Skip to content

install-gpu.md

Latest commit

 

History

History
69 lines (52 loc) · 2.45 KB

install-gpu.md

File metadata and controls

69 lines (52 loc) · 2.45 KB
Raw

GPU install (full score() pipeline)

The base cortex-score install is CPU-only - it covers score_from_predictions / score_from_prediction_bundle, which take a prediction tensor you already have. To run the full score("clip.mp4") pipeline you also need TRIBE v2 and its GPU stack.

TRIBE v2 is not a declared dependency of cortex-score: PyPI rejects published metadata that contains direct-URL (Git) dependencies, so it must be installed separately. cortex-score[gpu-deps] declares the compatible environment (torch / transformers / moviepy versions), and the pinned TRIBE commit is installed from a requirements file.

Requirements

  • A CUDA-capable GPU (TRIBE v2 weights are ~12 GB; plan for ≥16 GB VRAM).
  • ffmpeg and uvx (from uv) on PATH - TRIBE's preprocessing shells out to them.
  • A Hugging Face account with access to the gated Llama 3.2-3B weights.

Install

# 1. Base package + the TRIBE-compatible GPU dependency matrix
#    (torch / torchvision / torchaudio / transformers / moviepy / uv)
pip install "cortex-score[gpu-deps]"

# 2. TRIBE v2 itself + whisperx, pinned to the tested commit.
#    requirements/tribev2-gpu.txt lives in this repo - clone it first, or
#    use the raw URL:
#      pip install -r https://raw.githubusercontent.com/madhavcodez/cortex-score/main/requirements/tribev2-gpu.txt
pip install -r requirements/tribev2-gpu.txt

# 3. External tools (example: Debian/Ubuntu)
sudo apt-get install -y ffmpeg
curl -LsSf https://astral.sh/uv/install.sh | sh   # provides `uvx`

# 4. Authenticate for the gated model weights
huggingface-cli login

requirements/tribev2-gpu.txt pins TRIBE v2 to commit 34f52344e5ba96660fac877393e1954e399d3ef3, which matches cortex_score.runners.tribev2.TRIBEV2_PINNED_REVISION. Bumping one requires bumping the other and re-running the GPU smoke test.

Verify

cortex-score doctor

doctor reports Python, cortex-score, torch (+ CUDA), tribev2, ffmpeg, uvx, the Hugging Face token, and the cache directory. Every row should read ok (or report what to install) before you run score().

Run

from cortex_score import score

result = score("clip.mp4")
result.save("clip.score.json")

TRIBE v2 is licensed CC-BY-NC-4.0. Scores produced through the full score() path inherit the non-commercial restriction; it is emitted as a runtime warning on first load and recorded in every ScoreResult.license_restrictions.