Netclaw Inference Images

Public, reproducible container images for running open inference workloads on Petabridge hardware. Images are built on GitHub-hosted runners and published to the public GitHub Container Registry under ghcr.io/petabridge.

This repository contains build recipes and release automation, not model weights, credentials, hostnames, or private network configuration.

Built for NetClaw

These images are built first to power Netclaw, Petabridge's open-source AI agent. Netclaw runs on your hardware, connects to the communications tools your team already uses, and keeps agent execution and inference under your control. The images remain usable as standalone inference runtimes wherever their documented hardware and model requirements fit.

Images

Image	Platform	Intended hardware	Status
vllm-deepseek-v4-flash-0731-dspark-gb10	linux/arm64	Two NVIDIA DGX Sparks, TP=2	Source-build candidate; hardware qualification pending
tensorfold-glm-5-3-flash-gb10	linux/arm64	Two NVIDIA DGX Sparks, TP=2	Vendored MiaAI-Lab TensorFold recipe; hardware qualification pending, base digest outstanding (Apache-2.0 folder, see its README)
tensorfold-qwen3-8-flash-gb10	linux/arm64	Two NVIDIA DGX Sparks, TP=2	Vendored MiaAI-Lab TensorFold recipe plus the Petabridge image prompt reuse patch; hardware qualification pending (GPU test plan outstanding)
tensorfold-deepseek-v4-1-flash-1017-gb10	linux/arm64	Three NVIDIA DGX Sparks, TP=3	Source-build candidate of LakyFx's DS41F 1.2.5 / K2a TensorFold recipe; hardware qualification pending
llama-cpp-rocm-gfx1201-avx2	linux/amd64	AMD RDNA4 (gfx1201) hosts, AVX2 CPU baseline	Source-build candidate; hardware qualification pending
llama-cpp-rocm-gfx1201-avx2-rdna-boosts	linux/amd64	AMD RDNA4 (gfx1201) hosts, AVX2 CPU baseline	Experimental; vendored rdna-boosts blocks 00-08 on ROCm 7.2.3; hardware qualification pending

The DGX Spark image targets deepseek-ai/DeepSeek-V4-Flash-0731. Model weights are downloaded separately and should be pinned to the revision recorded in the image's dependency.lock.json.

What Petabridge builds

The first image started as a thin, labeled copy of an already exercised r0b0tlab runtime. The current recipe goes materially further:

1. Downloads vLLM source from the immutable upstream commit recorded in the lock file and verifies the source archive checksum.
2. Applies the repository-owned, checksum-locked DeepSeek V4 / SM121 source overlay derived from the cumulative public patch.
3. Builds and installs a new vLLM Python package from that source.
4. Verifies every installed runtime overlay file before the image is emitted.

The native CUDA, FlashInfer, B12X, NCCL, and vLLM extension artifacts are still inherited from one pinned r0b0tlab image digest. That boundary is intentional: the public upstream recipe does not contain enough material to reproduce its patched FlashInfer kernel parent byte-for-byte. Retaining the proven native stack gives us a testable source-owned vLLM image without pretending those native inputs are Petabridge-built. The exact boundary and every known version are documented in the image lock file.

Pulling images

Candidates use an immutable source-commit tag:

docker pull ghcr.io/petabridge/vllm-deepseek-v4-flash-0731-dspark-gb10:sha-<12-character-commit>

Deployments should use the published manifest digest rather than a tag:

ghcr.io/petabridge/vllm-deepseek-v4-flash-0731-dspark-gb10@sha256:<digest>

Release tags are added only by promoting an existing candidate digest. The promotion workflow never rebuilds an image.

Repository layout

images/
  nvidia/dgx-spark/   ARM64 / GB10 / SM121 images
  amd/rocm/           x64 ROCm images (RDNA4 / gfx1201)
scripts/              Manifest, source-overlay, release, and README validation
.github/workflows/    Public-runner validation, builds, and promotion

Each buildable image has a declarative image.json, a digest-pinned Dockerfile, dependency locks, retained licenses, and any downstream source patches or overlays needed to reproduce it.

CI/CD and supply chain

• Pull requests perform static validation only; they never receive registry credentials and never publish images.
• Candidate builds run only from master on GitHub-hosted ARM64 runners.
• GitHub Actions and external base images are pinned by full commit or digest.
• Candidate tags are immutable sha-<commit> tags.
• BuildKit publishes an SBOM and maximum-mode provenance attestation.
• Promotion accepts an exact digest and adds a version tag without rebuilding.

See the CI/CD contract and release notes for the operational and release process.

License and attribution

Repository-authored material is MIT licensed. vLLM is Apache-2.0, and the image retains upstream notices and third-party attribution alongside the runtime. Individual dependencies remain subject to their respective licenses.