CSGHub Files

✨ New Features

• AI Gateway & OpenAI-Compatible APIs
• Audio Transcription: Added OpenAI-compatible /v1/audio/transcriptions support with multipart request rewriting and audio token usage counting.
• Text/Image-to-Video: Added /v1/videos, /v1/videos/{id}, and /v1/videos/{id}/content APIs with provider adapters for OpenAI-compatible endpoints, LightX2V, MiniMax, and Seedance.
• Model Routing: Added provider-aware model IDs, composite model ID parsing, upstream catalog support, session routing, fallback retry for chat completions, and per-upstream availability reporting.
• Usage Limits: Added Redis-backed per-window usage limit checks for configured upstream policies.

• API Key Auth: AI Gateway inference endpoints now require user/org API keys instead of normal login sessions.

• API Key Management

• Added namespace-scoped API key management for users and organizations, including create, list, update, delete, built-in key retrieval, and built-in key refresh APIs.

• Added user/org API key authentication context propagation for downstream services.

• Inference & Evaluation

• Added configurable model architecture checks for inference, including admin APIs to view and update inference architecture rules.
• Added SGLang-based Qwen3-Guard stream inference configuration and Docker assets.
• Added AMD EvalScope evaluation configuration and Docker image support.

• Updated vLLM and AMD vLLM inference images/configuration to v0.19.0.

• Repository, Tags & Skills

• Added automatic industry tag scanning for model and dataset repositories using configured LLM prompts.
• Added source tracking for repository tags and safer tag replacement/removal behavior.
• Added skill mirror_from_saas routes and skill clone URL fields.
• Added a dedicated skill tag category seed.
• Improved SKILL.md validation and added broader validator tests.

🚀 Enhancements & Bug Fixes

• AI Gateway Reliability
• Fixed async model cache writes mutating live model lists.
• Fixed nil-user panic risk when listing CSGHub models.
• Improved sensitive-check whitelist lookup behavior.
• Made sensitive-check behavior configurable per LLM config where available.

• Improved SGLang Guard stream trace/session header handling.

• Resource Scheduling

• Added unavailable reasons to resource list responses.
• Added cluster offline/unavailable status handling.
• Prevented CPU-only workloads from being scheduled onto XPU nodes.

• Added replica-aware resource checks for Spaces.

• Repository & LFS

• Added repository size calculation trigger command.
• Added LFS pointer download nil-URL protection.

• Added LFS size checks before syncing files.

• Data Viewer

• Added file-size checks and optimizations before converting preview files.

• Finetune & Runner

• Fixed finetune jobs missing model and dataset revision data.

• Fixed potential runner panic paths in service/workflow handling.

• Proxy & Networking

• Set proxied Host headers without port where required.
• Sanitized logged authorization headers in internal proxy logs.

🛠 Maintenance

• Upgraded vulnerable dependencies reported by Dependabot.
• Improved accounting metering retry limit configurability.
• Added and refreshed unit tests across AI Gateway, API keys, resource checks, tags, skills, LFS, and database stores.
• Improved CI/test stability and separated CI build cache behavior.

Full Changelog: https://github.com/OpenCSGs/csghub-server/compare/v2.0.0-ce...v2.1.0-ce