Skip to content
- The Neutree CLI supports exporting access logs and model usage, making it easier to archive or perform offline analysis of inference call records and token usage.
- Adds MinerU (document parsing) and FunASR (speech recognition) model support, and supports quick deployment through model catalogs.
- When the Neutree CLI queries resources, table output now includes resource IDs, making it easier to copy resource IDs for subsequent commands.
- API keys support configuring the model invocation scope separately for endpoints and external endpoints.
- API key token quotas support Token, K, M, and B as units, making it easier to configure large token limits.
- The access log retention period is now adjusted to retain access logs for 30 days by default.
- Default monitoring collection items included the NVSwitch connection status, causing collection exceptions in some non-NVSwitch environments. The issue has been resolved in this release.
- GPU model labels for endpoints were displayed in the wrong order. The issue has been resolved in this release.
- Internal static node permissions might have been displayed in role permission configurations. The issue has been resolved in this release.
- When an endpoint was deployed by using a Recipe model catalog, if the validated hardware did not match any cluster GPU, the accelerator selector was hidden, preventing users from selecting an accelerator. The issue has been resolved in this release.
- When the AI gateway forwarded requests through model mapping, empty arrays in tool schemas might be rewritten as empty objects, causing some upstream services to return 400 errors. The issue has been resolved in this release.
- Access logs with large request bodies or response bodies might fail to be written. The issue has been resolved in this release.
- When deploying component images, Docker Hub upstream image paths might have been rewritten incorrectly. The issue has been resolved in this release.
- The scrollbar on the monitoring dashboard displayed unstable behavior. The issue has been resolved in this release.
- In external accelerator metric collection mode, GPU hardware information might have been displayed incorrectly because the node agent did not have GPU visibility configured. The issue has been resolved in this release.
- After a Helm upgrade, the Kong plugin schema was not automatically reloaded, causing gateway configuration synchronization to fail. The issue has been resolved in this release.
- After Router terminated abnormally or a request was interrupted, request statistics might have remained. The issue has been resolved in this release.
- After the Router connection pool reached its limit, some requests might have failed to continue being forwarded to the backend. The issue has been resolved in this release.