CSGHub Files

✨ New Features

• AI Gateway & Inference
• Added Laya inference engine support using the SystemOne protocol.
• Added native-first routing based on each upstream’s declared API format, with protocol adaptation between Chat Completions, Responses, and Messages APIs.
• Enforced API key model scopes across model listing, model details, and inference requests.

• Added an API for optimizing vLLM and SGLang engine arguments based on accelerator type, GPU count, runtime version, multi-node mode, and model context length.

• Repository Compliance

• Added license compliance classification for model and dataset repositories.
• Added compliance and commercial-use metadata to repository APIs and filters.

• Automatically re-evaluates compliance when license files, README content, or default branches change.

• Evaluation

• Added version-pinned model and dataset evaluations.

• Added support for evaluating multiple commits from the same model in one evaluation workflow.

• Hardware & Runtime Support

• Added MetaX MX-C600 support for vLLM, SGLang, and ms-swift.
• Added AMD ROCm environment detection for multi-node vLLM serving.

🚀 Enhancements & Bug Fixes

• AI Gateway
• Fixed duplicate and unavailable models appearing in model-list responses.
• Improved circuit-breaker health checks and unknown-upstream availability reporting.
• Fixed routing path rewrites and fallback behavior across Chat Completions, Responses, Messages, and Anthropic-compatible endpoints.
• Preserved reasoning_content for thinking-mode tool calls and correctly merged parallel function calls from the same turn.
• Improved prompt-caching compatibility, null message-content handling, and non-JSON upstream error handling.
• Fixed ASR connection tests by providing an input-audio sample.

• Secured MCP proxy calls through authenticated internal endpoints.

• Repositories & Mirrors

• Improved GitLab repository imports and repository synchronization reliability.
• Added writable namespace discovery.
• Fixed mirror request handling and repository visibility authorization.
• Allowed administrators to bypass namespace permission checks where repository administration already grants access.

• Added default-branch information to branch API responses.

• Deployments & Resources

• Added automatic reconciliation for unhealthy deployments.
• Improved deployment reconciliation when runtime status and Kubernetes service state disagree.
• Enforced permissions for private deployments and Space start/stop operations.
• Fixed CPU scenario filtering so only pure CPU resources are excluded.
• Improved default inference parameters according to GPU model and runtime version.

• Ensured canceled staged finetuning tasks execute their cleanup handlers.

• Security & Reliability

• Closed an SSRF bypass involving transitional IP addresses.
• Fixed SQL injection in DataViewer query filtering.
• Improved credential redaction across HTTP middleware, proxy, RPC, and structured logs.
• Fixed missing namespace UUIDs on user-scoped access tokens.
• Improved sensitive-content handling for controversial risk classifications.

🛠 Maintenance

• Upgraded NVIDIA and AMD vLLM runtimes and registered additional GLM and Qwen model profiles.
• Updated GitPython, PyJWT, and Requests to versions containing security fixes.
• Moved plugin logger output to stderr.
• Improved database test-container reuse and in-memory database behavior.
• Added the shared code-smell checklist for implementation and review workflows.

Full Changelog: https://github.com/OpenCSGs/csghub-server/compare/v2.5.0-ce...v2.6.0-ce