91 lines
4.5 KiB
Markdown
91 lines
4.5 KiB
Markdown
# Changelog
|
|
|
|
All notable changes to `circuitforge-core` are documented here.
|
|
Format follows [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
|
|
Versions follow [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
|
|
|
|
---
|
|
|
|
## [0.3.0] — 2026-04-02
|
|
|
|
### Added
|
|
|
|
**Hardware module** (`circuitforge_core.hardware`) — closes #5
|
|
- `detect_hardware()`: probes nvidia-smi / rocm-smi / Apple system_profiler / CPU fallback → `HardwareSpec`
|
|
- `select_tier(vram_mb)`: maps physical VRAM to a named `VramTier` (CPU / 2 / 4 / 6 / 8 / 16 / 24 GB)
|
|
- `generate_profile(spec)`: converts a `HardwareSpec` + service URLs → `LLMConfig` (llm.yaml-compatible)
|
|
- `HardwareSpec`, `LLMBackendConfig`, `LLMConfig` dataclasses
|
|
|
|
**cf-docuvision service** (`circuitforge_core.resources.docuvision`) — closes #8
|
|
- FastAPI HTTP service wrapping ByteDance/Dolphin-v2 (Qwen2.5-VL backbone, ~8 GB VRAM)
|
|
- `POST /extract`: accepts `image_b64` or `image_path` + `hint` (auto / table / text / form) → `ExtractResponse`
|
|
- Lazy model loading — model stays unloaded until first request
|
|
- JSON-structured output with 21 element types; plain-text fallback when model returns unstructured output
|
|
- `ProcessSpec` managed blocks wired into all four GPU profiles (6 / 8 / 16 / 24 GB)
|
|
- `--gpu-id` flag respected via `CUDA_VISIBLE_DEVICES`
|
|
|
|
**Documents module** (`circuitforge_core.documents`) — closes #7
|
|
- `ingest(image_bytes, hint) → StructuredDocument` — single call for all consumers
|
|
- Primary path: cf-docuvision HTTP service; automatic fallback to `LLMRouter` vision; graceful empty doc on total failure
|
|
- `StructuredDocument`, `Element`, `ParsedTable` frozen dataclasses with `.headings` / `.paragraphs` convenience properties
|
|
- `CF_DOCUVISION_URL` env var for service URL override
|
|
- `DocuvisionClient`: reusable HTTP client for cf-docuvision with `is_healthy()` probe
|
|
|
|
**Coordinator probe loop tests** — closes #13
|
|
- 4 async tests for `_run_instance_probe_loop`: healthy transition, timeout eviction, state cleanup, no-URL guard
|
|
|
|
---
|
|
|
|
## [0.2.0] — 2026-04-02
|
|
|
|
### Added
|
|
|
|
**Orchestrator — auto service lifecycle**
|
|
- `ServiceRegistry`: in-memory allocation tracker with state machine (`starting → running → idle → stopped`)
|
|
- `NodeSelector`: warm-first GPU scoring — prefers nodes already running the requested model, falls back to highest free VRAM
|
|
- `/api/services/{service}/allocate` coordinator endpoint: auto-selects best node, starts the llm_server process via agent, returns URL
|
|
- `CFOrchClient`: sync + async context managers for coordinator allocation/release
|
|
- Idle sweep in `AgentSupervisor`: stops instances that have been idle longer than `idle_stop_after_s` (default 600 s for vllm slot)
|
|
- Background health probe loop: coordinator polls all `starting` instances every 5 s via `GET /health`; promotes to `running` on 200, marks `stopped` after 300 s timeout (closes #10)
|
|
- Services table in coordinator dashboard HTML
|
|
- `idle_stop_after_s` field in service profiles
|
|
|
|
**LLM Router**
|
|
- cf-orch allocation support in `LLMRouter` backends
|
|
- VRAM lease acquisition/release wired through scheduler batch workers
|
|
|
|
**Scheduler**
|
|
- cf-orch VRAM lease per batch worker — prevents over-subscription
|
|
- `join()` on batch worker threads during shutdown
|
|
|
|
**HF inference server** (`llm_server.py`)
|
|
- Generic HuggingFace `transformers` inference server replacing Ouro/vllm-Docker-specific code
|
|
- `ProcessSpec` wiring in agent `service_manager.py`
|
|
- Handles transformers 5.x `BatchEncoding` return type from `apply_chat_template`
|
|
- Uses `dtype=` kwarg (replaces deprecated `torch_dtype=`)
|
|
|
|
### Fixed
|
|
|
|
- VRAM pre-flight threshold tightened: coordinator and `NodeSelector` now require full `service_max_mb` free (was `max_mb // 2`), preventing instances from starting on GPUs with insufficient headroom (closes #11 / related)
|
|
- `ServiceInstance` now seeded correctly on first `/allocate` call
|
|
- TTL sweep, immutability, and service-scoped release correctness in allocation path
|
|
- Coordinator logger added for allocation path visibility
|
|
|
|
### Changed
|
|
|
|
- Removed Ouro/vllm-Docker specifics from llm_server — now a generic HF inference endpoint
|
|
|
|
---
|
|
|
|
## [0.1.0] — 2026-03-01
|
|
|
|
### Added
|
|
|
|
- Package scaffold (`circuitforge_core`)
|
|
- DB base connection and migration runner
|
|
- Generalised tier system with BYOK (bring your own key) and local-vision unlocks
|
|
- LLM router extracted from Peregrine (fallback chain, vision-aware, BYOK support)
|
|
- Config module and vision router stub
|
|
- cf-orch orchestrator: coordinator (port 7700) + agent (port 7701)
|
|
- Agent registration + VRAM lease wiring
|
|
- Coordinator dashboard (HTML)
|