsPreset", title: "Large-Scale Preset",
-        showWhen: (b) => (b.hw !== "a3") && (b.pdMode === undefined || b.pdMode === "unified"),
+        showWhen: (b) => (!config.isNpuHw(b)) && (b.pdMode === undefined || b.pdMode === "unified"),
        stripPrefixes: [
          "--tp-size", "--tp", "--tensor-parallel-size",
          "--ep-size", "--ep", "--expert-parallel-size",
@@ -2376,6 +2444,89 @@ export const config = {
        "--port {{PORT}}",
      ],
    },
+    {
+      // Ascend 950PR/DT Series: 4 nodes × 8 cards, one rank per card (TP32).
+      // Unified PD, Balanced, DSPARK-only, with the product-line kernels armed
+      // per env: FIAS V2 BSND for the DSpark target-verify/draft attention
+      // paths and the fine-grained dual-stream MoE overlap. DP-attention runs
+      // at dp=1 (attn-TP 32), and the shared experts / dense MLP shard across
+      // attention-TP through the server flags (--shared-experts-tp-size 4).
+      // Checkpoint: the official Moonshot MXFP4 build (moonshotai/Kimi-K3,
+      // fetched from ModelScope by SGLANG_USE_MODELSCOPE=1 above). Its routed
+      // experts declare compressed-tensors "mxfp4-pack-quantized", which the
+      // loader detects from the checkpoint itself — the NPU MXFP4 MoE scheme
+      // serves them — so the recipe passes no --quantization flag (attention,
+      // shared experts and the dense MLP are ignored by the checkpoint and stay
+      // BF16). A ModelSlim (W4A8) checkpoint would need --quantization modelslim
+      // and a modelslim-style --model-path instead; the two are not mixable.
+      // Pool sizing is internal to the Ascend path: the KDA state and MLA KV
+      // pools are sized by the runtime, so the recipe sets neither
+      // --mamba-full-memory-ratio nor --max-mamba-cache-size, and the radix
+      // cache is off (one KDA state slot per request) — the ratio calculator
+      // does not apply.
+      // --disable-custom-all-reduce is carried verbatim from the recipe script.
+      // Ascend has no custom all-reduce kernel, so the NPU backend resolves the
+      // flag to True on its own (hardware_backend/npu/utils.py); spelling it out
+      // keeps the panel command identical to the command that was measured.
+      // The `unset` lines of the source script (ASCEND_CUSTOM_OPP_PATH,
+      // SGLANG_NPU_FUSED_MOE_MODE, ENABLE_PROFILING, the K3 trace files) are
+      // launch-script hygiene against stale profiling state, not recipe env —
+      // a panel command starts clean, so they don't carry over.
+      match: { hw: "a5", pdMode: "unified", strategy: "balanced" },
+      nnodes: 4,
+      verified: false,
+      verificationStatus: "in-progress",
+      env: [
+        "SGLANG_USE_MODELSCOPE=1",
+        "GLOO_SOCKET_IFNAME={{NETWORK_IFACE}}",
+        "HCCL_SOCKET_IFNAME={{NETWORK_IFACE}}",
+        "PYTORCH_NPU_ALLOC_CONF=expandable_segments:True",
+        "SGLANG_NPU_USE_FIAS_V2_BSND=True",
+        "SGLANG_NPU_FINE_GRAINED_MOE_DUAL_STREAM=True",
+        "SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1",
+        "SGLANG_ENABLE_SPEC_V2=1",
+        "SGLANG_RAGGED_VERIFY_MODE=static",
+        "SGLANG_DSPARK_FOLDED_PROPOSAL=0",
+        "SGLANG_DSPARK_FOLDED_SAMPLING=0",
+        "SGLANG_DSPARK_STACKED_CTX_KV=0",
+        "SGLANG_DSPARK_EMBED_IN_GRAPH=0",
+        "STREAMS_PER_DEVICE=32",
+        "DEEP_NORMAL_MODE_USE_INT8_QUANT=1",
+        "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=128",
+        "HCCL_BUFFSIZE=2000",
+        "DEEPEP_NORMAL_LONG_SEQ_ROUND=64",
+        "DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS=512",
+        "HCCL_OP_EXPANSION_MODE=AIV",
+      ],
+      flags: [
+        "--trust-remote-code",
+        "--model-path {{MODEL_NAME}}",
+        "--tokenizer-path {{MODEL_NAME}}",
+        "--attention-backend ascend",
+        "--device npu",
+        "--dtype bfloat16",
+        "--tp-size 32",
+        "--enable-dp-attention",
+        "--enable-dp-lm-head",
+        "--mem-fraction-static 0.9",
+        "--chunked-prefill-size 8192",
+        "--cuda-graph-bs-decode 32",
+        "--max-running-requests 32",
+        "--enable-shared-experts-attn-tp",
+        "--enable-dense-mlp-attn-tp",
+        "--shared-experts-tp-size 4",
+        "--reasoning-parser kimi_k3",
+        "--moe-a2a-backend deepep",
+        "--deepep-mode auto",
+        "--linear-attn-verify-backend triton",
+        "--disable-radix-cache",
+        "--disable-custom-all-reduce",
+        "--watchdog-timeout 9000",
+        "--model-loader-extra-config '{\"enable_multithread_load\": true}'",
+        "--host {{HOST_IP}}",
+        "--port {{PORT}}",
+      ],
+    },
  ],

  // Cross-node fabric env (substitute the NIC used by every rank).
@@ -2418,5 +2569,15 @@ export const config = {
      "  SGLANG_HOST_IP=<this-node-ip>   # this node's IP on that NIC",
      "If running outside the official image, source set_env.sh on every node first.",
    ],
+    a5: [
+      "Run the same command on all four nodes with --node-rank 0/1/2/3.",
+      "NPU collectives use HCCL. Pin the cross-node NIC on EVERY node:",
+      "  GLOO_SOCKET_IFNAME=<your-nic>   # bootstrap interface",
+      "  HCCL_SOCKET_IFNAME=<your-nic>   # HCCL transport interface",
+      "  SGLANG_HOST_IP=<this-node-ip>   # this node's IP on that NIC",
+      "If running outside the official image, source both set_env.sh scripts on every",
+      "node first: /usr/local/Ascend/ascend-toolkit/set_env.sh and",
+      "/usr/local/Ascend/nnal/atb/set_env.sh.",
+    ],
  },
 };