refactor(inference): make catalog own vLLM profiles and refresh llama.cpp pins

managed-inference/images/llama-cpp/Dockerfile

@@ -4,7 +4,7 @@
 ARG CUDA_DEV_IMAGE
 ARG CUDA_RUNTIME_IMAGE
 
 FROM ${CUDA_DEV_IMAGE} AS build
 
 ARG LLAMA_CPP_REVISION
 ARG LLAMA_CPP_ARCHIVE_SHA256
@@ -29,11 +29,11 @@
       build-essential=12.10ubuntu1 \
       ca-certificates=20260601~24.04.1 \
       cmake=3.28.3-1build7 \
-      curl=8.5.0-2ubuntu10.11 \
+      curl=8.5.0-2ubuntu10.12 \
       g++-14=14.2.0-4ubuntu2~24.04.1 \
       gcc-14=14.2.0-4ubuntu2~24.04.1 \
       golang-go=2:1.22~2build1 \
-      libcurl4-openssl-dev=8.5.0-2ubuntu10.11 \
+      libcurl4-openssl-dev=8.5.0-2ubuntu10.12 \
       libssl-dev=3.0.13-0ubuntu3.12 \
     && rm -rf /var/lib/apt/lists/*
 
@@ -98,7 +98,7 @@
     && chmod 0555 /opt/llama.cpp/licenses/go \
     && chmod 0444 /opt/llama.cpp/licenses/go/copyright
 
 FROM ${CUDA_RUNTIME_IMAGE} AS runtime
 
 ARG CUDA_DEV_IMAGE
 ARG CUDA_RUNTIME_IMAGE
@@ -123,7 +123,7 @@
 RUN apt-get update \
     && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
       ca-certificates=20260601~24.04.1 \
-      libcurl4t64=8.5.0-2ubuntu10.11 \
+      libcurl4t64=8.5.0-2ubuntu10.12 \
       libgomp1=14.2.0-4ubuntu2~24.04.1 \
     && groupadd --gid "${RUNTIME_GID}" nemoclaw-llama \
     && useradd --uid "${RUNTIME_UID}" --gid "${RUNTIME_GID}" \

managed-inference/images/llama-cpp/image.yaml

@@ -115,11 +115,11 @@ spec:
       build-essential: 12.10ubuntu1
       ca-certificates: 20260601~24.04.1
       cmake: 3.28.3-1build7
-      curl: 8.5.0-2ubuntu10.11
+      curl: 8.5.0-2ubuntu10.12
       g++-14: 14.2.0-4ubuntu2~24.04.1
       gcc-14: 14.2.0-4ubuntu2~24.04.1
       golang-go: 2:1.22~2build1
-      libcurl4-openssl-dev: 8.5.0-2ubuntu10.11
+      libcurl4-openssl-dev: 8.5.0-2ubuntu10.12
       libssl-dev: 3.0.13-0ubuntu3.12
     cmake:
       ggmlBackendDl: true
@@ -162,7 +162,7 @@ spec:
       - /usr/bin/sh
     packages:
       ca-certificates: 20260601~24.04.1
-      libcurl4t64: 8.5.0-2ubuntu10.11
+      libcurl4t64: 8.5.0-2ubuntu10.12
       libgomp1: 14.2.0-4ubuntu2~24.04.1
     writablePaths:
       - /tmp

managed-inference/models/vllm.deepseek-r1-distill-llama-70b.v1.yaml

@@ -0,0 +1,28 @@
+# 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: ServingModel
+metadata:
+  id: vllm.deepseek-r1-distill-llama-70b.v1
+  displayName: DeepSeek-R1 Distill Llama 70B
+spec:
+  id: deepseek-ai/DeepSeek-R1-Distill-Llama-70B
+  revision: b1c0b44b4369b597ad119a196caf79a9c40e141e
+  environmentValue: deepseek-r1-distill-70b
+  displayName: DeepSeek-R1 Distill Llama 70B
+  menuOrder: 20
+  servedName: deepseek-ai/DeepSeek-R1-Distill-Llama-70B
+  downloadSizeBytes: 141000000000
+  gated: true
+  installFastSafetensors: true
+  preparation:
+    ref: none/v1
+  probePolicyRef: nvidia.endpoint-validation.standard/v1
+  capabilities:
+    chatCompletions: true
+    streaming: true
+    toolCalls: true
+    structuredOutputs: false
+    reasoning: true
+    multimodal: false

managed-inference/models/vllm.deepseek-v4-flash-0731.v1.yaml

@@ -0,0 +1,46 @@
+# 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: ServingModel
+metadata:
+  id: vllm.deepseek-v4-flash-0731.v1
+  displayName: DeepSeek V4 Flash 0731
+spec:
+  id: deepseek-ai/DeepSeek-V4-Flash-0731
+  revision: 9e165c30e2704aec5d9d593cce3eebd58bbef1cb
+  environmentValue: deepseek-v4-flash-0731
+  displayName: DeepSeek V4 Flash 0731
+  menuOrder: 90
+  servedName: deepseek-v4-flash-0731
+  downloadSizeBytes: 166898661074
+  gated: false
+  installFastSafetensors: false
+  preparation:
+    ref: snapshot-copy-and-exact-text-replacement/v1
+    snapshotCopy:
+      sourcePath: encoding/encoding_dsv4.py
+      digest: sha256:abc0d26120250dda0ae077dc64aa28836026e61e970854aaeb792445e6a0dde6
+      targetPath: /usr/local/lib/python3.12/dist-packages/vllm/tokenizers/deepseek_v4_encoding.py
+    exactTextReplacement:
+      targetPath: /usr/local/lib/python3.12/dist-packages/vllm/tokenizers/deepseek_v4.py
+      expectedText: |2-
+                    elif reasoning_effort in ("max", "xhigh"):
+                        reasoning_effort = "max"
+                    else:
+                        reasoning_effort = "high"
+      replacementText: |2-
+                    elif reasoning_effort in ("max", "xhigh"):
+                        reasoning_effort = "max"
+                    elif reasoning_effort == "high":
+                        reasoning_effort = "high"
+                    else:
+                        reasoning_effort = "low"
+  probePolicyRef: nvidia.endpoint-validation.standard/v1
+  capabilities:
+    chatCompletions: true
+    streaming: true
+    toolCalls: true
+    structuredOutputs: false
+    reasoning: true
+    multimodal: false

managed-inference/models/vllm.deepseek-v4-flash.v1.yaml

@@ -0,0 +1,28 @@
+# 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: ServingModel
+metadata:
+  id: vllm.deepseek-v4-flash.v1
+  displayName: DeepSeek V4 Flash
+spec:
+  id: deepseek-ai/DeepSeek-V4-Flash
+  revision: 60d8d70770c6776ff598c94bb586a859a38244f1
+  environmentValue: deepseek-v4-flash
+  displayName: DeepSeek V4 Flash
+  menuOrder: 40
+  servedName: deepseek-ai/DeepSeek-V4-Flash
+  downloadSizeBytes: 352381000000
+  gated: false
+  installFastSafetensors: true
+  preparation:
+    ref: none/v1
+  probePolicyRef: nvidia.endpoint-validation.extended/v1
+  capabilities:
+    chatCompletions: true
+    streaming: true
+    toolCalls: true
+    structuredOutputs: false
+    reasoning: true
+    multimodal: false

managed-inference/models/vllm.muse-glimmer-30b-nvfp4-w4a4.v1.yaml

@@ -0,0 +1,28 @@
+# 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: ServingModel
+metadata:
+  id: vllm.muse-glimmer-30b-nvfp4-w4a4.v1
+  displayName: Muse Glimmer 30B NVFP4 W4A4
+spec:
+  id: Inferact/Muse-Glimmer-30B-NVFP4-W4A4
+  revision: d35cb79050f419c457611b1cee5c5d15b176f285
+  environmentValue: muse-glimmer-30b
+  displayName: Muse Glimmer 30B NVFP4 W4A4 [Experimental]
+  menuOrder: 70
+  servedName: muse-glimmer
+  downloadSizeBytes: 25447097878
+  gated: false
+  installFastSafetensors: false
+  preparation:
+    ref: none/v1
+  probePolicyRef: nvidia.endpoint-validation.standard/v1
+  capabilities:
+    chatCompletions: true
+    streaming: true
+    toolCalls: true
+    structuredOutputs: false
+    reasoning: true
+    multimodal: false

managed-inference/models/vllm.nemotron-3-nano-4b-fp8.v1.yaml

@@ -0,0 +1,28 @@
+# 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: ServingModel
+metadata:
+  id: vllm.nemotron-3-nano-4b-fp8.v1
+  displayName: NVIDIA Nemotron-3 Nano 4B FP8
+spec:
+  id: nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8
+  revision: 3fe6dab75665a93884214ad4b1b95cf02717d081
+  environmentValue: nemotron-3-nano-4b
+  displayName: NVIDIA Nemotron-3 Nano 4B FP8
+  menuOrder: 30
+  servedName: nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8
+  downloadSizeBytes: 5280000000
+  gated: false
+  installFastSafetensors: true
+  preparation:
+    ref: none/v1
+  probePolicyRef: nvidia.endpoint-validation.standard/v1
+  capabilities:
+    chatCompletions: true
+    streaming: true
+    toolCalls: true
+    structuredOutputs: false
+    reasoning: true
+    multimodal: false

managed-inference/models/vllm.nemotron-3-ultra-550b-a55b-nvfp4.v1.yaml

@@ -0,0 +1,28 @@
+# 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: ServingModel
+metadata:
+  id: vllm.nemotron-3-ultra-550b-a55b-nvfp4.v1
+  displayName: NVIDIA Nemotron 3 Ultra 550B NVFP4
+spec:
+  id: nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4
+  revision: 183968f87ae4cedce3039313cac1fd43d112c578
+  environmentValue: nemotron-3-ultra-550b-a55b
+  displayName: NVIDIA Nemotron 3 Ultra 550B NVFP4
+  menuOrder: 50
+  servedName: nvidia/nemotron-3-ultra-550b-a55b
+  downloadSizeBytes: 352381245521
+  gated: false
+  installFastSafetensors: false
+  preparation:
+    ref: none/v1
+  probePolicyRef: nvidia.endpoint-validation.standard/v1
+  capabilities:
+    chatCompletions: true
+    streaming: true
+    toolCalls: true
+    structuredOutputs: false
+    reasoning: true
+    multimodal: false

managed-inference/models/vllm.nemotron-3.5-lightning-30b-a3b-nvfp4.v1.yaml

@@ -0,0 +1,28 @@
+# 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: ServingModel
+metadata:
+  id: vllm.nemotron-3.5-lightning-30b-a3b-nvfp4.v1
+  displayName: NVIDIA Nemotron 3.5 Lightning 30B-A3B NVFP4
+spec:
+  id: nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4
+  revision: 0dcd680e5585c791728c83342b311d0a0026dbeb
+  environmentValue: nemotron-3.5-lightning-30b
+  displayName: NVIDIA Nemotron 3.5 Lightning 30B-A3B NVFP4 [Experimental]
+  menuOrder: 80
+  servedName: nvidia-nemotron-3.5-lightning-30b-a3b-nvfp4
+  downloadSizeBytes: 21561882284
+  gated: false
+  installFastSafetensors: false
+  preparation:
+    ref: none/v1
+  probePolicyRef: nvidia.endpoint-validation.standard/v1
+  capabilities:
+    chatCompletions: true
+    streaming: true
+    toolCalls: true
+    structuredOutputs: false
+    reasoning: true
+    multimodal: false

managed-inference/models/vllm.qwen3-6-27b-fp8.v1.yaml

@@ -0,0 +1,28 @@
+# 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: ServingModel
+metadata:
+  id: vllm.qwen3-6-27b-fp8.v1
+  displayName: Qwen3.6 27B FP8
+spec:
+  id: Qwen/Qwen3.6-27B-FP8
+  revision: e89b16ebf1988b3d6befa7de50abc2d76f26eb09
+  environmentValue: qwen3.6-27b
+  displayName: Qwen3.6 27B FP8
+  menuOrder: 10
+  servedName: Qwen/Qwen3.6-27B-FP8
+  downloadSizeBytes: 30900000000
+  gated: false
+  installFastSafetensors: true
+  preparation:
+    ref: none/v1
+  probePolicyRef: nvidia.endpoint-validation.standard/v1
+  capabilities:
+    chatCompletions: true
+    streaming: true
+    toolCalls: true
+    structuredOutputs: false
+    reasoning: true
+    multimodal: false

managed-inference/models/vllm.qwen3-6-35b-a3b-nvfp4.v1.yaml

@@ -0,0 +1,28 @@
+# 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: ServingModel
+metadata:
+  id: vllm.qwen3-6-35b-a3b-nvfp4.v1
+  displayName: Qwen3.6 35B-A3B NVFP4
+spec:
+  id: nvidia/Qwen3.6-35B-A3B-NVFP4
+  revision: 491c2f1ea524c639598bf8fa787a93fed5a6fbce
+  environmentValue: qwen3.6-35b-a3b-nvfp4
+  displayName: Qwen3.6 35B-A3B NVFP4
+  menuOrder: 60
+  servedName: nvidia/Qwen3.6-35B-A3B-NVFP4
+  downloadSizeBytes: 23500000000
+  gated: false
+  installFastSafetensors: true
+  preparation:
+    ref: none/v1
+  probePolicyRef: nvidia.endpoint-validation.standard/v1
+  capabilities:
+    chatCompletions: true
+    streaming: true
+    toolCalls: true
+    structuredOutputs: false
+    reasoning: true
+    multimodal: false

managed-inference/presets/llama-cpp.dgx-spark-gb10.single.muse-glimmer-30b.yaml

@@ -8,6 +8,9 @@ metadata:
   id: llama-cpp.dgx-spark-gb10.single.muse-glimmer-30b
   displayName: Meta Muse Glimmer 30B on one DGX Spark
   supportState: experimental
+  validation:
+    level: schema
+    evidence: managed-inference-catalog-compiler-v1
 
 spec:
   selection: automatic

managed-inference/presets/llama-cpp.dgx-spark-gb10.single.nemotron-3-nano-30b-a3b.yaml

@@ -8,6 +8,9 @@ metadata:
   id: llama-cpp.dgx-spark-gb10.single.nemotron-3-nano-30b-a3b
   displayName: NVIDIA Nemotron 3 Nano 30B-A3B on one DGX Spark
   supportState: experimental
+  validation:
+    level: schema
+    evidence: managed-inference-catalog-compiler-v1
 
 spec:
   selection: automatic

managed-inference/presets/llama-cpp.linux-amd64-nvidia.single.nemotron-3-nano-30b-a3b.yaml

@@ -8,6 +8,9 @@ metadata:
   id: llama-cpp.linux-amd64-nvidia.single.nemotron-3-nano-30b-a3b
   displayName: NVIDIA Nemotron 3 Nano 30B-A3B on one Linux x86_64 NVIDIA GPU
   supportState: experimental
+  validation:
+    level: schema
+    evidence: managed-inference-catalog-compiler-v1
 
 spec:
   selection: automatic

managed-inference/presets/local-model-profile.vllm.spark.v1.yaml

@@ -7,6 +7,9 @@ kind: ServingPreset
 metadata:
   id: local-model-profile.vllm.spark.v1
   displayName: Local model profile with vLLM
+  validation:
+    level: schema
+    evidence: managed-inference-catalog-compiler-v1
 
 spec:
   selection: disabled
@@ -98,4 +101,7 @@ spec:
 
   plan:
     backend: vllm
+    platform: spark
+    interactive: false
     recipeRef: vllm.qwen3-6-35b-a3b-nvfp4.spark-single.v1
+    installPolicyRef: vllm.fixed-authenticated/v1

managed-inference/presets/vllm.dgx-spark-gb10.dual.deepseek-v4-flash-0731.yaml

@@ -8,6 +8,9 @@ metadata:
   id: vllm.dgx-spark-gb10.dual.deepseek-v4-flash-0731
   displayName: DeepSeek V4 Flash 0731 on two DGX Sparks
   supportState: experimental
+  validation:
+    level: schema
+    evidence: managed-inference-catalog-compiler-v1
 
 spec:
   selection: automatic
@@ -76,6 +79,8 @@ spec:
 
   plan:
     backend: vllm
+    platform: spark
+    interactive: false
     recipeRef: vllm.deepseek-v4-flash-0731.spark-dual.v1
     bindings:
       sparkTopology: