File size: 3,178 Bytes
4655dd2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 | # Architecture
## High-Level
```
Browser (Next.js 14) <--HTTP/SSE/WS--> FastAPI (Python)
| Recharts | Transformers + Torch
| Shadcn/ui | pynvml/psutil
| Tailwind | fpdf2
```
## Frontend (`frontend/src`)
- `app/layout.tsx`: RTL, `Tajawal` font, `Sidebar` + `MobileNav`
- `app/page.tsx`: Dashboard (KPIs + ModelLoader + TelemetryCharts)
- `app/playground/page.tsx`: `Playground.tsx` (SSE streaming, TPS/TTFT)
- `app/benchmark/page.tsx`: `BenchmarkPanel.tsx` (preset + CustomDatasetPanel + progress SSE modal)
- `app/hardware/page.tsx`: `TelemetryCharts` + Power vs VRAM scatter
- `app/models/page.tsx`: Validate + scan `GET /api/model/list`
- `app/share/[token]/page.tsx`: Public report view via `GET /api/share/{token}`
- `lib/api.ts`: `API_BASE = process.env.NEXT_PUBLIC_API_URL`, `apiFetch`, `streamGenerate`, `runBenchmarkStream`
- `components/ui/*`: Button, Card, Input, Badge, Tabs, Select, Progress (Shadcn)
## Backend (`backend/app`)
- `main.py`: `FastAPI` + `CORSMiddleware(*)` + 7 routers
- `model_manager.py`: `ModelManager` singleton
- `load_model(path, dtype, quantization, device_map)` → `AutoModelForCausalLM.from_pretrained` with `BitsAndBytesConfig` + `torch_dtype` + `device_map`
- Demo fallback if `!HAS_TRANSFORMERS` or `CUDA` error
- `generate_stream(req)` → `TextIteratorStreamer` + `Thread` or `_demo_generate`
- `telemetry.py`: `TelemetryService` — `psutil` + `pynvml.nvmlDeviceGetMemoryInfo` + `history[600]`
- `benchmark.py`: `BENCHMARK_SUITES` 15 tasks + `evaluate_answer()` (Exact/Regex/LLM) + `run_benchmark()`
- `dataset_parser.py`: `scan_folder()` → `parse_file()` (CSV/JSON/JSONL/TXT/MD) + `detect_language()` + `infer_category()`
- `schemas.py`: Pydantic models (`ModelLoadRequest`, `GenerateRequest`, `TelemetrySnapshot`, `BenchmarkReport`)
- `routers/`:
- `model.py`: `/api/model/*`
- `inference.py`: `/api/generate` (SSE)
- `telemetry.py`: `/api/telemetry/` + `GET /history` + `WS /ws/telemetry`
- `benchmark.py`: `/api/benchmark/*` including `POST /run-stream` (SSE)
- `custom_benchmark.py`: `/api/benchmark/custom/*`
- `share.py`: `/api/share/*` (in-memory `share_store`)
- `export.py`: `/api/export/{json,csv,pdf}` (Unicode font via `tahoma.ttf`)
## Data Flow
1. **Load:** UI `POST /api/model/load` → `model_manager.load_model()` → `info` with `num_parameters` estimated from file size
2. **Chat:** UI `POST /api/generate stream=true` → `StreamingResponse` `data: {"type":"token",...}` → UI updates `tps`
3. **Benchmark:** UI `POST /api/benchmark/run-stream` → SSE `progress` per task → `task_done` → `done` with `BenchmarkReport`
4. **Telemetry:** Poll `GET /api/telemetry/` every 1.2s + `history` for charts
5. **Export:** `GET /api/export/pdf?report_id=` → `fpdf2` with `Arabic` font + `reshaper`
## Persistence
- `reports_store: Dict[str, BenchmarkReport]` (in-memory, lost on restart)
- `share_store: Dict[str, report_id]` (in-memory)
- For production, replace with `sqlite`/`postgres`.
## Security Notes
- No auth (local-first). For public, add proxy auth and restrict `folder_path` to allowlist.
|