- spec keeps definition fields only; observed hardware read from status.detail
- sliceable derived from slicedDetail (logical count / physical profiles)
- flat snapshot format isolated in instance-description for back-compat
- drop flavor.spec.sliceable; onceMaxRequest realigned (no ram/localStorage)
- lock compute ratio to memory ratio when coresPercentageOvercommit is false
- sliced mode submits whole cores / Gi; displayName-first type labels
- Add instance-type column to edit drawer for stopped instances
- Extract saveInstanceDataInDescription into shared util
- Floor sliced unit resources; keep CPU >= 1 core
- Expose applyInstanceType for both create and edit flows
- add activate/deactivate row actions (phase-driven) with new PUT APIs
- add displayName field; list shows displayName || name
- rename list OS column title to Platform
- share a two-line flavor renderer between the create dropdown and the list
- normalize 'CPU Only' label to 'CPU-only'
- Card/table list of instance types per Kubernetes cluster with delete
- Create drawer with flavor picker and per-GPU resource form (CPU/RAM/storage)
- Fix unit CPU to 1 for non-acceleratable flavors; N/A memory for generic types
- Add routes and i18n across all locales
- Regroup meta into aligned rows (CPU/RAM/Memory, Arch/Max/Sliced) and drop the OS row
- Use aggregated endpoint for instance types
- Share formatManufacturer helper for consistent vendor casing
- Scale sliced CPU in millicores and RAM in MiB for precise, k8s-valid
fractional allocations; show rounded decimals in the disabled inputs
- Disable the by-ratio option when the type has no sliced capacity
- Use the dark logo map under the dark theme in template cards
- Add whole/by-ratio segmented toggle to the GPU Instance form; sliced
mode picks a VRAM ratio (10-100 or finer 1-10 when max<10), scales
CPU/RAM by the ratio (floored, min 1), pins compute at 100%
- Availability, card Max, and Instance Type cell reflect sliceable types
- Add ready/deleting status columns for storage and storage types
Reserve a stable block for the empty/loading state (matching the NoResult
height) and fade the empty state in, so entering a list page no longer jumps
between spinner, empty result, and rows. Apply scroll-table to the antd main
lists and pass emptyMinHeight to the SealTable-based lists.
Managers start in a loading state and gate the empty state on the
cluster/worker fetch, so the 'no clusters/workers' guidance no longer
flashes before the real infra data lands.
My Models empty state now guides managers to create a cluster/worker or
deploy a model (queries gated to canSeeAdmin/canSeeOrgAdmin); plain users
keep the consumer copy. Add an Experimental badge to the GPU-instance
cluster type card.
Remove the cluster-detail route and the name-link navigation; the
cluster name is now plain text. Mount the plugin-provided
clusterDetail.AccessDrawer so the enterprise row action can open it.
The basic form seeds a default cluster on open, which fired the full cluster-change handler and triggered an evaluate request before any model was picked. Split off a seed callback that refreshes scoped options without evaluating, and guard the evaluate handler with a model-selected check as a backstop.
Read creator_id / creator_name / creator_deleted from the breakdown item
root into the flat user_* fields, so the GPU Instances / Storage chart
export marks a deleted owner independently of the grouped instance/volume.
Deleted users / models / API keys / instances / volumes now sink below the
live entries in each usage tab's filter, while the current account stays on
top and live entries keep their order (stable sort).
Group the GPU Instances and Storage chart exports by date+instance /
date+volume, add Instance/Storage and (org-admin only) User columns, and
mark deleted instances/volumes and their owner user independently in both
the preview and the exported sheet.
Replace unreliable Form.useWatch on the unregistered gpuInstanceOptions
path with explicit shared clusterType state. Fixes the type selector
being unclickable and flickering, and ensures a model cluster no longer
submits a stale gpuInstanceOptions payload.
Switch all echarts-based chart imports to the dedicated core-ui charts
entry so echarts stays out of the synchronous entry bundle and loads
only on chart routes. Requires @gpustack/core-ui with the ./charts export.
core-ui no longer bundles katex.min.css (it would base64-inline ~1.4MB
of fonts into the render-blocking global css). Wrap FullMarkdown locally
to co-locate the katex stylesheet so it loads only in the routes that
render math.
page-box now imports the header-slot bridge from core-ui and re-exports it, plus a dev warning when a header slot has multiple owners. cluster-detail breadcrumb becomes a BaseSelect to switch clusters, and the tab tables are keyed by id so they refetch on switch.
- breakdown sub-tables now send all active filters (route/user/api_key), matching the trend chart
- summary tab filters the token trend by user and unions user options from both meta APIs (deduped by id)
- stabilize the filters reference so meta load no longer retriggers a second fetch on mount
- Add "Dynamic add-item form fields" guidance to CLAUDE.md (pick
component by field schema: LabelSelector / ListInput / MetadataList)
- Fold StatusTag status-display recipe into create-crud-page skill
- Remove DESIGN.md (content migrated to the above)
jialin