Danchi-1/lodevem

Low-resource Device Virtual Emulator — A benchmarking harness for evaluating compressed PyTorch models against simulated low-cost Android device profiles, without requiring physical target hardware. # lodevem > **Lo**w-resource **De**vice **V**irtual **Em**ulator — benchmark PyTorch models against simulated low-cost Android device profiles, without physical target hardware. --- ## What It Does `lodevem` answers the question: *"If a farmer in rural Ghana with a Nokia C1 (1GB RAM) runs my cocoa disease model, will it work? How fast? Will it crash?"* Or, *"Can this $100 budget Android phone generate tokens from a quantized 2B parameter LLM without running out of memory?"* You bring your model files (already compressed however you like). lodevem benchmarks each one against a library of 16 real device profiles spanning budget Android phones, Android Go devices, and KaiOS feature phones — and produces a results table ready for your paper. **lodevem does not compress your model.** That is your responsibility as the researcher. --- ## How It Works ``` Your .pt model files (cocoa_fp32.pt, cocoa_int8.pt, tiny_llm.pt ...) │ ▼ ┌─────────────────┐ │ lodevem │ benchmarks each model against each device profile └────────┬────────┘ │ ┌────────┴─────────┐ ▼ ▼ nn-Meter psutil / Docker Latency Prediction RAM Measurement (per device SoC) (per device profile) │ │ └────────┬─────────┘ ▼ Results Table (console + CSV file) ``` **New in 0.2.0:** Full support for Large Language Models (LLMs) via Hugging Face. Automatically measures Time-To-First-Token (TTFT) and Tokens-Per-Second (TPS) on simulated budget devices. ### Two measurement modes — selected automatically | Environment | Mode | What it does | |---|---|---| | Kaggle, Colab, any machine | **Lite** (default) | Uses `psutil` to measure real peak RAM. No containers needed. | | Linux with Docker running | **Full** | Spins up a RAM-capped container. Hard OOM detection enforced by the kernel. | You don't choose the mode — lodevem detects Docker automatically and uses whichever is appropriate. --- ## Installation ```bash pip install lodevem ``` **Note:* …