--- title: Virtual-LPU sdk: static emoji: 📊 colorFrom: blue colorTo: indigo short_description: Find bottlenecks in PyTorch training in real time. license: apache-2.0 --- app_file: index.html pinned: false # LPU AI Accelerator V7 — README# LPU AI Accelerator V7 — README **Intelligent Training Control Center & Adaptive Runtime.** Runs beside your training (VS Code → your python → V010) as an independent service: auto-detects the training process, shows live hardware + training state in a control-center dashboard, classifies bottlenecks with measured evidence, and governs parameter changes through a 3-level safety model with full audit trail and rollback. * Build on: LPU Virtual Runtime **V6** core (reused verbatim; only `version.py` bumped to 7.0.0 and an additive protocol extension). * Server: FastAPI + WebSocket + SQLite/JSONL (torch-free footprint). * Dashboard: single self-contained HTML (no build step). * Integration SDK: one stdlib-only file the trainer imports (fail-open). * 56 automated tests green in the reference sandbox; 2 hardware-gated rows must run on the 2×P100 rig via `scripts/rig_tests.ps1`. ## Quick start (Windows) ```powershell & "C:\Ipara.co\Scripts\python.exe" -m pip install -r requirements.txt & "C:\Ipara.co\Scripts\python.exe" .\run_lpu.py --desktop # NATIVE window (no browser) & "C:\Ipara.co\Scripts\python.exe" .\run_lpu.py # browser dashboard ``` Then press **DETECT PROJECT** (or let auto-detection find your running training) and **CONNECT TO TRAINING**. For deep telemetry inside the trainer, add `adapters/lpu_accelerate.py` to your project — see `docs/INSTALL_WINDOWS.md` §4. ## Documents | File | Content | |---|---| | `docs/AUDIT_REPORT_V7.md` | Phase 1 source audit (capabilities, gaps, V7 map) | | `docs/IMPLEMENTATION_AND_TEST_REPORT.md` | Step-by-step build + test evidence | | `docs/INSTALL_WINDOWS.md` | Install, run, SDK integration, security, rig tests | | `docs/API_REFERENCE.md` | Every REST + WebSocket endpoint | | `docs/test_run_detail.txt` | Raw pytest output (reference sandbox) | ## Non-negotiables honored (mission §21) Read-only discovery (hash-tested) · measured data only, unmeasured = "not measured" · level-3 parameters never auto-applied · every change journaled with old/new/reason and rollback · stopping the accelerator never stops training · loopback-only by default · token auth mandatory when exposed.