Skip to content

CHANGELOG.md

Latest commit

 

History

History
115 lines (90 loc) · 5.83 KB

CHANGELOG.md

File metadata and controls

115 lines (90 loc) · 5.83 KB
Raw

Changelog

All notable changes to DGX Model Manager are documented here.

[2.0.0] - 2026-08-10

Added

  • Authenticated HTTPS control plane with Viewer, Operator, and Admin roles.
  • One-time host-local administrator bootstrap token.
  • Argon2id password hashing, opaque browser sessions, and CSRF protection.
  • Scoped API tokens and administrative audit history.
  • DGX Spark / GB10 platform and unified-memory awareness.
  • Hugging Face cache inventory with custom scan directories.
  • Hugging Face repository search, file-size inspection, download progress, and local deletion.
  • Ollama inventory, pull, and delete support.
  • Compose Builder with DGX-aware runtime estimates and optimization profiles.
  • Collision-aware host-port allocation for generated Compose deployments.
  • Saved Compose deployment lifecycle: validate, start, stop, logs, archive.
  • Optional LiteLLM route management for supported generated deployments.
  • Optional Ollama wildcard management.
  • Optional multi-node DGX Spark agent architecture.
  • Diagnostics for host, Docker, Compose, application logs, LiteLLM journal, and container discovery.
  • Public Field Manual (docs.html).
  • Offline release validator and reproducible release archive builder.

Model metadata

  • Structural SentenceTransformers detection.
  • Whisper/STT, TTS, audio, embedding, text-generation, and multimodal classification improvements.
  • Nested vision/audio configuration recognition.
  • Exact Safetensors parameter counting for non-quantized checkpoints without loading tensor data.
  • Safetensors dtype inference.
  • Separation of base dtype from checkpoint storage precision.
  • Quantizer implementation and quantization format tracked independently.
  • Mixed quantization represented explicitly with all discovered bit widths.
  • Logical parameter resolution for quantized Hugging Face derivatives from declared base models.
  • Persistent private Hugging Face metadata cache with offline-tolerant enrichment.
  • Metadata-only Hugging Face cache directories excluded from installed model inventory.
  • Hugging Face cache physical-size accounting de-duplicates shared inode-backed blobs/snapshot links.
  • Sub-gigabyte model sizes retained and displayed using adaptive units.

Compose Builder

  • Generated-plan integrity guard prevents browser-supplied arbitrary YAML from becoming a privileged Compose injection path.
  • Saved deployments remain stopped until explicitly launched.
  • Preferred engine ports are checked against both saved deployment metadata and real host bind availability.
  • Port reassignment is recorded in generator decision notes.
  • vLLM relies on checkpoint quantization auto-detection.
  • Stable served-model names are generated for routable OpenAI-compatible deployments.
  • Deployment log viewer enlarged for operational use.

LiteLLM

  • LiteLLM health detection recognizes authenticated endpoints.
  • Optional server-side authentication using LITELLM_MASTER_KEY.
  • LiteLLM master key can be supplied to Model Manager through a dedicated systemd LoadCredential= credential.
  • The Model Manager does not need access to the full LiteLLM environment file.
  • Authenticated /v1/models discovery for Routing.
  • Credential-aware Settings health tests.
  • Refined secret redaction avoids false positives such as max_tokens.
  • LiteLLM config is redacted before reaching the browser.
  • Optional narrowly scoped sudo rule for LiteLLM restart.

Serving-engine safety

  • Running externally managed serving services can be detected without being claimed as v2-managed.
  • Removed unsafe fallback that stopped the first container publishing an engine's conventional port.
  • Stop operations now apply only to v2-managed Compose deployments.
  • Serving Engines UI disables Stop for externally managed services.

Users and access

  • Backend prevents disabling the currently signed-in account.
  • Backend prevents disabling or demoting the last active administrator.
  • UI mirrors those invariants.
  • Signed-in user is identified with a You badge.

Settings

  • Canonical service display names used consistently.
  • Endpoint tests display persistent inline success/failure state and latency in addition to toast notifications.
  • LiteLLM endpoint tests authenticate when a systemd credential is available.

Security

  • HTTPS-first browser access.
  • Authenticated operational reads.
  • CSRF protection for mutations.
  • Private/loopback-first service target policy.
  • Public/unsafe target rejection by default.
  • Secret redaction before configuration reaches the browser.
  • Runtime configuration, databases, TLS keys, tokens, environment files, logs, and generated artifacts excluded from publication.
  • Compose generator refuses arbitrary browser-edited YAML.
  • External-service ownership boundary prevents accidental shutdown of workloads not created by v2.

Release / packaging

  • Public config.example.json separated from local runtime configuration.
  • Maintainer-only hardware acceptance plan and release checklist excluded from Git/public archives.
  • Generated release packages are written under ignored dist/.
  • Per-file SHA-256 manifest generated inside release archives.
  • Archive-level SHA-256 files generated for .tar.gz and .zip.

Hardware acceptance

Live DGX Spark / GB10 acceptance validated the primary local control-plane workflows including model inventory, Hugging Face operations, collision-aware Compose generation/save/archive, authenticated LiteLLM reads, engine ownership detection, local-node display, diagnostics, access controls, and settings.

The following were intentionally not exercised against live production-serving resources before publication:

  • launching/stopping a new v2-managed GPU-serving stack while another active inference workload was in use;
  • mutating/restarting LiteLLM routing during active client use;
  • second-physical-DGX remote-node lifecycle testing.

These remain explicit validation gaps for post-release field testing.