feat(inference): add Nemotron 3.5 Lightning Spark profile

docs/inference/set-up-vllm.mdx

@@ -267,6 +267,24 @@ When you start managed vLLM outside the installer express flow, NemoClaw uses th
 For the managed single-node DGX Spark `nvidia/Qwen3.6-35B-A3B-NVFP4` profile, NemoClaw enables async scheduling and does not enable multi-token prediction (MTP) speculative decoding.
 The profile retains `--gpu-memory-utilization 0.4`, a `262144`-token context window, `4` concurrent sequences, an `8192`-token batch limit, chunked prefill, prefix caching, and its registered parsers and acceleration backends.
 
+NVIDIA Nemotron 3.5 Lightning is an additional explicit-only single-DGX Spark profile and does not change the Qwen default.
+Select its stable catalog ID after confirming that `$$nemoclaw profiles list` reports it as compatible.
+
+```bash
+$$nemoclaw onboard --profile vllm.dgx-spark-gb10.single.nemotron-3.5-lightning-30b-a3b-nvfp4
+```
+
+The profile serves the public [`nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4`](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4) checkpoint through a pinned ARM64 vLLM image.
+Its profile limits for one concurrent sequence are a 65,536-token context window, a 4,096-token batch limit, 0.7 GPU-memory utilization, FP8 KV cache, and one-token MTP speculative decoding.
+It uses vLLM's Python frontend because the pinned runtime's Rust frontend does not register the required `step3p5` reasoning and tool-call parsers.
+Physical single-DGX Spark validation covered plain chat, automatic and named structured tool calls, the 10-turn 32K-input/2K-output release workload, OpenClaw chat and file tools, cached restart, cleanup, and restoration of the original Qwen route.
+
+<Warning title="Experimental Nemotron 3.5 Lightning vLLM Profile">
+This profile is an explicit opt-in for one DGX Spark and remains Experimental.
+Promotion requires broader validation of the pinned Python frontend and the `step3p5` reasoning and tool-call parsers.
+The fixed recipe intentionally rejects model and serve-argument overrides; use another catalog profile instead of modifying its validated resource boundary.
+</Warning>
+
 Muse Glimmer is an additional single-DGX Spark choice and does not change the Qwen default.
 Select `muse-glimmer-30b` to serve `Inferact/Muse-Glimmer-30B-NVFP4-W4A4` with the `muse-glimmer` alias.
 The registered recipe configures a 32,768-token context window, 1 concurrent sequence, a 4,096-token batch limit, and 0.75 GPU-memory utilization.

managed-inference/presets/vllm.dgx-spark-gb10.single.nemotron-3.5-lightning-30b-a3b-nvfp4.yaml

@@ -0,0 +1,106 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+apiVersion: nemoclaw.nvidia.com/managed-inference/v1
+kind: ServingPreset
+
+metadata:
+  id: vllm.dgx-spark-gb10.single.nemotron-3.5-lightning-30b-a3b-nvfp4
+  displayName: NVIDIA Nemotron 3.5 Lightning 30B-A3B NVFP4 on one DGX Spark
+  supportState: experimental
+
+spec:
+  selection: explicit-only
+  priority: 450
+
+  requirements:
+    all:
+      - readiness:
+          scope: everyNode
+          kind: qualification
+          id: host.platform.dgx_spark
+          status: qualified
+      - readiness:
+          scope: everyNode
+          kind: capability
+          id: host.platform.supported
+          state: present
+      - readiness:
+          scope: everyNode
+          kind: capability
+          id: host.platform.dgx_spark
+          state: present
+      - readiness:
+          scope: everyNode
+          kind: capability
+          id: host.docker.available
+          state: present
+      - readiness:
+          scope: everyNode
+          kind: capability
+          id: host.docker.daemon_reachable
+          state: present
+      - readiness:
+          scope: everyNode
+          kind: capability
+          id: host.docker.runtime_supported
+          state: present
+      - readiness:
+          scope: everyNode
+          kind: capability
+          id: host.docker.storage_compatible
+          state: present
+      - readiness:
+          scope: everyNode
+          kind: capability
+          id: host.gpu.nvidia_available
+          state: present
+      - readiness:
+          scope: everyNode
+          kind: capability
+          id: host.gpu.container_toolkit_available
+          state: present
+      - readiness:
+          scope: everyNode
+          kind: capability
+          id: host.gpu.cdi_healthy
+          state: present
+      - readiness:
+          scope: everyNode
+          kind: observation
+          id: host.os.platform
+          comparison:
+            operator: equals
+            value: linux
+      - readiness:
+          scope: everyNode
+          kind: observation
+          id: host.os.architecture
+          comparison:
+            operator: equals
+            value: arm64
+      - readiness:
+          scope: everyNode
+          kind: observation
+          id: host.docker.runtime
+          comparison:
+            operator: equals
+            value: docker
+      - readiness:
+          scope: everyNode
+          kind: observation
+          id: host.gpu.count
+          comparison:
+            operator: at-least
+            value: 1
+      - readiness:
+          scope: everyNode
+          kind: observation
+          id: host.gpu.driver_version
+          comparison:
+            operator: version-at-least
+            value: 580.65.06
+
+  plan:
+    backend: vllm
+    recipeRef: vllm.nemotron-3.5-lightning-30b-a3b-nvfp4.spark-single.v1

managed-inference/recipes/vllm.nemotron-3.5-lightning-30b-a3b-nvfp4.spark-single.v1.yaml

@@ -0,0 +1,88 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+apiVersion: nemoclaw.nvidia.com/managed-inference/v1
+kind: ServingRecipe
+
+metadata:
+  id: vllm.nemotron-3.5-lightning-30b-a3b-nvfp4.spark-single.v1
+  displayName: NVIDIA Nemotron 3.5 Lightning 30B-A3B NVFP4 with vLLM
+
+spec:
+  backend: vllm
+
+  model:
+    id: nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4
+    revision: 0dcd680e5585c791728c83342b311d0a0026dbeb
+    servedName: nvidia-nemotron-3.5-lightning-30b-a3b-nvfp4
+    downloadSizeBytes: 21561882284
+    gated: false
+    installFastSafetensors: false
+    preparation:
+      ref: none/v1
+
+  runtime:
+    image: vllm/vllm-openai@sha256:3af90144a0926e5c5fe46ee16e5201e763dd854538b9d7ce433755f11dadaf78
+    imageDownloadSizeBytes: 12691539532
+    pullTimeoutSeconds: 3600
+    architecture: arm64
+    networkMode: bridge
+    ipcMode: host
+    sharedMemoryBytes: 34359738368
+    gpuRequest: all
+    devices: []
+    ulimits:
+      memlock: -1
+      stackBytes: 67108864
+    modelCache:
+      source: huggingface-cache
+      target: /root/.cache/huggingface
+    temporaryFilesystems: []
+    environment:
+      # The Lightning 3.5 Spark validation used vLLM's Python frontend;
+      # `step3p5` parsers are not available in the Rust frontend.
+      VLLM_USE_RUST_FRONTEND: "0"
+
+  execution:
+    materializerRef: vllm.host-local/v1
+    lifecycleRef: vllm.host-local.lifecycle/v1
+
+  serve:
+    authentication: bearer
+    executable: /usr/local/bin/vllm
+    arguments:
+      - name: --max-model-len
+        value: 65536
+      - name: --max-num-seqs
+        value: 1
+      - name: --gpu-memory-utilization
+        value: 0.7
+      - name: --kv-cache-dtype
+        value: fp8
+      - name: --moe-backend
+        value: marlin
+      - name: --max-num-batched-tokens
+        value: 4096
+      - name: --speculative-config
+        value: '{"method":"mtp","num_speculative_tokens":1,"moe_backend":"flashinfer_cutlass"}'
+      - name: --mamba-backend
+        value: flashinfer
+      - name: --mamba-ssm-cache-dtype
+        value: float16
+      - name: --enable-mamba-cache-stochastic-rounding
+      - name: --mamba-cache-philox-rounds
+        value: 5
+      - name: --enable-prefix-caching
+      - name: --mamba-cache-mode
+        value: align
+      - name: --compilation-config
+        value: '{"cudagraph_capture_sizes":[1,2,4,8,16,24,32,40,48,56,64,72,80,88,96,104,112,120,128,136,144,152,160,168,176,184,192,200,208,216,224,232,240,248,256,1024,2048,4096]}'
+      - name: --enable-auto-tool-choice
+      - name: --tool-call-parser
+        value: step3p5
+      - name: --reasoning-parser
+        value: step3p5
+
+  readiness:
+    timeoutSeconds: 1800
+    expectedModel: nvidia-nemotron-3.5-lightning-30b-a3b-nvfp4

src/lib/inference/serving/adapter-registry.test.ts

@@ -36,6 +36,14 @@ function shippedRecipe(): ManagedInferenceServingRecipe {
   return structuredClone(recipe as ManagedInferenceServingRecipe);
 }
 
+function shippedHostLocalRecipe(): ManagedInferenceServingRecipe {
+  const recipe = loadManagedInferenceCatalog().recipes.find(
+    ({ spec }) => spec.execution.materializerRef === HOST_LOCAL_VLLM_MATERIALIZER_REF,
+  );
+  expect(recipe).toBeDefined();
+  return structuredClone(recipe as ManagedInferenceServingRecipe);
+}
+
 describe("managed inference adapter registries", () => {
   it("registers one versioned descriptor for each shipped managed cluster mechanic", () => {
     expect(listManagedInferenceTopologyQualificationDescriptors()).toMatchObject([
@@ -143,6 +151,45 @@ describe("managed inference adapter registries", () => {
     expect(getManagedInferenceRecipeRegistrationError(wrongShape)).toMatch(/TP times PP/u);
   });
 
+  it("accepts shell-quoted JSON values for host-local vLLM arguments", () => {
+    const recipe = shippedHostLocalRecipe();
+    const structuredArgument = {
+      ...recipe,
+      spec: {
+        ...recipe.spec,
+        serve: {
+          ...recipe.spec.serve,
+          arguments: [
+            ...recipe.spec.serve.arguments,
+            {
+              name: "--speculative-config",
+              value: '{"method":"mtp","num_speculative_tokens":1}',
+            },
+          ],
+        },
+      },
+    } as ManagedInferenceServingRecipe;
+    const unsafeArgument = {
+      ...structuredArgument,
+      spec: {
+        ...structuredArgument.spec,
+        serve: {
+          ...structuredArgument.spec.serve,
+          arguments: structuredArgument.spec.serve.arguments.map((argument) =>
+            argument.name === "--speculative-config"
+              ? { ...argument, value: '{"method":"mtp"};touch/tmp/unsafe' }
+              : argument,
+          ),
+        },
+      },
+    } as ManagedInferenceServingRecipe;
+
+    expect(getManagedInferenceRecipeRegistrationError(structuredArgument)).toBeUndefined();
+    expect(getManagedInferenceRecipeRegistrationError(unsafeArgument)).toMatch(
+      /bounded safe text/u,
+    );
+  });
+
   it("rejects schema-valid values that the registered adapter cannot execute", () => {
     const recipe = shippedRecipe();
     const unsupportedCache = {

src/lib/inference/serving/adapter-registry.ts

@@ -81,6 +81,9 @@ const STABLE_ID = /^[A-Za-z0-9][A-Za-z0-9._/-]{0,159}$/u;
 const PINNED_IMAGE =
   /^(?:[a-z0-9]+(?:[._-][a-z0-9]+)*(?::[0-9]+)?\/)?(?:[a-z0-9]+(?:[._-][a-z0-9]+)*\/)*[a-z0-9]+(?:[._-][a-z0-9]+)*@sha256:[0-9a-f]{64}$/u;
 const SAFE_ENVIRONMENT_NAME = /^[A-Z][A-Z0-9_]{0,127}$/u;
+// Structured vLLM flags are passed as individual argv values after shell
+// quoting. Keep the allowlist narrow while admitting JSON objects and arrays.
+const HOST_LOCAL_SAFE_VALUE = /^[A-Za-z0-9_@%+=:,./{}[\]"-]+$/u;
 const SHA256_DIGEST = /^sha256:[0-9a-f]{64}$/u;
 const MANAGED_CLUSTER_MATERIALIZER_OWNED_ENVIRONMENT = new Set([
   "GLOO_SOCKET_IFNAME",
@@ -410,13 +413,13 @@ function validateHostLocalVllmMaterializerRecipe(
         typeof value === "string" &&
         (Buffer.byteLength(value, "utf8") > 16_384 ||
           value.includes("\0") ||
-          !/^[A-Za-z0-9_@%+=:,./-]+$/u.test(value)),
+          !HOST_LOCAL_SAFE_VALUE.test(value)),
     ) ||
     Object.values(recipe.spec.runtime.environment).some(
       (value) =>
         Buffer.byteLength(value, "utf8") > 4_096 ||
         value.includes("\0") ||
-        !/^[A-Za-z0-9_@%+=:,./-]+$/u.test(value),
+        !HOST_LOCAL_SAFE_VALUE.test(value),
     )
   ) {
     return "host-local vLLM serving values must be bounded safe text";

src/lib/inference/serving/catalog-loader.test.ts

@@ -32,21 +32,25 @@ const EMPTY_CATALOG: CompiledServingCatalog = {
 const EXPECTED_MANAGED_RECIPE_IDS = [
   "llama-cpp.nemotron-3-nano-30b-a3b.spark-single.v1",
   "vllm.deepseek-v4-flash-0731.spark-dual.v1",
+  "vllm.nemotron-3.5-lightning-30b-a3b-nvfp4.spark-single.v1",
   "vllm.qwen3-6-35b-a3b-nvfp4.spark-single.v1",
 ];
 const EXPECTED_MANAGED_PRESET_IDS = [
   "llama-cpp.dgx-spark-gb10.single.nemotron-3-nano-30b-a3b",
   "llama-cpp.linux-amd64-nvidia.single.nemotron-3-nano-30b-a3b",
   "local-model-profile.vllm.spark.v1",
   "vllm.dgx-spark-gb10.dual.deepseek-v4-flash-0731",
+  "vllm.dgx-spark-gb10.single.nemotron-3.5-lightning-30b-a3b-nvfp4",
 ];
 const EXPECTED_MANAGED_SOURCE_IDS = [
   "llama-cpp.dgx-spark-gb10.single.nemotron-3-nano-30b-a3b",
   "llama-cpp.linux-amd64-nvidia.single.nemotron-3-nano-30b-a3b",
   "local-model-profile.vllm.spark.v1",
   "vllm.dgx-spark-gb10.dual.deepseek-v4-flash-0731",
+  "vllm.dgx-spark-gb10.single.nemotron-3.5-lightning-30b-a3b-nvfp4",
   "llama-cpp.nemotron-3-nano-30b-a3b.spark-single.v1",
   "vllm.deepseek-v4-flash-0731.spark-dual.v1",
+  "vllm.nemotron-3.5-lightning-30b-a3b-nvfp4.spark-single.v1",
   "vllm.qwen3-6-35b-a3b-nvfp4.spark-single.v1",
 ];