csghub-server Files

v2.3.0-ce

New Features

AI Gateway

• Added OpenAI-compatible Responses API support, including streaming/non-streaming adapters, response ID handling, usage accounting, and token counting.
• Added OpenAI-compatible image editing via /v1/images/edits, with multipart request rewriting, prompt moderation, tracing, storage integration, and usage recording.
• Added asynchronous generation processing for long-running generation tasks, including scheduled status refresh and metering event publication.
• Added TPU accelerator support across hardware definitions and runner resource handling.

Repositories & Mirrors

• Added batch repository extra-information lookup via /api/v1/repos/extra.
• Added ZIP archive downloads for code repositories.
• Added mirror task cancellation support and mirror-backed fork repository creation.

Forwarded Services

• Added configurable forward proxy routes for third-party backend services.

Enhancements & Bug Fixes

AI Gateway

• Fixed circuit breaker state transitions so open circuits can reliably move to half-open and close again.
• Added circuit-state Prometheus metrics and improved health-check logging with upstream context and response snippets.
• Improved LLM tracing for chat, embedding, multimodal, and streaming requests, including bounded captured content and richer metadata.
• Improved video, audio, image, and Responses API usage counting.
• Fixed request cancellation and retry behavior for upstream proxy calls and streamed responses.
• Improved model target handling, provider/model trimming, Qwen3.5 metadata parsing, and unsupported tool/developer-role normalization.

Deployments & Resources

• Improved resource availability checks for offline nodes, invalid accelerator type, missing prices, CPU-only scheduling, and cluster-level VXPU scheduler configuration.
• Added cluster-specific scheduler lookup for evaluation, finetune, and dataflow jobs.
• Added GGUF runtime entrypoint validation for llama.cpp and ktransformers deployments.
• Added finetune swift_command validation and support for SFT, RLHF, and PT modes.
• Improved admin deploy listing with deleted-deploy filtering and deploy-name search.

Repositories & User Pages

• Added filtering support to user repository lists for models, datasets, codes, spaces, MCP servers, and skills.
• Escaped repository commit messages in API responses.
• Removed branch create/delete routes from public repository routers.
• Fixed code API responses that could return null.

Mirrors & Data Viewer

• Preserved LFS metadata during mirror sync.
• Improved mirror task state transitions, progress updates, priority ordering, and cancellation safety.
• Cleaned up old dataset viewer data when no convertible files remain.

Security & Reliability

• Hardened JWT refresh claim handling.
• Improved HTTP request logging so canceled requests and panics still emit logs, including client-closed 499 status.
• Improved reverse proxy handling for canceled upstream requests.
• Added image URL validation to reject unsafe hosts, ports, redirects, and non-image content types.
• Preserved trace ID injection through slog.With and grouped loggers.
• Returned empty JSON arrays instead of null for empty list responses.

Maintenance

• Added and updated focused unit tests across AIGateway, repositories, mirrors, deployment/resource handling, JWT, logging, URL validation, and DataViewer workflows.
• Added configurable Prometheus metric keys.
• Added inference image/config updates for diffusers, FunASR, llama.cpp, ROCm, and related serving scripts.
• Updated runtime and example configuration for AIGateway tracing, async generation, archive limits, Prometheus metrics, and forwarded services.
• Updated Go dependencies for tracing support.

Full Changelog

https://github.com/OpenCSGs/csghub-server/compare/v2.2.0-ce...v2.3.0-ce