Skip to main content

v26.11.2

On this page

Released 2026-11-08 · Release line 26.11 · Commit 3cf500e60

This release makes GPUWorks easier to use in offline and private-network environments, and fixes reasoning output in the Playground and benchmark metrics, agent connection settings, and API key revocation. All 26.11 users should upgrade, especially those deployed on private networks or offline. This release has no database migration.

Upgrade Notes

  • Private-network sources are allowed by default: HTTP sources on ordinary private and loopback addresses, reached from the management node, change from denied by default to allowed by default. To keep the old behavior, add outbound_restrict_private_hosts: true to server.yaml before upgrading; existing outbound_allow_private_hosts entries keep working. Metadata, link-local, and special reserved addresses stay denied in both modes.
  • Storage paths: storage_probe_roots no longer limits the paths of new or edited storage; it only drives a startup warning about existing storage.
  • Upgrade the server first, then the agents: the benchmark metric fixes (when the agent generates the load), the Docker startup wait, and runveil load run on the agent, so agents also need v26.11.2.

Added

  • gpuworks-agent runveil load -i FILE REF: load a docker save or OCI archive (.tar / .tar.gz) onto a runveil (udocker) node, without the management node reaching an HTTP source and without stopping the agent. See Offline Deployment.
  • server.yaml gains outbound_restrict_private_hosts (default false); when enabled, only private targets granted in outbound_allow_private_hosts are allowed.
  • server.yaml gains storage_denied_paths: replaces the built-in list of directories storage may not use; [] denies none. See Configuration and Environment Variables.
  • Benchmark reports add time to first visible answer and a count of requests that produced output but no answer text; a streamed request whose usage reports output tokens but carries no recognizable text raises the warning BENCHMARK_STREAM_OUTPUT_UNRECOGNIZED. See Run Benchmarks and Read the Report.

Changed

  • Storage paths no longer need to sit under storage_probe_roots (default /opt/gpuworks/models, /opt/gpuworks/cache, /mnt/gpuworks); paths such as /data/models work directly. System directories are still denied.
  • Changes to outbound_allow_private_hosts and outbound_restrict_private_hosts take effect on the management node's next outbound request, without restarting the server.
  • A source denied by the outbound policy is reported as IMAGE_ARCHIVE_URL_POLICY_DENIED instead of "unreachable".
  • When the Docker daemon is not ready at agent startup, the agent waits up to 2 minutes; if Docker still does not answer, it falls back to runveil with the reason daemon_not_ready and switches back once Docker is ready. Newly generated systemd units start after docker.service and containerd.service.

Fixed

  • A model registered from a node's local directory, paired with an inference template that declares public weight sources such as Hugging Face or ModelScope, kept failing evaluation with a source conflict and could not be deployed. (GWKS-216)
  • x.part / x.part.state files left by a cancelled platform download were still treated as an unfinished download after another tool finished x, so the model could not be registered or scanned.
  • Installing or restarting a nohup agent inherited stale GPUWORKS_CONNECT_MODE, GPUWORKS_SERVER, GPUWORKS_TOKEN, and similar variables from the shell, reverting the connection settings already written to the configuration.
  • An agent that started before Docker fell back to runveil permanently, and multi-node deployments were refused because runveil cannot expose RDMA devices.
  • Newer vLLM versions stream reasoning under delta.reasoning (so does the cuda12.9-vllm0.31.0-r2 runner shipped with 26.11), and the Playground and AI assistant showed no reasoning. They now read reasoning_content, reasoning, and thinking.
  • The benchmark load generator did not read delta.reasoning: TTFT absorbed the whole reasoning time, TPOT came out too low, and requests that spent their output budget reasoning were left out of the TTFT/TPOT samples.

Security

  • Revoking an API key prefix that does not exist returns 404 API_KEY_NOT_FOUND instead of reporting it revoked; an authentication already in flight before a revocation cannot admit new requests after it. (GWKS-203)
  • klauspost/compress moves to 1.20.2 for GO-2026-6865; the library decompresses imported OCI image layers and downloaded archives. (GWKS-214)
  • Built with Go 1.26.10 for GO-2026-6809 in crypto/x509, reachable through server_dial.

Downloads

Archives live under https://dl.gpuworks.ai/releases/v26.11.2/; the same paths on Alibaba Cloud OSS and Huawei Cloud OBS hold identical files.

FileSHA-256
gpuworks-v26.11.2-linux-amd64.tar.gz6b1762f0047819b12dd1750a514e8d38f4a602ca2efa2445b31d412f510edc66
gpuworks-v26.11.2-linux-arm64.tar.gz1498dce13f019b106e35adb2ab701197430372af6d6aa29b9071e841b937af74
gpuworks-v26.11.2-darwin-arm64.tar.gz00e5ef7e08cea99f26d20644ef6e1b219aab0a83f49ad45f3654118aba8a6cb8
gpuworks-v26.11.2-darwin-amd64.tar.gzbe2ce55b2e1cd61aa171ef21fb752c34326b50de4e8cc5747ec795fc3fdfd60e

Full Changelog

v26.11.1...v26.11.2

Was this page helpful?