One GPU. Full LLM workflow. Real benchmarks. No cloud required.

llm-gpu-lab is a hands-on, end-to-end LLM toolkit you can run on a single NVIDIA GPU. It walks you from "I have a 4080 and curiosity" to a fine-tuned small LLM, a GGUF deployment, and a self-contained HTML benchmark report — without any paid hosted inference and without any cloud training.

Every number in the report comes from a JSON artifact written on the machine that ran it. There are no fake screenshots, no hand-edited numbers, and no "expected" results substituted for measured ones.

A developer with one consumer NVIDIA GPU (RTX 4080 / 4070 Ti SUPER / 3090 class, 12–24 GB) who wants to learn — or demonstrate — the full local LLM workflow:

1. environment diagnosis
2. tokenizer training
3. tiny GPT pretraining from scratch
4. generation from the trained checkpoint
5. LoRA / QLoRA supervised fine-tuning of a small open model
6. lightweight evaluation
7. (optional) lm-evaluation-harness integration
8. GGUF export and llama.cpp serving
9. GPU benchmarking and a reproducible HTML report

If you already know all of those, this is also a clean, opinionated template you can extend.

Two parallel paths share the tokenizer + eval + benchmark + report infrastructure: a from-scratch TinyGPT pretraining path (top) and an open-Hub model + LoRA / QLoRA SFT path (middle), with an optional GGUF export + llama.cpp serve branch off the SFT path. Every box writes a JSON artifact under results/<gpu>/, and the final HTML report renders all of them together.

# 1. Create a venv (uv recommended; plain python -m venv also works)
uv venv --python 3.11 .venv
source .venv/Scripts/activate

# 2. Install PyTorch with the CUDA wheel that matches your driver
pip install --index-url https://download.pytorch.org/whl/cu124 "torch>=2.4"

# 3a. Install from PyPI (most users)
pip install "llm-gpu-lab[nlp,hub]"

# 3b. ...or from source if you want to hack on it
git clone https://github.com/DaoyuanLi2816/llm-gpu-lab && cd llm-gpu-lab
pip install -e ".[dev,nlp,hub]"

# 4. Verify the environment
python -m llm_gpu_lab doctor --out results/rtx4080/environment.json

# 5. Run the full smoke pipeline
make smoke

# 6. Open the HTML report
xdg-open results/rtx4080/report.html      # Linux
open    results/rtx4080/report.html      # macOS
start   results\rtx4080\report.html      # Windows

The smoke pipeline takes about 30 seconds of compute on an RTX 4080 plus a one-time ~270 MB Hugging Face download for SmolLM2-135M.

These are the literal numbers from results/rtx4080/benchmark_summary.json and results/rtx4080/pretrain_metrics.json after running make smoke on the maintainer's machine. Re-run on yours and the report will overwrite them with your numbers.

These deliberately include a small failure mode: the basic eval scores 50 % because some of its arithmetic prompts use the first regex-matched number as the answer, and SmolLM2-135M echoes the operands first. We keep that honest — it is a real artifact of the chosen extraction rule, not a bug we hid.

python -m llm_gpu_lab doctor           --out results/rtx4080/environment.json
python -m llm_gpu_lab train-tokenizer  --config configs/pretrain/tiny_10m_smoke.yaml
python -m llm_gpu_lab pretrain         --config configs/pretrain/tiny_10m_smoke.yaml
python -m llm_gpu_lab generate         --checkpoint artifacts/checkpoints/tiny_10m_smoke/final.safetensors \
                                       --prompts examples/prompts/generation_prompts.txt \
                                       --out results/rtx4080/generation_samples.json
python -m llm_gpu_lab sft              --config configs/sft/smollm2_135m_lora_fallback.yaml
python -m llm_gpu_lab eval             --config configs/eval/smoke_eval.yaml
python -m llm_gpu_lab bench-gpu        --out results/rtx4080/benchmark_summary.json
python -m llm_gpu_lab export-gguf      --config configs/export/gguf_q4_k_m.yaml
python -m llm_gpu_lab serve-llamacpp   --model artifacts/gguf/smollm2_135m_lora.Q4_K_M.gguf --port 8080
python -m llm_gpu_lab report           --results-dir results/rtx4080 --out results/rtx4080/report.html
python -m llm_gpu_lab lm-eval          --base-model HuggingFaceTB/SmolLM2-135M-Instruct --task arc_easy --limit 20

Every command writes a machine-readable JSON artifact under results/rtx4080/. The HTML report only renders sections for artifacts that actually exist on disk.

llm-gpu-lab/
├── src/llm_gpu_lab/        # Python package (CLI, models, train, eval, …)
├── configs/                # YAML configs for pretrain / sft / eval / export
├── scripts/                # Shell helpers (setup_env, setup_llamacpp, smoke)
├── examples/prompts/       # eval prompts (jsonl) + generation prompts (txt)
├── docs/                   # quickstart, design, troubleshooting, licenses, …
├── tests/                  # pytest suite (CPU-friendly, GPU tests auto-skip)
├── results/rtx4080/        # committed JSON / HTML artifacts from real runs
├── pyproject.toml          # Python project metadata + ruff + pytest config
├── Makefile                # `make smoke`, `make test`, `make lint`, …
└── .github/workflows/ci.yml# CPU-only CI (lint + tests + audit)

Apache-2.0 — see LICENSE. The licence choice and the licences of major dependencies are explained in docs/licenses.md.

Start with docs/troubleshooting.md. It covers the issues we actually hit while building this repo: CUDA not available, bitsandbytes import errors, OOM during QLoRA, HF rate limits, llama.cpp build failures, GGUF conversion edge cases, Windows

• Unicode console issues, and more.

See docs/roadmap.md — that is also the only file where TODO / FIXME are allowed (enforced by scripts/audit_placeholders.sh).

CONTRIBUTING.md. Short version:

• make lint, make test, make audit must pass.
• Numbers in docs must come from a JSON in results/<gpu>/. No fake benchmark numbers.
• Public datasets / models only, with licences documented.