Installation¶

pip install figma-extractor

Or with conda / mamba from Anaconda.org:

conda install -c vickykumar7125 figma-extractor
# or
mamba install -c vickykumar7125 figma-extractor

That installs the core extractor and the figma-extractor command. It does not install LangChain, Transformers, PyTorch, or bitsandbytes. Pip selects one universal wheel, figma_extractor-<version>-py3-none-any.whl, on Linux, Windows, and macOS. The conda package is noarch Python and installs the same runtime dependencies. There is no CUDA-specific wheel, because this package does not contain compiled CUDA code.

Python 3.11 or newer is required.

Optional extras¶

pip install "figma-extractor[openai]"
pip install "figma-extractor[huggingface]"
pip install "figma-extractor[huggingface-local]"
pip install "figma-extractor[all-llm]"
pip install "figma-extractor[dev]"

Extras: llm, openai, anthropic, google, vertex, ollama, huggingface, huggingface-local, huggingface-quant, groq, xai, nvidia, cohere, together, deepseek, all-llm, dev, test.

huggingface-quant adds bitsandbytes. It does not select a CUDA build of PyTorch. Conda installs the extraction core only; use pip extras in the same environment when you need LLM providers.

PyTorch¶

Install PyTorch from the official index that matches the machine, as a separate command. Then install this package, or the other way around. Pip cannot choose between the CUDA 12.9, 13.0, and 13.2 indexes by itself.

Machine	Command
NVIDIA driver CUDA 13.2 or newer	`pip install -r requirements/cuda132.txt` from a checkout, or the pins in that file
NVIDIA driver CUDA 13.0 or 13.1	`requirements/cuda130.txt`
NVIDIA driver CUDA 12.9	`requirements/cuda129.txt`
CPU	`requirements/cpu.txt`
Intel XPU	`requirements/xpu.txt`
AMD ROCm on Linux	`requirements/gpu.txt`
macOS Apple Silicon	`requirements/macos.txt`

Those files also install the chat-provider stack. For extraction only, stop after pip install figma-extractor or conda install -c vickykumar7125 figma-extractor.

From a clone, python setup.py reads nvidia-smi and installs the matching file. python setup.py --core installs the editable package only.

Source¶

pip install .

If a wheel is not available for the platform, pip builds the sdist. The sdist needs a normal Python toolchain. It does not need a CUDA compiler.

Conda recipe (maintainers)¶

The repository recipe lives in conda-recipe/. From a checkout with conda-build:

conda build conda-recipe -c conda-forge --output-folder ./conda-bld
anaconda --site anaconda.org upload -u vickykumar7125 ./conda-bld/noarch/figma-extractor-*.conda

Do not commit Anaconda tokens. Store ANACONDA_API_TOKEN as a GitHub Actions secret for the Release workflow.