Skip to main content

v26.11.3

On this page

Released 2026-11-12 · Release line 26.11 · Commit 1335143d3

This release fixes two Hygon DCU deployment problems: cards that report only the series BW matched no built-in template, and instances placed on any card other than card 0 found no GPU under Docker and nerdctl. It also fixes deployments being refused when the selected model and the weight source disagree. 26.11 users with Hygon DCUs should upgrade. This release has no database migration.

Upgrade Notes

  • Upgrade the server first, then the agents: the Hygon card numbering fix is in the agent, so agents must also move to v26.11.3; the built-in templates' card range comes with the server.
  • Built-in Hygon specs cover more cards: built-in Hygon specs that pass on real K100 / K100_AI hardware now also apply to the whole BW family (a bare BW that reports only the series, and BW100, BW150, BW1000, BW1100 and so on). Apart from Qwen3-0.6B these specs have not each been run on BW. Custom templates copied from built-in ones keep their old card range; edit them if needed.

Changed

  • The GPU name predicate of built-in Hygon specs widens from BW1000 / BW1100 to the whole BW family. On some drivers hy-smi --showproductname reports only Card Series: BW, and every built-in Hygon spec refused such nodes.
  • The manual installation guide downloads the latest release instead of a nightly build.

Fixed

  • Hygon cards for which hy-smi reports only BW failed every built-in template with "no GPU passed the deployment hardware check". (#10)
  • Docker and nerdctl hand a Hygon container only the selected cards, so HIP numbers them from 0 inside, but the agent set HIP_VISIBLE_DEVICES / HYGON_VISIBLE_DEVICES to the host index; any selection other than card 0 failed with No HIP GPUs are available. runveil binds the whole host /dev and keeps host indexes.
  • A deployment whose form carried the selected model together with a different weight source was refused as a source conflict in evaluation and creation; the selected model now wins. (GWKS-234)

Downloads

Archives live under https://dl.gpuworks.ai/releases/v26.11.3/; the same paths on Alibaba Cloud OSS and Huawei Cloud OBS hold identical files.

FileSHA-256
gpuworks-v26.11.3-linux-amd64.tar.gz601aeb1776e31a111842cf5ff26c4eccc1a789b0ee08b5fed129211be32711e9
gpuworks-v26.11.3-linux-arm64.tar.gz6bff1c2b6e24b5e5becb8804675cd0bbc5bd428404f99b0d230e4a28b7c23128
gpuworks-v26.11.3-darwin-arm64.tar.gzd442a03ec03f11febae628359058ac39f3ff11c301782d38ff2442955cadd894
gpuworks-v26.11.3-darwin-amd64.tar.gz578811ab7dfdb8aadcd3ac6369c0db995d78c52107bb72449122c880e84f5489

Full Changelog

v26.11.2...v26.11.3

Was this page helpful?