Circuit-Forge/circuitforge-core
5
0
Fork 0
Code Issues 5 Pull requests Projects Releases 8 Packages Wiki Activity Actions 2

8 releases 14 tags

RSS feed
  • v0.20.0 fb3a4c697d

    v0.20.0 — LLMRouter dict init + Ollama embed preflight
    Some checks failed
    CI / test (push) Waiting to run
    Details
    Mirror / mirror (push) Has been cancelled
    Details
    Release — PyPI / release (push) Has been cancelled
    Details
    Stable

    pyr0ball released this 2026-05-05 14:59:49 -07:00 | 7 commits to main since this release

    Closes #59 and #60.

    #59 — dict init: LLMRouter.__init__ now accepts Path | dict. Pagepiper ingest scripts pass a runtime-constructed config dict directly instead of writing a temp file.

    #60 — Ollama preflight: _check_ollama_model_pulled() checks GET /api/tags before the first embed() call. Raises RuntimeError("Fix: ollama pull <model>") if the configured embedding model is not pulled. Result cached per backend URL (one HTTP call per router lifetime). Silently skips for non-Ollama backends.

    Downloads
  • v0.17.0 185057d8ca

    v0.17.0 — Reranker module
    Some checks failed
    CI / test (push) Has been cancelled
    Details
    Mirror / mirror (push) Has been cancelled
    Details
    Release — PyPI / release (push) Has been cancelled
    Details
    Stable

    pyr0ball released this 2026-04-26 09:04:39 -07:00 | 30 commits to main since this release

    What's new

    circuitforge_core.reranker — shared reranker module for RAG pipelines across the orchard (closes #54)

    Five backends: BGE, Qwen3 (generative yes/no logit scorer), CrossEncoder (sentence-transformers), Cohere (BYOK cloud), Remote (cf-orch HTTP delegate). cf-reranker FastAPI service on port 8011.

    Auto cf-orch routing: make_reranker() detects CF_ORCH_URL and routes to cf-orch automatically — cloud apps get Qwen3 reranking with zero code changes.

    See CHANGELOG.md for full details.

    Downloads
  • v0.16.0 b52c578911

    v0.16.0 — shared CloudSession module
    Some checks failed
    CI / test (push) Waiting to run
    Details
    Mirror / mirror (push) Has been cancelled
    Details
    Release — PyPI / release (push) Has been cancelled
    Details
    Stable

    pyr0ball released this 2026-04-24 17:30:38 -07:00 | 32 commits to main since this release

    What's New

    circuitforge_core.cloud_session

    Shared cloud session resolution for all CF products — extracted from product-level hand-rolled implementations into a single tested module.

    • CloudUser dataclass: user_id, tier, product, has_byok, meta
    • CloudSessionFactory: JWT validation (Directus HS256), Heimdall tier resolution (5-min cache), guest session fallback with per-session anon UUID cookie
    • dependency() / require_tier(): FastAPI-compatible dependency factories — wired with Depends(get_session) and Depends(require_tier("premium"))
    • Auto-provision: calls Heimdall /admin/provision on first cloud login (idempotent)
    • BYOK detector hook: products pass a callable; has_byok surfaces in CloudUser

    Usage

    from circuitforge_core.cloud_session import CloudSessionFactory, CloudUser

_factory = CloudSessionFactory(product="myproduct", byok_detector=_detect_byok)
get_session = _factory.dependency()
require_tier = _factory.require_tier

    Full changelog: v0.15.0...v0.16.0

    Downloads
  • v0.10.0 38c2bd702a

    v0.10.0 — Community Module
    Some checks failed
    Mirror / mirror (push) Has been cancelled
    Details
    Release — PyPI / release (push) Has been cancelled
    Details
    Stable

    pyr0ball released this 2026-04-12 17:24:31 -07:00 | 58 commits to main since this release

    circuitforge-core v0.10.0

    Added

    circuitforge_core.community — shared community signal module (BSL 1.1)

    • CommunityDB: psycopg2 connection pool with run_migrations(). Idempotent on every startup.
    • CommunityPost: frozen dataclass capturing user-authored community posts with item snapshots.
    • SharedStore: base class for product-specific community stores with typed pg read/write helpers.
    • Migration 001: community_posts schema. Migration 002: community_reactions stub.
    • psycopg2-binary added to [community] optional extras.

    See CHANGELOG.md for full history.

    Downloads
  • v0.5.0 9544f695e6

    v0.5.0 — Cross-platform manage.py Stable

    pyr0ball released this 2026-04-02 23:05:22 -07:00 | 103 commits to main since this release

    What's new

    circuitforge_core.manage (closes #6)

    Replaces bash-only manage.sh with a cross-platform Python process manager that works on Linux, macOS, and Windows natively — no WSL2, no Docker required.

    Docker mode (auto-detected, or --mode docker): wraps docker compose / docker-compose

    Native mode (--mode native, default on Windows): PID-file process management

    • platformdirs for OS-appropriate PID and log paths
    • Cross-platform kill (SIGTERM/SIGKILL on Unix, taskkill /F on Windows)
    • Polling log tail — no tail -f dependency

    Getting started in a product:

    python -m circuitforge_core.manage install-shims
# → writes manage.sh, manage.ps1, manage.toml.example
cp manage.toml.example manage.toml  # edit for your services
./manage.sh start

    Tests

    279 total (+36 new): config (6), docker_mode (9), native_mode (21)

    Downloads
  • v0.4.0 6e3474b97b

    v0.4.0 — Agent watchdog + Ollama adopt-if-running Stable

    pyr0ball released this 2026-04-02 22:13:01 -07:00 | 106 commits to main since this release

    What's new

    Agent watchdog (closes #15)

    Coordinator restarts no longer require manual agent restarts. Known nodes are persisted to SQLite and restored on startup; remote nodes reappear within ~10 s automatically. Agent processes now run a 30 s reconnect loop so they re-register whenever the coordinator comes back.

    Ollama adopt-if-running (closes #16)

    adopt: true in any ProcessSpec lets cf-orch claim an already-running process rather than spawning a new one — the primary use case being Ollama running as a system daemon. All GPU profiles now have an Ollama managed block with adopt: true and health_path: /api/tags so Ollama VRAM is accounted for in the allocator and the service appears in the dashboard.

    Tests

    243 total (+27 new): NodeStore (8), AgentSupervisor watchdog (8), Ollama adopt (11)

    Downloads
  • v0.3.0 a36f469d60

    v0.3.0 — Hardware detection, cf-docuvision, documents pipeline Stable

    pyr0ball released this 2026-04-02 18:56:49 -07:00 | 110 commits to main since this release

    What's new

    Hardware module (circuitforge_core.hardware)

    • detect_hardware() → VRAM tier selection → generate_profile() for llm.yaml-compatible config
    • Supports nvidia-smi, rocm-smi, Apple silicon, CPU fallback

    cf-docuvision service (circuitforge_core.resources.docuvision)

    • FastAPI wrapper for ByteDance/Dolphin-v2 (Qwen2.5-VL, ~8 GB VRAM)
    • POST /extract with hint modes: auto / table / text / form
    • ProcessSpec managed: blocks wired into all GPU profiles (6/8/16/24 GB)

    Documents module (circuitforge_core.documents)

    • ingest(image_bytes, hint) → StructuredDocument — single entry point for all products
    • cf-docuvision primary path with LLMRouter vision fallback
    • Consumed by: kiwi (receipts/recipes), falcon (forms), peregrine (resume images), godwit (identity docs)

    Tests

    • 216 tests total (+58 new): hardware (31), docuvision (14), documents (22), coordinator probe (4)

    Closes #5, #7, #8, #13

    Downloads
  • v0.2.0 482c430cdb

    v0.2.0 — Orchestrator auto-lifecycle + HF inference server Stable

    pyr0ball released this 2026-04-02 17:25:06 -07:00 | 113 commits to main since this release

    What's New

    Orchestrator — auto service lifecycle

    • Health probe loop (closes #10): coordinator background task polls all starting instances every 5 s via GET /health; promotes to running on 200, marks stopped after 300 s
    • NodeSelector: warm-first GPU scoring — prefers nodes already running the requested model; falls back to highest free VRAM
    • ServiceRegistry: in-memory state machine tracking allocations across starting → running → idle → stopped
    • /api/services/{service}/allocate: auto-selects best node, starts llm_server via agent, returns URL
    • CFOrchClient: sync + async context managers for coordinator allocation/release
    • Idle sweep in AgentSupervisor: stops instances idle longer than idle_stop_after_s (default 600 s)
    • Services table in coordinator dashboard

    HF Inference Server

    • Generic HuggingFace transformers inference endpoint (replaces Ouro/vllm-Docker-specific code)
    • Handles transformers 5.x BatchEncoding from apply_chat_template
    • Uses dtype= kwarg (deprecates torch_dtype=)

    Bug Fixes

    • VRAM pre-flight (closes #11 / related): tightened threshold from max_mb // 2 to full max_mb — prevents starting instances on GPUs without sufficient headroom
    • ServiceInstance seeded correctly on first /allocate call
    • TTL sweep, immutability, and service-scoped release correctness

    LLM Router + Scheduler

    • cf-orch allocation support in LLMRouter backends
    • VRAM lease acquisition/release through scheduler batch workers
    • join() on batch workers during shutdown

    See CHANGELOG.md for the full history.

    Downloads