Safetensors Studio & Bench

Run, test, and benchmark local Safetensors models with surgical VRAM observability.

Safetensors Studio & Bench is a full-stack, local-first platform to load any *.safetensors + config.json + tokenizer.json folder, stream chat with real TPS/TTFT, run automated benchmarks (Reasoning/Coding/Arabic/Summarization), and profile GPU/CPU/RAM live — no cloud required.

📖 Arabic README: README_AR.md | 📚 Docs: docs/ | 🤝 Contribute: CONTRIBUTING.md

✨ Key Features

Area	What you get
Safetensors Loader	Local folder picker • `Float16 / Bfloat16 / Float32` • `4-bit NF4 / 8-bit` via `bitsandbytes` • `device_map: auto / balanced / cpu / cuda:0` + offloading
Live Playground	Token Streaming (SSE) • Real `Tokens/sec` & `TTFT` • Temperature / Top-P / Max Tokens • Copy & clear
Automated Benchmark	15 preset tasks: Reasoning (4), Coding (4), Arabic Quality (4), Summarization (3) • `Exact / Regex / LLM-as-Judge` • Custom datasets: drop any folder with `.csv/.json/.jsonl/.txt/.md` → auto-detects language (Arabic/English) & category
Hardware Profiling	`pynvml + psutil` → live VRAM/CPU/RAM/Power via `Recharts` • VRAM peak timeline • TPS vs VRAM scatter • 600-point history
Reports & Sharing	Export `PDF (Arabic-capable)` / `JSON` / `CSV` • Shareable link `/share/{token}` • Full tables + `by_category` insights

🌍 Supported Languages

Language	Coverage
English	UI, prompts, docs, code, benchmarks (`Reasoning`, `Coding`, `Summarization`)
Arabic (العربية)	Full UI RTL, benchmarks (`Arabic Quality`, `Summarization`), PDF with `tahoma.ttf` + `arabic-reshaper`, auto-detection `\u0600-\u06FF` for custom datasets
Mixed	Auto `language_counts: {ar, en}` per folder, per report

Add a new language: add prompts to backend/app/benchmark.py:14 and categories to dataset_parser.py:11. See SUPPORTED_LANGUAGES.md.

🚀 Quick Start

1. Requirements

Node 22+ / Python 3.12+
8GB RAM minimum (CPU mode), 12GB+ VRAM recommended for 7B models on GPU

2. Backend

cd backend
pip install -r requirements.txt          # torch CPU + transformers are included
# optional GPU: pip install torch --index-url https://download.pytorch.org/whl/cu121
uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload
# → http://localhost:8000/docs

3. Frontend

cd frontend
npm install
npm run dev
# → http://localhost:3000
# env: frontend/.env.local → NEXT_PUBLIC_API_URL=http://localhost:8000

4. Docker (alternative)

docker compose up --build
# frontend http://localhost:3000  backend http://localhost:8000

5. Load a model

Place a HF-style folder anywhere, e.g. C:\models\mistral-7b containing model.safetensors, config.json, tokenizer.json
In UI / or /models or new: top bar of /benchmark → paste path → Validate → Load
Or via API: POST /api/model/load {"model_path":"C:\\models\\my-model","dtype":"float16","quantization":"4bit","device_map":"auto"}

No GPU? The platform auto-enters Demo Mode (simulated generation) so you can still test all features.

📁 Project Structure

backend/app/
  main.py                 # FastAPI + CORS + routers
  model_manager.py        # Safetensors loader (torch dtype/quant/device_map)
  telemetry.py            # pynvml/psutil + 600-point history
  benchmark.py            # 15 tasks + evaluate_answer()
  dataset_parser.py       # CSV/JSON/JSONL/TXT/MD → BenchmarkTask + lang detect
  schemas.py
  routers/{model,inference,telemetry,benchmark,custom_benchmark,share,export}.py
frontend/src/
  app/{page, playground, benchmark, hardware, models, share/[token]}
  components/{ModelLoader, ModelPathSelector, Playground, BenchmarkPanel, CustomDatasetPanel, TelemetryCharts}
  lib/{api.ts, utils.ts}
example_dataset/          # 9-task mixed sample (AR/EN)
docs/                     # English developer docs

🔌 API Reference (core)

Method	Path	Description
`POST`	`/api/model/load`	Load Safetensors folder
`GET`	`/api/model/status`	Loaded info
`GET`	`/api/model/validate?path=`	Check `*.safetensors + config.json`
`POST`	`/api/generate` (SSE)	Stream tokens with `ttft_ms`, `tokens_per_sec`
`GET`	`/api/telemetry/`	Snapshot (VRAM/GPU/CPU/RAM)
`WS`	`/ws/telemetry`	Live push (fallback to HTTP polling)
`POST`	`/api/benchmark/run`	Sync benchmark (blocking)
`POST`	`/api/benchmark/run-stream`	SSE progress (`type: progress/task_done/done`)
`POST`	`/api/benchmark/custom/scan?folder_path=`	Scan custom folder
`POST`	`/api/benchmark/custom/run-from-folder`	Run on custom folder
`POST`	`/api/benchmark/custom/upload`	Multipart upload
`POST`	`/api/share/{report_id}`	Create share token
`GET`	`/api/share/{token}`	Fetch shared report
`GET`	`/api/export/{json,csv,pdf}?report_id=`	Export

Full spec: http://localhost:8000/docs

🧪 Benchmark Suites

Suite	Tasks	Judge
`reasoning`	arithmetic, sequence, logic puzzle, fraction	Regex `\b7\b`
`coding`	fibonacci, loop output, reverse_string, sorted	Regex `def\s+`
`arabic`	spelling, synonym, i'rab, summarization	Regex `ذهبت`, `فرح`
`summarization`	tech text, Transformer, bullet points	Regex `7 مليار`

Custom folder example (CSV):

prompt,expected,expected_regex,category,name
"ما مرادف سعيد؟",فرح,فرح,arabic,syn

📊 Hardware Profiling

telemetry.py:19 polls pynvml.nvmlDeviceGetMemoryInfo + psutil.virtual_memory every 1.2s
Recharts Area/Line/Scatter: VRAM timeline, CPU/RAM %, Power vs VRAM, TPS vs VRAM
vram_peak_mb tracked per task and globally

📄 Export & Sharing

PDF: Unicode via tahoma.ttf + arabic-reshaper + python-bidi (fallback to ? sanitization)
Share: in-memory share_store[token]=report_id → /share/{token} page

🛠️ Development

See docs/DEVELOPMENT.md and docs/ARCHITECTURE.md.

# Frontend
npm run build   # production
npm run lint

# Backend
python -m pytest  # (add tests)
pip install -r requirements.txt

🤗 Hugging Face

Use any HF model: download snapshots and point model_path to it.
Model card template: docs/HUGGINGFACE.md
Dataset parser auto-handles HF datasets exported as JSONL/CSV.

🤝 Contributing

Please read CONTRIBUTING.md and CODE_OF_CONDUCT.md. PRs for new languages, benchmarks, and quant backends are welcome!

🔒 Security

See SECURITY.md.

📝 License

MIT — see LICENSE.

Built with ❤️ for local AI — no cloud, full control.