From 1a8f600f455a3bc71822ec541d262746a158f2d4 Mon Sep 17 00:00:00 2001 From: Julie Yaunches Date: Fri, 17 Jul 2026 20:56:11 -0400 Subject: [PATCH 1/5] docs: add v0.0.87 release notes Signed-off-by: Julie Yaunches --- docs/changelog/2026-07-17.mdx | 26 +++++++++++++++++++ docs/get-started/dgx-station-preparation.mdx | 15 ++++++++--- .../recover-rebuild-sandboxes.mdx | 2 ++ 3 files changed, 39 insertions(+), 4 deletions(-) diff --git a/docs/changelog/2026-07-17.mdx b/docs/changelog/2026-07-17.mdx index 94ef116eff7..f2ba18974cd 100644 --- a/docs/changelog/2026-07-17.mdx +++ b/docs/changelog/2026-07-17.mdx @@ -3,6 +3,32 @@ * SPDX-License-Identifier: Apache-2.0 */} +## v0.0.87 + +NemoClaw v0.0.87 adds bounded DGX Station factory-image qualification paths, makes managed Deep Agents Code startup restart-safe, and improves rebuild recovery, managed vLLM storage checks, sandbox backups, and strict-provider compatibility. + +- DGX Station Express recognizes the exact April 2026 NVIDIA Colossus BaseOS and June 2026 NVIDIA AI Developer Tools GB300 factory profiles for qualification. + The installer preserves each factory kernel, driver, Docker, and NVIDIA Container Toolkit stack, applies only the bounded access or runtime preparation required by that exact profile, and rejects identity, package, service, GPU, or runtime drift. + DGX Station remains Deferred while physical qualification continues. + For more information, refer to [Prepare DGX Station to Install NemoClaw](/user-guide/openclaw/get-started/prerequisites/dgx-station-preparation) and [Platform Support and Launch Claims](/user-guide/openclaw/reference/platform-support). +- Managed Deep Agents Code onboarding now persists the `nemoclaw-dcode-entrypoint` startup command when the OpenShell Docker driver recreates a sandbox. + The recreated container also receives the required `nproc=512:512` and `nofile=65536:65536` limits, so the managed runtime remains available after a gateway restart without weakening its process and file-descriptor boundaries. + For more information, refer to [Security Best Practices](/user-guide/deepagents/security/best-practices). +- Rebuild recovery verifies that a restored Hermes sandbox returns to healthy gateway and managed MCP state before reporting success. + OpenClaw rebuilds also clear stale managed-provider session pins after an inference switch, allowing restored sessions to use the current configured model while preserving intentional pins to other providers. + For more information, refer to [Recover and Rebuild Sandboxes](/user-guide/openclaw/manage-sandboxes/operate-sandboxes/recover-and-rebuild-sandboxes) and [Switch Inference Providers](/user-guide/openclaw/inference/manage-inference/switch-providers). +- Managed vLLM storage preflight estimates cold image and model downloads from pinned image metadata and model payload sizes. + It checks Docker storage and the Hugging Face cache separately when they use different filesystems, rechecks capacity after a cold image pull, warns and continues during express or other non-interactive setup, and requires confirmation during interactive setup. + For more information, refer to [Set Up vLLM](/user-guide/openclaw/inference/local-inference/set-up-vllm). +- Sandbox backup creation now streams archive data and validates entries incrementally instead of buffering the complete archive in host memory. + Large backups therefore retain the existing traversal checks and partial-state behavior without requiring memory proportional to the archive size. + For more information, refer to [Create and Restore Snapshots](/user-guide/openclaw/manage-sandboxes/state-and-backups/create-and-restore-snapshots). +- Hermes registers NemoClaw tools with the single function-schema envelope required by strict OpenAI-compatible providers. + Google Gemini no longer rejects the managed Hermes tool list because of a nested schema, and audio transcription retains its declared parameters. + For more information, refer to [Use Google Gemini](/user-guide/hermes/inference/hosted-inference/use-google-gemini). +- Replacement-image rebuild failures preserve bounded, redacted Docker diagnostics when process output arrives as buffered data, making host-specific build failures actionable without exposing credentials or private host paths. + For more information, refer to [Recover and Rebuild Sandboxes](/user-guide/openclaw/manage-sandboxes/operate-sandboxes/recover-and-rebuild-sandboxes). + ## v0.0.86 NemoClaw v0.0.86 enables the Station express recipe on qualified stock DGX OS GB300 systems, makes interrupted Station setup resumable, and fixes model validation, managed vLLM cache checks, sandbox builds, and upgrade guidance. diff --git a/docs/get-started/dgx-station-preparation.mdx b/docs/get-started/dgx-station-preparation.mdx index 6693056974d..09363966d05 100644 --- a/docs/get-started/dgx-station-preparation.mdx +++ b/docs/get-started/dgx-station-preparation.mdx @@ -16,7 +16,12 @@ Complete this preparation before following the Quickstart on a qualifying DGX St DGX Station GB300 express install accepts either the generic Ubuntu 24.04 ARM64 image or stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0`. A stock DGX OS image qualifies only when a safe, root-owned `/etc/dgx-release` marker identifies `DGX Server for GALAXY-GB300`. -An unknown version, malformed or unsafe marker, NVIDIA BaseOS image, or other Station generation stops before host preparation. +NemoClaw also recognizes these exact no-OTA factory profiles for bounded qualification: + +- April 2026 NVIDIA Colossus BaseOS with `DGX_PRETTY_NAME=NVIDIA DGX Server`, `DGX_SWBUILD_VERSION=7.5.0-GB300ws-GB200ws`, and `DGX_SWBUILD_DATE=2026-04-02-08-20-16`. +- June 2026 NVIDIA AI Developer Tools with `DGX_PRETTY_NAME=NVIDIA DGX GB300WS`, `DGX_SWBUILD_VERSION=7.5.0`, and `DGX_SWBUILD_DATE=2026-06-16-11-48-10`. + +An unknown version, malformed or unsafe marker, unmatched no-OTA factory image, or other Station generation stops before host preparation. On an unqualified system, set `NEMOCLAW_PROVIDER` or `NEMOCLAW_NO_EXPRESS=1` explicitly to continue without Station host automation. ## Prepare Generic Ubuntu @@ -41,11 +46,13 @@ After changing pinned packages, the installer exits with status `10`; reboot, si ## Validate Stock DGX OS -On a qualifying stock DGX OS image, express install validates the factory stack in place against the local default Docker daemon. +On a qualifying stock DGX OS or exact no-OTA factory image, express install validates the factory stack in place against the local default Docker daemon. It requires a loaded GB300 driver with zero volatile corrected and uncorrected ECC errors, active Docker and containerd services, working Docker Buildx, `nvidia-ctk`, the `nvidia.com/gpu=all` CDI device, and successful `nvidia-smi` device-visibility probes through both CDI and `--gpus all`. -This path does not install or replace host packages, enable or restart services, generate CDI configuration, or rewrite the Docker runtime. +The stock DGX OS and AI Developer Tools paths do not install or replace host packages, enable or restart services, generate CDI configuration, or rewrite the Docker runtime. +The Colossus BaseOS path requires the exact reviewed package and failed-service fingerprints, preserves the factory kernel, driver, DKMS, Docker, and NVIDIA Container Toolkit packages, and prepares only Docker access and packaged CDI. +It registers the NVIDIA Docker runtime only when the launch probe proves that the runtime is missing, and it restores the prior daemon configuration if registration fails. The preparation helper writes a private audit log under `~/station-bootstrap-logs`. -Aside from the installer's separately disclosed `docker`-group membership grant when the current trusted account needs Docker access, its only persistent stock-runtime effect is that the digest-pinned visibility-test image can remain in the Docker cache. +Aside from the installer's separately disclosed `docker`-group membership grant when the current trusted account needs Docker access, the stock DGX OS and AI Developer Tools paths leave only the digest-pinned visibility-test image in the Docker cache. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. Repeat qualification from a clean host and CI coverage remain pending. diff --git a/docs/manage-sandboxes/recover-rebuild-sandboxes.mdx b/docs/manage-sandboxes/recover-rebuild-sandboxes.mdx index 8fea4c64981..0a6e9f496f3 100644 --- a/docs/manage-sandboxes/recover-rebuild-sandboxes.mdx +++ b/docs/manage-sandboxes/recover-rebuild-sandboxes.mdx @@ -100,6 +100,8 @@ $$nemoclaw rebuild The rebuild command preserves the mounted workspace and registered policies while recreating the container. +After state restoration, NemoClaw clears a session's stale model and provider pin when it still targets the managed `inference` provider but no longer matches the configured default model. +The session then follows the current default selected through `inference set`, while sessions pinned to another provider remain unchanged. From 2aa8b1415ea51615ff6d0756749319877b34ddaf Mon Sep 17 00:00:00 2001 From: Julie Yaunches Date: Fri, 17 Jul 2026 21:05:39 -0400 Subject: [PATCH 2/5] docs: sync Station qualification matrix Signed-off-by: Julie Yaunches --- ci/platform-matrix.json | 4 ++-- docs/inference/choose-inference-provider.mdx | 2 +- docs/reference/platform-support.mdx | 4 ++-- 3 files changed, 5 insertions(+), 5 deletions(-) diff --git a/ci/platform-matrix.json b/ci/platform-matrix.json index 648a38fe3c8..9790e723d99 100644 --- a/ci/platform-matrix.json +++ b/ci/platform-matrix.json @@ -66,7 +66,7 @@ "prd_priority": "P1", "ci_tested": false, "prerequisites_notes": "Deferred. For evaluation on a qualifying DGX Station GB300, see [Additional Setup for DGX Station](prerequisites/dgx-station-preparation) before the Quickstart.", - "notes": "The PRD marks this platform as P1. Workstation form-factor with NVIDIA GPUs and the same Docker + NVIDIA Container Toolkit + CDI requirements as DGX Spark. Station remains Deferred. For canonical hardware qualification, image requirements, preparation, repair limits, and reboot handoff, see [Prepare DGX Station to Install NemoClaw](../get-started/prerequisites/dgx-station-preparation). On a Station GB300, the installer accepts either a clean generic Ubuntu 24.04 ARM64 image or stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0` when a safe, root-owned `/etc/dgx-release` marker identifies `DGX Server for GALAXY-GB300`. The stock DGX OS path validates the factory GB300 driver and volatile ECC state, Docker and containerd, Buildx, `nvidia-ctk`, CDI, and GPU visibility through both CDI and `--gpus all` before it proceeds; it does not replace host packages or rewrite the runtime. Unknown DGX OS versions, NVIDIA BaseOS images, unsafe release markers, and other Station generations stop before host preparation. The installer offers express install with the pinned `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` recipe, including an approximately 352 GB model download, without follow-up provider, model, policy, or sandbox-name choices. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Pass `--station-deepseek` to use `deepseek-ai/DeepSeek-V4-Flash` while retaining the one-confirmation express flow. The flag requires an interactive terminal, and `/dev/tty` must be available when the installer runs through `curl | bash`. For headless setup on a qualifying Station, select `NEMOCLAW_PROVIDER=install-vllm` and `NEMOCLAW_VLLM_MODEL=deepseek-v4-flash` instead. Direct managed-vLLM onboarding still defaults to `deepseek-ai/DeepSeek-V4-Flash` when no model override is set." + "notes": "The PRD marks this platform as P1. Workstation form-factor with NVIDIA GPUs and the same Docker + NVIDIA Container Toolkit + CDI requirements as DGX Spark. Station remains Deferred. For canonical hardware qualification, image requirements, preparation, repair limits, and reboot handoff, see [Prepare DGX Station to Install NemoClaw](../get-started/prerequisites/dgx-station-preparation). On a Station GB300, the installer accepts either a clean generic Ubuntu 24.04 ARM64 image or stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0` when a safe, root-owned `/etc/dgx-release` marker identifies `DGX Server for GALAXY-GB300`. It also recognizes two exact no-OTA factory profiles for bounded qualification: April 2026 NVIDIA Colossus BaseOS (`DGX_SWBUILD_VERSION=7.5.0-GB300ws-GB200ws`, `DGX_SWBUILD_DATE=2026-04-02-08-20-16`) and June 2026 NVIDIA AI Developer Tools (`DGX_SWBUILD_VERSION=7.5.0`, `DGX_SWBUILD_DATE=2026-06-16-11-48-10`). The stock DGX OS and AI Developer Tools paths validate the factory stack in place without replacing host packages, enabling or restarting services, generating CDI configuration, or rewriting the Docker runtime. The Colossus BaseOS path preserves the factory kernel, driver, DKMS, Docker, and NVIDIA Container Toolkit packages and prepares only Docker access and packaged CDI. Each path validates the factory GB300 driver and volatile ECC state, Docker and containerd, Buildx, `nvidia-ctk`, CDI, and GPU visibility through both CDI and `--gpus all` before it proceeds. Unknown DGX OS versions, unmatched no-OTA factory images, unsafe release markers, and other Station generations stop before host preparation. The installer offers express install with the pinned `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` recipe, including an approximately 352 GB model download, without follow-up provider, model, policy, or sandbox-name choices. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Pass `--station-deepseek` to use `deepseek-ai/DeepSeek-V4-Flash` while retaining the one-confirmation express flow. The flag requires an interactive terminal, and `/dev/tty` must be available when the installer runs through `curl | bash`. For headless setup on a qualifying Station, select `NEMOCLAW_PROVIDER=install-vllm` and `NEMOCLAW_VLLM_MODEL=deepseek-v4-flash` instead. Direct managed-vLLM onboarding still defaults to `deepseek-ai/DeepSeek-V4-Flash` when no model override is set." }, { "name": "NVIDIA RTX (consumer and Pro workstation GPUs)", @@ -149,7 +149,7 @@ "name": "Local vLLM (managed install/start)", "status": "caveated", "endpoint_type": "Local OpenAI-compatible", - "notes": "Appears by default on DGX Spark and qualifying DGX Station GB300 hosts. DGX Station remains Deferred. The Station path accepts a clean generic Ubuntu 24.04 ARM64 image or stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0` after the release marker and factory GPU runtime pass validation. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pins runtime images to immutable digests. DGX Spark and DGX Station models without a model-specific runtime use the `linux/arm64` digest published under `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use the matching `linux/arm64` or `linux/amd64` digest published under `nvcr.io/nvidia/vllm:26.03.post1-py3`. The DGX Station express installer selects `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` with a pinned Hugging Face revision and the multi-platform index digest published under `vllm/vllm-openai:v0.22.0`; that index resolves to the `linux/arm64` manifest on Station. Direct managed-vLLM profile defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `deepseek-ai/DeepSeek-V4-Flash`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls from `nvcr.io` require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing." + "notes": "Appears by default on DGX Spark and qualifying DGX Station GB300 hosts. DGX Station remains Deferred. The Station path accepts a clean generic Ubuntu 24.04 ARM64 image, stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0`, or the exact April 2026 NVIDIA Colossus BaseOS and June 2026 NVIDIA AI Developer Tools no-OTA factory profiles documented in [Prepare DGX Station to Install NemoClaw](/user-guide/openclaw/get-started/prerequisites/dgx-station-preparation), after the release marker and factory GPU runtime pass validation. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pins runtime images to immutable digests. DGX Spark and DGX Station models without a model-specific runtime use the `linux/arm64` digest published under `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use the matching `linux/arm64` or `linux/amd64` digest published under `nvcr.io/nvidia/vllm:26.03.post1-py3`. The DGX Station express installer selects `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` with a pinned Hugging Face revision and the multi-platform index digest published under `vllm/vllm-openai:v0.22.0`; that index resolves to the `linux/arm64` manifest on Station. Direct managed-vLLM profile defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `deepseek-ai/DeepSeek-V4-Flash`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls from `nvcr.io` require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing." } ], diff --git a/docs/inference/choose-inference-provider.mdx b/docs/inference/choose-inference-provider.mdx index a83b755927c..ba3bba979ca 100644 --- a/docs/inference/choose-inference-provider.mdx +++ b/docs/inference/choose-inference-provider.mdx @@ -32,7 +32,7 @@ Use this status table to distinguish validated provider integrations from adapte | Local Ollama | Tested with limitations | Local Ollama API | Available when Ollama is installed or running on the host. Validated default models: `qwen3.6:35b` (high VRAM), `nemotron-3-nano:30b` (medium VRAM), `qwen3.5:9b` (low VRAM fallback). | | Local NVIDIA NIM | Experimental | Local OpenAI-compatible | Requires `NEMOCLAW_EXPERIMENTAL=1` and a NIM-capable NVIDIA GPU. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence with `assertCdiNvidiaGpuSpecPresent`, `src/lib/onboard/fatal-runtime-preflight.ts`). NIM images pull from `nvcr.io` and require NGC registry login. NemoClaw gates this path behind the experimental flag because it does not auto-select a NIM image for the host today. You must explicitly pick from the validated image list. On Linux arm64 DGX Spark and DGX Station hosts, onboarding warns that some NIM images may not publish a `linux/arm64` manifest; the warning is advisory, and the selected image pull can still fail when the registry has no matching platform manifest. Managed vLLM has host-specific default models and is not gated on the same boxes. Validated images referenced in `src/lib/inference/config.ts` and `nemoclaw/src/index.ts`: `nvidia/nemotron-3-super-120b-a12b` (default cloud model), `nvidia/nemotron-3-nano-30b-a3b`, `nvidia/llama-3.3-nemotron-super-49b-v1.5`. | | Local vLLM (already running) | Tested with limitations | Local OpenAI-compatible | Appears in the onboarding menu when NemoClaw detects a server already on `localhost:8000`. No flag required. Model is whatever the existing server serves. | -| Local vLLM (managed install/start) | Tested with limitations | Local OpenAI-compatible | Appears by default on DGX Spark and qualifying DGX Station GB300 hosts. DGX Station remains Deferred. The Station path accepts a clean generic Ubuntu 24.04 ARM64 image or stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0` after the release marker and factory GPU runtime pass validation. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pins runtime images to immutable digests. DGX Spark and DGX Station models without a model-specific runtime use the `linux/arm64` digest published under `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use the matching `linux/arm64` or `linux/amd64` digest published under `nvcr.io/nvidia/vllm:26.03.post1-py3`. The DGX Station express installer selects `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` with a pinned Hugging Face revision and the multi-platform index digest published under `vllm/vllm-openai:v0.22.0`; that index resolves to the `linux/arm64` manifest on Station. Direct managed-vLLM profile defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `deepseek-ai/DeepSeek-V4-Flash`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls from `nvcr.io` require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing. | +| Local vLLM (managed install/start) | Tested with limitations | Local OpenAI-compatible | Appears by default on DGX Spark and qualifying DGX Station GB300 hosts. DGX Station remains Deferred. The Station path accepts a clean generic Ubuntu 24.04 ARM64 image, stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0`, or the exact April 2026 NVIDIA Colossus BaseOS and June 2026 NVIDIA AI Developer Tools no-OTA factory profiles documented in [Prepare DGX Station to Install NemoClaw](/user-guide/openclaw/get-started/prerequisites/dgx-station-preparation), after the release marker and factory GPU runtime pass validation. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pins runtime images to immutable digests. DGX Spark and DGX Station models without a model-specific runtime use the `linux/arm64` digest published under `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use the matching `linux/arm64` or `linux/amd64` digest published under `nvcr.io/nvidia/vllm:26.03.post1-py3`. The DGX Station express installer selects `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` with a pinned Hugging Face revision and the multi-platform index digest published under `vllm/vllm-openai:v0.22.0`; that index resolves to the `linux/arm64` manifest on Station. Direct managed-vLLM profile defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `deepseek-ai/DeepSeek-V4-Flash`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls from `nvcr.io` require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing. | {/* provider-status:end */} ## Hosted Providers diff --git a/docs/reference/platform-support.mdx b/docs/reference/platform-support.mdx index 66a976b0bbf..ff7c57b0954 100644 --- a/docs/reference/platform-support.mdx +++ b/docs/reference/platform-support.mdx @@ -79,7 +79,7 @@ For install requirements and the shorter setup-oriented platform view, refer to | OS | Container runtime | Status | PRD priority | CI | Notes | |----|-------------------|--------|--------------|----|-------| | DGX OS (Spark) | Docker | Tested | P1 | Yes | Use the standard installer and `$$nemoclaw onboard`. For an end-to-end walkthrough with local inference, see the [NVIDIA Spark playbook](https://build.nvidia.com/spark/nemoclaw). | -| DGX OS (Station) | Docker | Deferred | P1 | No | The PRD marks this platform as P1. Workstation form-factor with NVIDIA GPUs and the same Docker + NVIDIA Container Toolkit + CDI requirements as DGX Spark. Station remains Deferred. For canonical hardware qualification, image requirements, preparation, repair limits, and reboot handoff, see [Prepare DGX Station to Install NemoClaw](../get-started/prerequisites/dgx-station-preparation). On a Station GB300, the installer accepts either a clean generic Ubuntu 24.04 ARM64 image or stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0` when a safe, root-owned `/etc/dgx-release` marker identifies `DGX Server for GALAXY-GB300`. The stock DGX OS path validates the factory GB300 driver and volatile ECC state, Docker and containerd, Buildx, `nvidia-ctk`, CDI, and GPU visibility through both CDI and `--gpus all` before it proceeds; it does not replace host packages or rewrite the runtime. Unknown DGX OS versions, NVIDIA BaseOS images, unsafe release markers, and other Station generations stop before host preparation. The installer offers express install with the pinned `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` recipe, including an approximately 352 GB model download, without follow-up provider, model, policy, or sandbox-name choices. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Pass `--station-deepseek` to use `deepseek-ai/DeepSeek-V4-Flash` while retaining the one-confirmation express flow. The flag requires an interactive terminal, and `/dev/tty` must be available when the installer runs through `curl \| bash`. For headless setup on a qualifying Station, select `NEMOCLAW_PROVIDER=install-vllm` and `NEMOCLAW_VLLM_MODEL=deepseek-v4-flash` instead. Direct managed-vLLM onboarding still defaults to `deepseek-ai/DeepSeek-V4-Flash` when no model override is set. | +| DGX OS (Station) | Docker | Deferred | P1 | No | The PRD marks this platform as P1. Workstation form-factor with NVIDIA GPUs and the same Docker + NVIDIA Container Toolkit + CDI requirements as DGX Spark. Station remains Deferred. For canonical hardware qualification, image requirements, preparation, repair limits, and reboot handoff, see [Prepare DGX Station to Install NemoClaw](../get-started/prerequisites/dgx-station-preparation). On a Station GB300, the installer accepts either a clean generic Ubuntu 24.04 ARM64 image or stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0` when a safe, root-owned `/etc/dgx-release` marker identifies `DGX Server for GALAXY-GB300`. It also recognizes two exact no-OTA factory profiles for bounded qualification: April 2026 NVIDIA Colossus BaseOS (`DGX_SWBUILD_VERSION=7.5.0-GB300ws-GB200ws`, `DGX_SWBUILD_DATE=2026-04-02-08-20-16`) and June 2026 NVIDIA AI Developer Tools (`DGX_SWBUILD_VERSION=7.5.0`, `DGX_SWBUILD_DATE=2026-06-16-11-48-10`). The stock DGX OS and AI Developer Tools paths validate the factory stack in place without replacing host packages, enabling or restarting services, generating CDI configuration, or rewriting the Docker runtime. The Colossus BaseOS path preserves the factory kernel, driver, DKMS, Docker, and NVIDIA Container Toolkit packages and prepares only Docker access and packaged CDI. Each path validates the factory GB300 driver and volatile ECC state, Docker and containerd, Buildx, `nvidia-ctk`, CDI, and GPU visibility through both CDI and `--gpus all` before it proceeds. Unknown DGX OS versions, unmatched no-OTA factory images, unsafe release markers, and other Station generations stop before host preparation. The installer offers express install with the pinned `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` recipe, including an approximately 352 GB model download, without follow-up provider, model, policy, or sandbox-name choices. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Pass `--station-deepseek` to use `deepseek-ai/DeepSeek-V4-Flash` while retaining the one-confirmation express flow. The flag requires an interactive terminal, and `/dev/tty` must be available when the installer runs through `curl \| bash`. For headless setup on a qualifying Station, select `NEMOCLAW_PROVIDER=install-vllm` and `NEMOCLAW_VLLM_MODEL=deepseek-v4-flash` instead. Direct managed-vLLM onboarding still defaults to `deepseek-ai/DeepSeek-V4-Flash` when no model override is set. | | Linux | Docker | Tested | P0 | Yes | Primary tested path. Ubuntu 24.04 has host-level onboarding validation. A digest-pinned Ubuntu 26.04 userspace lane builds the CLI and runs preflight, installer, and platform contracts on eligible main pushes; Docker-host, AppArmor, Landlock, and live onboarding validation on 26.04 remain pending. Other distros (Ubuntu 22.04, Fedora, Rocky, Alma, NixOS, Arch) may work but are not validated. | | macOS (Apple Silicon) | Colima, Docker Desktop | Tested with limitations | P0 | Yes | Start the container runtime (Colima or Docker Desktop) before running the installer. Homebrew Colima users must install both Colima and the Docker CLI (`brew install colima docker`) before `docker info` can work. Xcode Command Line Tools (`xcode-select --install`) are typically required for Node native modules during install. NemoClaw recommends them but does not enforce them during preflight. | | NVIDIA RTX (consumer and Pro workstation GPUs) | Docker | Deferred | P1 | No | The PRD marks this platform as P1. Covers RTX consumer cards and RTX Pro workstation cards on Linux hosts that meet the generic-Linux-GPU requirements (NVIDIA Container Toolkit + CDI present). The provider menu emits managed vLLM behind `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm` for this host class today; the end-to-end onboard path on this hardware is not yet validated in CI. | @@ -104,7 +104,7 @@ NemoClaw routes inference through the OpenShell gateway. Each row below is a pro | Local Ollama | Tested with limitations | Local Ollama API | Available when Ollama is installed or running on the host. Validated default models: `qwen3.6:35b` (high VRAM), `nemotron-3-nano:30b` (medium VRAM), `qwen3.5:9b` (low VRAM fallback). | | Local NVIDIA NIM | Experimental | Local OpenAI-compatible | Requires `NEMOCLAW_EXPERIMENTAL=1` and a NIM-capable NVIDIA GPU. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence with `assertCdiNvidiaGpuSpecPresent`, `src/lib/onboard/fatal-runtime-preflight.ts`). NIM images pull from `nvcr.io` and require NGC registry login. NemoClaw gates this path behind the experimental flag because it does not auto-select a NIM image for the host today. You must explicitly pick from the validated image list. On Linux arm64 DGX Spark and DGX Station hosts, onboarding warns that some NIM images may not publish a `linux/arm64` manifest; the warning is advisory, and the selected image pull can still fail when the registry has no matching platform manifest. Managed vLLM has host-specific default models and is not gated on the same boxes. Validated images referenced in `src/lib/inference/config.ts` and `nemoclaw/src/index.ts`: `nvidia/nemotron-3-super-120b-a12b` (default cloud model), `nvidia/nemotron-3-nano-30b-a3b`, `nvidia/llama-3.3-nemotron-super-49b-v1.5`. | | Local vLLM (already running) | Tested with limitations | Local OpenAI-compatible | Appears in the onboarding menu when NemoClaw detects a server already on `localhost:8000`. No flag required. Model is whatever the existing server serves. | -| Local vLLM (managed install/start) | Tested with limitations | Local OpenAI-compatible | Appears by default on DGX Spark and qualifying DGX Station GB300 hosts. DGX Station remains Deferred. The Station path accepts a clean generic Ubuntu 24.04 ARM64 image or stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0` after the release marker and factory GPU runtime pass validation. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pins runtime images to immutable digests. DGX Spark and DGX Station models without a model-specific runtime use the `linux/arm64` digest published under `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use the matching `linux/arm64` or `linux/amd64` digest published under `nvcr.io/nvidia/vllm:26.03.post1-py3`. The DGX Station express installer selects `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` with a pinned Hugging Face revision and the multi-platform index digest published under `vllm/vllm-openai:v0.22.0`; that index resolves to the `linux/arm64` manifest on Station. Direct managed-vLLM profile defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `deepseek-ai/DeepSeek-V4-Flash`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls from `nvcr.io` require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing. | +| Local vLLM (managed install/start) | Tested with limitations | Local OpenAI-compatible | Appears by default on DGX Spark and qualifying DGX Station GB300 hosts. DGX Station remains Deferred. The Station path accepts a clean generic Ubuntu 24.04 ARM64 image, stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0`, or the exact April 2026 NVIDIA Colossus BaseOS and June 2026 NVIDIA AI Developer Tools no-OTA factory profiles documented in [Prepare DGX Station to Install NemoClaw](/user-guide/openclaw/get-started/prerequisites/dgx-station-preparation), after the release marker and factory GPU runtime pass validation. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pins runtime images to immutable digests. DGX Spark and DGX Station models without a model-specific runtime use the `linux/arm64` digest published under `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use the matching `linux/arm64` or `linux/amd64` digest published under `nvcr.io/nvidia/vllm:26.03.post1-py3`. The DGX Station express installer selects `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` with a pinned Hugging Face revision and the multi-platform index digest published under `vllm/vllm-openai:v0.22.0`; that index resolves to the `linux/arm64` manifest on Station. Direct managed-vLLM profile defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `deepseek-ai/DeepSeek-V4-Flash`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls from `nvcr.io` require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing. | {/* provider-status-full:end */} ## Messaging Integrations From 59b0a6c25440428fe6f1430a87f1864aa6fa3ee4 Mon Sep 17 00:00:00 2001 From: Julie Yaunches Date: Fri, 17 Jul 2026 21:07:52 -0400 Subject: [PATCH 3/5] docs: note Station resume receipt fix Signed-off-by: Julie Yaunches --- docs/changelog/2026-07-17.mdx | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/docs/changelog/2026-07-17.mdx b/docs/changelog/2026-07-17.mdx index f2ba18974cd..5f3afb1f1e7 100644 --- a/docs/changelog/2026-07-17.mdx +++ b/docs/changelog/2026-07-17.mdx @@ -5,12 +5,15 @@ ## v0.0.87 -NemoClaw v0.0.87 adds bounded DGX Station factory-image qualification paths, makes managed Deep Agents Code startup restart-safe, and improves rebuild recovery, managed vLLM storage checks, sandbox backups, and strict-provider compatibility. +NemoClaw v0.0.87 adds bounded DGX Station factory-image qualification paths, fixes Station post-reboot resume, makes managed Deep Agents Code startup restart-safe, and improves rebuild recovery, managed vLLM storage checks, sandbox backups, and strict-provider compatibility. - DGX Station Express recognizes the exact April 2026 NVIDIA Colossus BaseOS and June 2026 NVIDIA AI Developer Tools GB300 factory profiles for qualification. The installer preserves each factory kernel, driver, Docker, and NVIDIA Container Toolkit stack, applies only the bounded access or runtime preparation required by that exact profile, and rejects identity, package, service, GPU, or runtime drift. DGX Station remains Deferred while physical qualification continues. For more information, refer to [Prepare DGX Station to Install NemoClaw](/user-guide/openclaw/get-started/prerequisites/dgx-station-preparation) and [Platform Support and Launch Claims](/user-guide/openclaw/reference/platform-support). +- Station Express onboarding now accepts both the current six-field installer resume receipt and the legacy three-field format after host preparation requires a reboot. + The current format validates the agent, sandbox name, and policy tier in addition to the pinned revision, model, and receipt generation, while malformed, unsupported, or extended receipts remain fail-closed for troubleshooting. + For more information, refer to [Prepare DGX Station to Install NemoClaw](/user-guide/openclaw/get-started/prerequisites/dgx-station-preparation). - Managed Deep Agents Code onboarding now persists the `nemoclaw-dcode-entrypoint` startup command when the OpenShell Docker driver recreates a sandbox. The recreated container also receives the required `nproc=512:512` and `nofile=65536:65536` limits, so the managed runtime remains available after a gateway restart without weakening its process and file-descriptor boundaries. For more information, refer to [Security Best Practices](/user-guide/deepagents/security/best-practices). From 2fa17d80878482eeb8e69b133c0171f164464dad Mon Sep 17 00:00:00 2001 From: Julie Yaunches Date: Fri, 17 Jul 2026 21:13:21 -0400 Subject: [PATCH 4/5] docs: clarify Station factory boundaries Signed-off-by: Julie Yaunches --- ci/platform-matrix.json | 2 +- docs/get-started/dgx-station-preparation.mdx | 5 +++-- docs/reference/platform-support.mdx | 2 +- 3 files changed, 5 insertions(+), 4 deletions(-) diff --git a/ci/platform-matrix.json b/ci/platform-matrix.json index 9790e723d99..2303386db17 100644 --- a/ci/platform-matrix.json +++ b/ci/platform-matrix.json @@ -66,7 +66,7 @@ "prd_priority": "P1", "ci_tested": false, "prerequisites_notes": "Deferred. For evaluation on a qualifying DGX Station GB300, see [Additional Setup for DGX Station](prerequisites/dgx-station-preparation) before the Quickstart.", - "notes": "The PRD marks this platform as P1. Workstation form-factor with NVIDIA GPUs and the same Docker + NVIDIA Container Toolkit + CDI requirements as DGX Spark. Station remains Deferred. For canonical hardware qualification, image requirements, preparation, repair limits, and reboot handoff, see [Prepare DGX Station to Install NemoClaw](../get-started/prerequisites/dgx-station-preparation). On a Station GB300, the installer accepts either a clean generic Ubuntu 24.04 ARM64 image or stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0` when a safe, root-owned `/etc/dgx-release` marker identifies `DGX Server for GALAXY-GB300`. It also recognizes two exact no-OTA factory profiles for bounded qualification: April 2026 NVIDIA Colossus BaseOS (`DGX_SWBUILD_VERSION=7.5.0-GB300ws-GB200ws`, `DGX_SWBUILD_DATE=2026-04-02-08-20-16`) and June 2026 NVIDIA AI Developer Tools (`DGX_SWBUILD_VERSION=7.5.0`, `DGX_SWBUILD_DATE=2026-06-16-11-48-10`). The stock DGX OS and AI Developer Tools paths validate the factory stack in place without replacing host packages, enabling or restarting services, generating CDI configuration, or rewriting the Docker runtime. The Colossus BaseOS path preserves the factory kernel, driver, DKMS, Docker, and NVIDIA Container Toolkit packages and prepares only Docker access and packaged CDI. Each path validates the factory GB300 driver and volatile ECC state, Docker and containerd, Buildx, `nvidia-ctk`, CDI, and GPU visibility through both CDI and `--gpus all` before it proceeds. Unknown DGX OS versions, unmatched no-OTA factory images, unsafe release markers, and other Station generations stop before host preparation. The installer offers express install with the pinned `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` recipe, including an approximately 352 GB model download, without follow-up provider, model, policy, or sandbox-name choices. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Pass `--station-deepseek` to use `deepseek-ai/DeepSeek-V4-Flash` while retaining the one-confirmation express flow. The flag requires an interactive terminal, and `/dev/tty` must be available when the installer runs through `curl | bash`. For headless setup on a qualifying Station, select `NEMOCLAW_PROVIDER=install-vllm` and `NEMOCLAW_VLLM_MODEL=deepseek-v4-flash` instead. Direct managed-vLLM onboarding still defaults to `deepseek-ai/DeepSeek-V4-Flash` when no model override is set." + "notes": "The PRD marks this platform as P1. Workstation form-factor with NVIDIA GPUs and the same Docker + NVIDIA Container Toolkit + CDI requirements as DGX Spark. Station remains Deferred. For canonical hardware qualification, image requirements, preparation, repair limits, and reboot handoff, see [Prepare DGX Station to Install NemoClaw](../get-started/prerequisites/dgx-station-preparation). On a Station GB300, the installer accepts either a clean generic Ubuntu 24.04 ARM64 image or stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0` when a safe, root-owned `/etc/dgx-release` marker identifies `DGX Server for GALAXY-GB300`. It also recognizes two exact no-OTA factory profiles for bounded qualification. Both require `DGX_PLATFORM=DGX Server for GALAXY-GB300` and require `DGX_OTA_VERSION` and `DGX_OTA_DATE` to be absent. The April 2026 NVIDIA Colossus BaseOS profile requires `DGX_PRETTY_NAME=NVIDIA DGX Server`, `DGX_SWBUILD_VERSION=7.5.0-GB300ws-GB200ws`, and `DGX_SWBUILD_DATE=2026-04-02-08-20-16`; the June 2026 NVIDIA AI Developer Tools profile requires `DGX_PRETTY_NAME=NVIDIA DGX GB300WS`, `DGX_SWBUILD_VERSION=7.5.0`, and `DGX_SWBUILD_DATE=2026-06-16-11-48-10`. The stock DGX OS and AI Developer Tools paths validate the factory stack in place without replacing host packages, enabling or restarting services, generating CDI configuration, or rewriting the Docker runtime. The Colossus BaseOS path preserves the factory kernel, driver, DKMS, Docker, and NVIDIA Container Toolkit packages and prepares only Docker access and packaged CDI. Only the Colossus BaseOS path registers the NVIDIA Docker runtime when the launch probe proves it is missing, with restoration of the prior daemon configuration if registration fails. Each path validates the factory GB300 driver and volatile ECC state, Docker and containerd, Buildx, `nvidia-ctk`, CDI, and GPU visibility through both CDI and `--gpus all` before it proceeds. Unknown DGX OS versions, unmatched no-OTA factory images, unsafe release markers, and other Station generations stop before host preparation. The installer offers express install with the pinned `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` recipe, including an approximately 352 GB model download, without follow-up provider, model, policy, or sandbox-name choices. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Pass `--station-deepseek` to use `deepseek-ai/DeepSeek-V4-Flash` while retaining the one-confirmation express flow. The flag requires an interactive terminal, and `/dev/tty` must be available when the installer runs through `curl | bash`. For headless setup on a qualifying Station, select `NEMOCLAW_PROVIDER=install-vllm` and `NEMOCLAW_VLLM_MODEL=deepseek-v4-flash` instead. Direct managed-vLLM onboarding still defaults to `deepseek-ai/DeepSeek-V4-Flash` when no model override is set." }, { "name": "NVIDIA RTX (consumer and Pro workstation GPUs)", diff --git a/docs/get-started/dgx-station-preparation.mdx b/docs/get-started/dgx-station-preparation.mdx index 09363966d05..106b58e47ea 100644 --- a/docs/get-started/dgx-station-preparation.mdx +++ b/docs/get-started/dgx-station-preparation.mdx @@ -16,7 +16,8 @@ Complete this preparation before following the Quickstart on a qualifying DGX St DGX Station GB300 express install accepts either the generic Ubuntu 24.04 ARM64 image or stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0`. A stock DGX OS image qualifies only when a safe, root-owned `/etc/dgx-release` marker identifies `DGX Server for GALAXY-GB300`. -NemoClaw also recognizes these exact no-OTA factory profiles for bounded qualification: +NemoClaw also recognizes these exact no-OTA factory profiles for bounded qualification. +Each profile requires `DGX_PLATFORM=DGX Server for GALAXY-GB300`, requires both `DGX_OTA_VERSION` and `DGX_OTA_DATE` to be absent, and requires every profile-specific field below: - April 2026 NVIDIA Colossus BaseOS with `DGX_PRETTY_NAME=NVIDIA DGX Server`, `DGX_SWBUILD_VERSION=7.5.0-GB300ws-GB200ws`, and `DGX_SWBUILD_DATE=2026-04-02-08-20-16`. - June 2026 NVIDIA AI Developer Tools with `DGX_PRETTY_NAME=NVIDIA DGX GB300WS`, `DGX_SWBUILD_VERSION=7.5.0`, and `DGX_SWBUILD_DATE=2026-06-16-11-48-10`. @@ -50,7 +51,7 @@ On a qualifying stock DGX OS or exact no-OTA factory image, express install vali It requires a loaded GB300 driver with zero volatile corrected and uncorrected ECC errors, active Docker and containerd services, working Docker Buildx, `nvidia-ctk`, the `nvidia.com/gpu=all` CDI device, and successful `nvidia-smi` device-visibility probes through both CDI and `--gpus all`. The stock DGX OS and AI Developer Tools paths do not install or replace host packages, enable or restart services, generate CDI configuration, or rewrite the Docker runtime. The Colossus BaseOS path requires the exact reviewed package and failed-service fingerprints, preserves the factory kernel, driver, DKMS, Docker, and NVIDIA Container Toolkit packages, and prepares only Docker access and packaged CDI. -It registers the NVIDIA Docker runtime only when the launch probe proves that the runtime is missing, and it restores the prior daemon configuration if registration fails. +Only the Colossus BaseOS path registers the NVIDIA Docker runtime when the launch probe proves that the runtime is missing, and it restores the prior daemon configuration if registration fails. The preparation helper writes a private audit log under `~/station-bootstrap-logs`. Aside from the installer's separately disclosed `docker`-group membership grant when the current trusted account needs Docker access, the stock DGX OS and AI Developer Tools paths leave only the digest-pinned visibility-test image in the Docker cache. diff --git a/docs/reference/platform-support.mdx b/docs/reference/platform-support.mdx index ff7c57b0954..564e5baa7b8 100644 --- a/docs/reference/platform-support.mdx +++ b/docs/reference/platform-support.mdx @@ -79,7 +79,7 @@ For install requirements and the shorter setup-oriented platform view, refer to | OS | Container runtime | Status | PRD priority | CI | Notes | |----|-------------------|--------|--------------|----|-------| | DGX OS (Spark) | Docker | Tested | P1 | Yes | Use the standard installer and `$$nemoclaw onboard`. For an end-to-end walkthrough with local inference, see the [NVIDIA Spark playbook](https://build.nvidia.com/spark/nemoclaw). | -| DGX OS (Station) | Docker | Deferred | P1 | No | The PRD marks this platform as P1. Workstation form-factor with NVIDIA GPUs and the same Docker + NVIDIA Container Toolkit + CDI requirements as DGX Spark. Station remains Deferred. For canonical hardware qualification, image requirements, preparation, repair limits, and reboot handoff, see [Prepare DGX Station to Install NemoClaw](../get-started/prerequisites/dgx-station-preparation). On a Station GB300, the installer accepts either a clean generic Ubuntu 24.04 ARM64 image or stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0` when a safe, root-owned `/etc/dgx-release` marker identifies `DGX Server for GALAXY-GB300`. It also recognizes two exact no-OTA factory profiles for bounded qualification: April 2026 NVIDIA Colossus BaseOS (`DGX_SWBUILD_VERSION=7.5.0-GB300ws-GB200ws`, `DGX_SWBUILD_DATE=2026-04-02-08-20-16`) and June 2026 NVIDIA AI Developer Tools (`DGX_SWBUILD_VERSION=7.5.0`, `DGX_SWBUILD_DATE=2026-06-16-11-48-10`). The stock DGX OS and AI Developer Tools paths validate the factory stack in place without replacing host packages, enabling or restarting services, generating CDI configuration, or rewriting the Docker runtime. The Colossus BaseOS path preserves the factory kernel, driver, DKMS, Docker, and NVIDIA Container Toolkit packages and prepares only Docker access and packaged CDI. Each path validates the factory GB300 driver and volatile ECC state, Docker and containerd, Buildx, `nvidia-ctk`, CDI, and GPU visibility through both CDI and `--gpus all` before it proceeds. Unknown DGX OS versions, unmatched no-OTA factory images, unsafe release markers, and other Station generations stop before host preparation. The installer offers express install with the pinned `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` recipe, including an approximately 352 GB model download, without follow-up provider, model, policy, or sandbox-name choices. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Pass `--station-deepseek` to use `deepseek-ai/DeepSeek-V4-Flash` while retaining the one-confirmation express flow. The flag requires an interactive terminal, and `/dev/tty` must be available when the installer runs through `curl \| bash`. For headless setup on a qualifying Station, select `NEMOCLAW_PROVIDER=install-vllm` and `NEMOCLAW_VLLM_MODEL=deepseek-v4-flash` instead. Direct managed-vLLM onboarding still defaults to `deepseek-ai/DeepSeek-V4-Flash` when no model override is set. | +| DGX OS (Station) | Docker | Deferred | P1 | No | The PRD marks this platform as P1. Workstation form-factor with NVIDIA GPUs and the same Docker + NVIDIA Container Toolkit + CDI requirements as DGX Spark. Station remains Deferred. For canonical hardware qualification, image requirements, preparation, repair limits, and reboot handoff, see [Prepare DGX Station to Install NemoClaw](../get-started/prerequisites/dgx-station-preparation). On a Station GB300, the installer accepts either a clean generic Ubuntu 24.04 ARM64 image or stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0` when a safe, root-owned `/etc/dgx-release` marker identifies `DGX Server for GALAXY-GB300`. It also recognizes two exact no-OTA factory profiles for bounded qualification. Both require `DGX_PLATFORM=DGX Server for GALAXY-GB300` and require `DGX_OTA_VERSION` and `DGX_OTA_DATE` to be absent. The April 2026 NVIDIA Colossus BaseOS profile requires `DGX_PRETTY_NAME=NVIDIA DGX Server`, `DGX_SWBUILD_VERSION=7.5.0-GB300ws-GB200ws`, and `DGX_SWBUILD_DATE=2026-04-02-08-20-16`; the June 2026 NVIDIA AI Developer Tools profile requires `DGX_PRETTY_NAME=NVIDIA DGX GB300WS`, `DGX_SWBUILD_VERSION=7.5.0`, and `DGX_SWBUILD_DATE=2026-06-16-11-48-10`. The stock DGX OS and AI Developer Tools paths validate the factory stack in place without replacing host packages, enabling or restarting services, generating CDI configuration, or rewriting the Docker runtime. The Colossus BaseOS path preserves the factory kernel, driver, DKMS, Docker, and NVIDIA Container Toolkit packages and prepares only Docker access and packaged CDI. Only the Colossus BaseOS path registers the NVIDIA Docker runtime when the launch probe proves it is missing, with restoration of the prior daemon configuration if registration fails. Each path validates the factory GB300 driver and volatile ECC state, Docker and containerd, Buildx, `nvidia-ctk`, CDI, and GPU visibility through both CDI and `--gpus all` before it proceeds. Unknown DGX OS versions, unmatched no-OTA factory images, unsafe release markers, and other Station generations stop before host preparation. The installer offers express install with the pinned `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` recipe, including an approximately 352 GB model download, without follow-up provider, model, policy, or sandbox-name choices. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Pass `--station-deepseek` to use `deepseek-ai/DeepSeek-V4-Flash` while retaining the one-confirmation express flow. The flag requires an interactive terminal, and `/dev/tty` must be available when the installer runs through `curl \| bash`. For headless setup on a qualifying Station, select `NEMOCLAW_PROVIDER=install-vllm` and `NEMOCLAW_VLLM_MODEL=deepseek-v4-flash` instead. Direct managed-vLLM onboarding still defaults to `deepseek-ai/DeepSeek-V4-Flash` when no model override is set. | | Linux | Docker | Tested | P0 | Yes | Primary tested path. Ubuntu 24.04 has host-level onboarding validation. A digest-pinned Ubuntu 26.04 userspace lane builds the CLI and runs preflight, installer, and platform contracts on eligible main pushes; Docker-host, AppArmor, Landlock, and live onboarding validation on 26.04 remain pending. Other distros (Ubuntu 22.04, Fedora, Rocky, Alma, NixOS, Arch) may work but are not validated. | | macOS (Apple Silicon) | Colima, Docker Desktop | Tested with limitations | P0 | Yes | Start the container runtime (Colima or Docker Desktop) before running the installer. Homebrew Colima users must install both Colima and the Docker CLI (`brew install colima docker`) before `docker info` can work. Xcode Command Line Tools (`xcode-select --install`) are typically required for Node native modules during install. NemoClaw recommends them but does not enforce them during preflight. | | NVIDIA RTX (consumer and Pro workstation GPUs) | Docker | Deferred | P1 | No | The PRD marks this platform as P1. Covers RTX consumer cards and RTX Pro workstation cards on Linux hosts that meet the generic-Linux-GPU requirements (NVIDIA Container Toolkit + CDI present). The provider menu emits managed vLLM behind `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm` for this host class today; the end-to-end onboard path on this hardware is not yet validated in CI. | From bf696e4a178937b3affd384986ff9dcbde677b89 Mon Sep 17 00:00:00 2001 From: Julie Yaunches Date: Fri, 17 Jul 2026 21:16:36 -0400 Subject: [PATCH 5/5] docs: link Station guides by agent Signed-off-by: Julie Yaunches --- ci/platform-matrix.json | 2 +- docs/inference/choose-inference-provider.mdx | 2 +- docs/reference/platform-support.mdx | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-) diff --git a/ci/platform-matrix.json b/ci/platform-matrix.json index 2303386db17..55c4c0d0a7b 100644 --- a/ci/platform-matrix.json +++ b/ci/platform-matrix.json @@ -149,7 +149,7 @@ "name": "Local vLLM (managed install/start)", "status": "caveated", "endpoint_type": "Local OpenAI-compatible", - "notes": "Appears by default on DGX Spark and qualifying DGX Station GB300 hosts. DGX Station remains Deferred. The Station path accepts a clean generic Ubuntu 24.04 ARM64 image, stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0`, or the exact April 2026 NVIDIA Colossus BaseOS and June 2026 NVIDIA AI Developer Tools no-OTA factory profiles documented in [Prepare DGX Station to Install NemoClaw](/user-guide/openclaw/get-started/prerequisites/dgx-station-preparation), after the release marker and factory GPU runtime pass validation. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pins runtime images to immutable digests. DGX Spark and DGX Station models without a model-specific runtime use the `linux/arm64` digest published under `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use the matching `linux/arm64` or `linux/amd64` digest published under `nvcr.io/nvidia/vllm:26.03.post1-py3`. The DGX Station express installer selects `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` with a pinned Hugging Face revision and the multi-platform index digest published under `vllm/vllm-openai:v0.22.0`; that index resolves to the `linux/arm64` manifest on Station. Direct managed-vLLM profile defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `deepseek-ai/DeepSeek-V4-Flash`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls from `nvcr.io` require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing." + "notes": "Appears by default on DGX Spark and qualifying DGX Station GB300 hosts. DGX Station remains Deferred. The Station path accepts a clean generic Ubuntu 24.04 ARM64 image, stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0`, or the exact April 2026 NVIDIA Colossus BaseOS and June 2026 NVIDIA AI Developer Tools no-OTA factory profiles documented for [OpenClaw](/user-guide/openclaw/get-started/prerequisites/dgx-station-preparation), [Hermes](/user-guide/hermes/get-started/prerequisites/dgx-station-preparation), and [Deep Agents](/user-guide/deepagents/get-started/prerequisites/dgx-station-preparation), after the release marker and factory GPU runtime pass validation. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pins runtime images to immutable digests. DGX Spark and DGX Station models without a model-specific runtime use the `linux/arm64` digest published under `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use the matching `linux/arm64` or `linux/amd64` digest published under `nvcr.io/nvidia/vllm:26.03.post1-py3`. The DGX Station express installer selects `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` with a pinned Hugging Face revision and the multi-platform index digest published under `vllm/vllm-openai:v0.22.0`; that index resolves to the `linux/arm64` manifest on Station. Direct managed-vLLM profile defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `deepseek-ai/DeepSeek-V4-Flash`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls from `nvcr.io` require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing." } ], diff --git a/docs/inference/choose-inference-provider.mdx b/docs/inference/choose-inference-provider.mdx index ba3bba979ca..34d587c9c1a 100644 --- a/docs/inference/choose-inference-provider.mdx +++ b/docs/inference/choose-inference-provider.mdx @@ -32,7 +32,7 @@ Use this status table to distinguish validated provider integrations from adapte | Local Ollama | Tested with limitations | Local Ollama API | Available when Ollama is installed or running on the host. Validated default models: `qwen3.6:35b` (high VRAM), `nemotron-3-nano:30b` (medium VRAM), `qwen3.5:9b` (low VRAM fallback). | | Local NVIDIA NIM | Experimental | Local OpenAI-compatible | Requires `NEMOCLAW_EXPERIMENTAL=1` and a NIM-capable NVIDIA GPU. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence with `assertCdiNvidiaGpuSpecPresent`, `src/lib/onboard/fatal-runtime-preflight.ts`). NIM images pull from `nvcr.io` and require NGC registry login. NemoClaw gates this path behind the experimental flag because it does not auto-select a NIM image for the host today. You must explicitly pick from the validated image list. On Linux arm64 DGX Spark and DGX Station hosts, onboarding warns that some NIM images may not publish a `linux/arm64` manifest; the warning is advisory, and the selected image pull can still fail when the registry has no matching platform manifest. Managed vLLM has host-specific default models and is not gated on the same boxes. Validated images referenced in `src/lib/inference/config.ts` and `nemoclaw/src/index.ts`: `nvidia/nemotron-3-super-120b-a12b` (default cloud model), `nvidia/nemotron-3-nano-30b-a3b`, `nvidia/llama-3.3-nemotron-super-49b-v1.5`. | | Local vLLM (already running) | Tested with limitations | Local OpenAI-compatible | Appears in the onboarding menu when NemoClaw detects a server already on `localhost:8000`. No flag required. Model is whatever the existing server serves. | -| Local vLLM (managed install/start) | Tested with limitations | Local OpenAI-compatible | Appears by default on DGX Spark and qualifying DGX Station GB300 hosts. DGX Station remains Deferred. The Station path accepts a clean generic Ubuntu 24.04 ARM64 image, stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0`, or the exact April 2026 NVIDIA Colossus BaseOS and June 2026 NVIDIA AI Developer Tools no-OTA factory profiles documented in [Prepare DGX Station to Install NemoClaw](/user-guide/openclaw/get-started/prerequisites/dgx-station-preparation), after the release marker and factory GPU runtime pass validation. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pins runtime images to immutable digests. DGX Spark and DGX Station models without a model-specific runtime use the `linux/arm64` digest published under `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use the matching `linux/arm64` or `linux/amd64` digest published under `nvcr.io/nvidia/vllm:26.03.post1-py3`. The DGX Station express installer selects `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` with a pinned Hugging Face revision and the multi-platform index digest published under `vllm/vllm-openai:v0.22.0`; that index resolves to the `linux/arm64` manifest on Station. Direct managed-vLLM profile defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `deepseek-ai/DeepSeek-V4-Flash`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls from `nvcr.io` require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing. | +| Local vLLM (managed install/start) | Tested with limitations | Local OpenAI-compatible | Appears by default on DGX Spark and qualifying DGX Station GB300 hosts. DGX Station remains Deferred. The Station path accepts a clean generic Ubuntu 24.04 ARM64 image, stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0`, or the exact April 2026 NVIDIA Colossus BaseOS and June 2026 NVIDIA AI Developer Tools no-OTA factory profiles documented for [OpenClaw](/user-guide/openclaw/get-started/prerequisites/dgx-station-preparation), [Hermes](/user-guide/hermes/get-started/prerequisites/dgx-station-preparation), and [Deep Agents](/user-guide/deepagents/get-started/prerequisites/dgx-station-preparation), after the release marker and factory GPU runtime pass validation. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pins runtime images to immutable digests. DGX Spark and DGX Station models without a model-specific runtime use the `linux/arm64` digest published under `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use the matching `linux/arm64` or `linux/amd64` digest published under `nvcr.io/nvidia/vllm:26.03.post1-py3`. The DGX Station express installer selects `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` with a pinned Hugging Face revision and the multi-platform index digest published under `vllm/vllm-openai:v0.22.0`; that index resolves to the `linux/arm64` manifest on Station. Direct managed-vLLM profile defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `deepseek-ai/DeepSeek-V4-Flash`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls from `nvcr.io` require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing. | {/* provider-status:end */} ## Hosted Providers diff --git a/docs/reference/platform-support.mdx b/docs/reference/platform-support.mdx index 564e5baa7b8..080baa8e9fb 100644 --- a/docs/reference/platform-support.mdx +++ b/docs/reference/platform-support.mdx @@ -104,7 +104,7 @@ NemoClaw routes inference through the OpenShell gateway. Each row below is a pro | Local Ollama | Tested with limitations | Local Ollama API | Available when Ollama is installed or running on the host. Validated default models: `qwen3.6:35b` (high VRAM), `nemotron-3-nano:30b` (medium VRAM), `qwen3.5:9b` (low VRAM fallback). | | Local NVIDIA NIM | Experimental | Local OpenAI-compatible | Requires `NEMOCLAW_EXPERIMENTAL=1` and a NIM-capable NVIDIA GPU. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence with `assertCdiNvidiaGpuSpecPresent`, `src/lib/onboard/fatal-runtime-preflight.ts`). NIM images pull from `nvcr.io` and require NGC registry login. NemoClaw gates this path behind the experimental flag because it does not auto-select a NIM image for the host today. You must explicitly pick from the validated image list. On Linux arm64 DGX Spark and DGX Station hosts, onboarding warns that some NIM images may not publish a `linux/arm64` manifest; the warning is advisory, and the selected image pull can still fail when the registry has no matching platform manifest. Managed vLLM has host-specific default models and is not gated on the same boxes. Validated images referenced in `src/lib/inference/config.ts` and `nemoclaw/src/index.ts`: `nvidia/nemotron-3-super-120b-a12b` (default cloud model), `nvidia/nemotron-3-nano-30b-a3b`, `nvidia/llama-3.3-nemotron-super-49b-v1.5`. | | Local vLLM (already running) | Tested with limitations | Local OpenAI-compatible | Appears in the onboarding menu when NemoClaw detects a server already on `localhost:8000`. No flag required. Model is whatever the existing server serves. | -| Local vLLM (managed install/start) | Tested with limitations | Local OpenAI-compatible | Appears by default on DGX Spark and qualifying DGX Station GB300 hosts. DGX Station remains Deferred. The Station path accepts a clean generic Ubuntu 24.04 ARM64 image, stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0`, or the exact April 2026 NVIDIA Colossus BaseOS and June 2026 NVIDIA AI Developer Tools no-OTA factory profiles documented in [Prepare DGX Station to Install NemoClaw](/user-guide/openclaw/get-started/prerequisites/dgx-station-preparation), after the release marker and factory GPU runtime pass validation. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pins runtime images to immutable digests. DGX Spark and DGX Station models without a model-specific runtime use the `linux/arm64` digest published under `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use the matching `linux/arm64` or `linux/amd64` digest published under `nvcr.io/nvidia/vllm:26.03.post1-py3`. The DGX Station express installer selects `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` with a pinned Hugging Face revision and the multi-platform index digest published under `vllm/vllm-openai:v0.22.0`; that index resolves to the `linux/arm64` manifest on Station. Direct managed-vLLM profile defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `deepseek-ai/DeepSeek-V4-Flash`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls from `nvcr.io` require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing. | +| Local vLLM (managed install/start) | Tested with limitations | Local OpenAI-compatible | Appears by default on DGX Spark and qualifying DGX Station GB300 hosts. DGX Station remains Deferred. The Station path accepts a clean generic Ubuntu 24.04 ARM64 image, stock DGX OS `7.2.0`, `7.4.0`, or `7.5.0`, or the exact April 2026 NVIDIA Colossus BaseOS and June 2026 NVIDIA AI Developer Tools no-OTA factory profiles documented for [OpenClaw](/user-guide/openclaw/get-started/prerequisites/dgx-station-preparation), [Hermes](/user-guide/hermes/get-started/prerequisites/dgx-station-preparation), and [Deep Agents](/user-guide/deepagents/get-started/prerequisites/dgx-station-preparation), after the release marker and factory GPU runtime pass validation. On Station GB300 only, direct-GPU sandbox policies add the exact read-only BDF directory for each host-discovered NVIDIA display-class PCI device, plus only existing CPU, memory, NUMA topology, and NVIDIA module initialization paths; they do not add `/sys`, the PCI parent subtree, or sysfs write access. One physical DGX OS `7.5.0` GB300 validation completed with released OpenShell `0.0.85`, local Nemotron Ultra serving, sandbox `cuInit(0)`, and a Hermes write/read file-tool task. In that validation, reads of `/sys/fs/cgroup/cgroup.controllers` and `/sys/class/net/lo/address` remained denied. DGX Station remains Deferred pending repeat qualification from a clean host and CI coverage. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pins runtime images to immutable digests. DGX Spark and DGX Station models without a model-specific runtime use the `linux/arm64` digest published under `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use the matching `linux/arm64` or `linux/amd64` digest published under `nvcr.io/nvidia/vllm:26.03.post1-py3`. The DGX Station express installer selects `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` with a pinned Hugging Face revision and the multi-platform index digest published under `vllm/vllm-openai:v0.22.0`; that index resolves to the `linux/arm64` manifest on Station. Direct managed-vLLM profile defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `deepseek-ai/DeepSeek-V4-Flash`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls from `nvcr.io` require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing. | {/* provider-status-full:end */} ## Messaging Integrations