v2.5.0-ce

✨ New Features

• AI Gateway & Multimodal APIs

  • Added Anthropic-compatible Messages API support through /v1/messages, including native and protocol-adapted upstream routing.
  • Added the OpenAI-compatible /v1/audio/translations API.
  • Extended the Responses API to hosted models and chat-adapter tools.
  • Added PaddleOCR-VL document parsing with preprocessing and embedded-image extraction.
  • Added configurable image, audio, and video moderation for repository discussions.

• Finetuning & Evaluation

  • Added staged download, training, and upload workflows to reduce finetuning resource costs.
  • Added an MS-SWIFT agentic job runtime.
  • Upgraded EvalScope and added synchronized dataset rules for supported evaluation frameworks.

• Repository & Package Management

  • Added synchronization for MCP and Skill repositories.
  • Added S3-backed synchronization for Skill and Code packages.

• Scheduling & Resource Management

  • Added an API for updating space resource scenario constraints.
  • Added Ascend NPU runtime images and Ascend 910 scheduling support.
  • Added a TGI ROCm inference runtime for AMD GPUs.
  • Added hardware-mode information to space resource responses.

• Security & Traffic Control

  • Added configurable Redis-backed sliding-window and token-bucket API rate limiting.

🚀 Enhancements & Bug Fixes

• AI Gateway Reliability

  • Added protocol-specific upstream health checks and connection probes for text, image, audio, rerank, Responses, and Anthropic endpoints.
  • Prevented unknown upstream health states from incorrectly blocking traffic or marking models unavailable.
  • Fixed health-check defaults being overwritten during database loading.
  • Improved PaddleOCR-VL pipeline selection, document preprocessing, and OCR response handling.
  • Captured generated audio duration in text-to-speech usage data.

• Repositories, Mirrors & Imports

  • Fixed Skill and Code import failures and optimized package publishing during mirror synchronization.
  • Rejected invalid mirror metadata source URLs.
  • Fixed mirror deletion when the underlying repository had already been removed.
  • Safely handled large Hugging Face NDJSON commit payloads.
  • Paginated repository-file scans and preserved scan errors instead of silently discarding them.
  • Improved repository-list performance by limiting expensive count queries.

• Deployments & Finetuning

  • Preserved node selections when creating finetuning jobs.
  • Reduced deployment-update delays when a cluster is offline.
  • Added cluster-state checks and force-stop support for serverless deployments.
  • Added status filtering and multi-status queries for deployment listings.
  • Enforced repository and version permissions for private inference endpoints.

• Authorization & API Safety

  • Added authentication requirements to previously unprotected user APIs.
  • Protected the event-stream endpoint from unauthenticated resource exhaustion.
  • Removed API keys from verification logs and corrected verification error handling.
  • Made namespace-path lookup deterministic.

• Evaluation & Resources

  • Fixed EvalScope dataset identifiers to match the upstream 1.10.0 definitions.
  • Added token-usage details to Claw evaluation summaries.
  • Added XPU memory-size metadata for physical and partitioned accelerators.

🛠 Maintenance

• Updated CI actions and pinned the test toolchain to Go 1.26.
• Upgraded GitPython to 3.1.61 for Space runtimes.
• Updated MS-SWIFT, AMD finetuning, CSGClaw, and inference runtime configurations.
• Expanded unit and integration coverage across AI Gateway protocols, moderation, finetuning workflows, mirroring, deployment controls, and rate limiting.

Full Changelog: v2.4.0-ce...v2.5.0-ce