Follow-up from #414.
developers/mindmap.md (english and chinese) covers HAMi's NVIDIA vGPU scheduling internals in detail: DevicePlugin deployment, Mutating Webhook, KubeScheduler extension, Spread and Binpack scheduling. The content is still technically accurate for what it covers.
The scope is narrower than HAMi today. It does not cover Cambricon MLU, Hygon DCU, Kunlunxin XPU, MetaX GPU, Ascend NPU, or Iluvatar Corex. #414 added a one line note to the page clarifying this, but did not expand the actual diagram.
Doing this properly needs real architecture knowledge of each backend's device plugin and scheduling integration, not something that should be guessed at from the docs alone. Someone with that knowledge should decide whether to:
- extend the existing mind map to add branches for each backend, or
- split it into one mind map per backend, or
- keep it NVIDIA-only and rename it to say so explicitly
Also relevant: neither image has an editable source file (see diagrams-inventory.md), so any update means a full redraw, not an edit.
Follow-up from #414.
developers/mindmap.md (english and chinese) covers HAMi's NVIDIA vGPU scheduling internals in detail: DevicePlugin deployment, Mutating Webhook, KubeScheduler extension, Spread and Binpack scheduling. The content is still technically accurate for what it covers.
The scope is narrower than HAMi today. It does not cover Cambricon MLU, Hygon DCU, Kunlunxin XPU, MetaX GPU, Ascend NPU, or Iluvatar Corex. #414 added a one line note to the page clarifying this, but did not expand the actual diagram.
Doing this properly needs real architecture knowledge of each backend's device plugin and scheduling integration, not something that should be guessed at from the docs alone. Someone with that knowledge should decide whether to:
Also relevant: neither image has an editable source file (see diagrams-inventory.md), so any update means a full redraw, not an edit.