Skip to content

v2.4.0-ce

Latest
Latest

Choose a tag to compare

View all tags
@Rader Rader released this 09 Aug 08:16
· 3 commits to main since this release
v2.4.0-ce
5d20fb5
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

This release expands multimodal inference capabilities, improves repository and deployment workflows, strengthens synchronization reliability, and introduces new administration and evaluation features.

New Features

AI Gateway

  • Added a PaddleOCR-backed OpenAI-compatible /v1/ocr endpoint.
  • Added text-to-speech, batch speech generation, and custom voice management APIs.
  • Added LongCat avatar and audio-driven video generation support.
  • Added streaming audio transcription and audio-duration reporting.
  • Added reasoning support to the Responses API chat adapter.

Repositories & Skills

  • Added HTTP range downloads for repository files.
  • Added Hugging Face-compatible model tree APIs.
  • Added code repository imports from Git URLs.
  • Added repository ownership transfer for skills.
  • Added repository last-commit size metadata.

Evaluation & Deployments

  • Added Claw evaluation creation, status, log, and result workflows.
  • Added deployment log queries using limit and since.
  • Added resource scenario constraints and automatic sandbox resource allocation.
  • Added configurable eager-mode behavior for vLLM deployments.

Administration

  • Added user deletion functionality to the admin console.

Enhancements & Bug Fixes

AI Gateway

  • Improved Responses API moderation, logging, tracing, streaming event conversion, and usage accounting.
  • Simplified model identifiers while preserving stable routing for inference and serverless models.
  • Added backend upstream testing to avoid browser CORS restrictions.
  • Improved tool-calling detection from engine configuration.
  • Fixed audio token accounting and FunASR audio-duration handling.
  • Added AMD runtime support for image generation and FunASR.
  • Improved image moderation by reading protected content as a stream instead of fetching its URL anonymously.
  • Unified request-count and latency metric dimensions.

Mirror & Repository Synchronization

  • Improved multi-source synchronization and preemptive task scheduling.
  • Fixed mirror jobs becoming stuck during preemption, cancellation, and state transitions.
  • Normalized namespace mappings to prevent case-sensitive conflicts.
  • Added earlier namespace, repository-name, and path-change validation.
  • Allowed metadata scan failures to be skipped without rejecting repository pushes.
  • Improved mirror worker reliability and LFS synchronization behavior.

Deployments & Resources

  • Added reconciliation for stale, oscillating, and timed-out deployments.
  • Prevented suspended pods from being incorrectly marked as failed.
  • Prevented duplicate active deployment names within the same owner and deployment type.
  • Cleaned up deployment instances after Knative service deletion events.
  • Fixed deployment filtering when the requested status is zero.
  • Fixed space-resource filtering by status.
  • Improved custom GPU label selection and VXPU memory calculations.
  • Excluded clusters without usable space-resource definitions from cluster results.

Organizations & Metadata

  • Fixed namespace conflicts between users and organizations.
  • Improved validation and error messages when creating existing namespaces.
  • Fixed Casdoor username inconsistencies.
  • Added pagination to tags and moderation word-set endpoints.
  • Fixed incomplete tag results on later homepage pages.
  • Added paddle_static, text-ranking, text-to-audio, OCR, and audio-driven-video metadata support.

Dataset Experience

  • Added clearer status messaging when format-conversion tasks complete only partially.

Runtime Reliability

  • Fixed Loki timeout conflicts for recent-log queries.
  • Improved error propagation between remote runners and API handlers.
  • Updated PaddleOCR, FunASR, vLLM, SGLang, LlamaFactory, notebook, and multimodal inference runtime configurations.

Maintenance

Backend

  • Eliminated repeated regular-expression compilation in repository queries.
  • Updated Go dependencies and inference container definitions.
  • Expanded unit coverage for AI Gateway adapters, deployments, mirrors, repository APIs, moderation, resource scheduling, and organization behavior.
  • Refreshed generated API documentation and error translations.

Frontend

  • Refactored render handlers to use injected RenderBase dependencies and expanded the associated test suites.

New Contributors

Assets 2
Loading