Skip to content

development.md

Latest commit

 

History

History
81 lines (59 loc) · 2.64 KB

development.md

File metadata and controls

81 lines (59 loc) · 2.64 KB
Raw

Development and local testing

Prerequisites

  • Python 3.12 for parity with the image
  • Node.js 24 and pnpm 11 for release tooling
  • Docker with Buildx
  • An NVIDIA GPU, compatible driver, and NVIDIA Container Toolkit for container runtime tests

macOS can run host-side tests and build linux/amd64 images through Buildx, but it cannot execute this CUDA worker locally.

Host-side setup

python3.12 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
corepack enable
pnpm install --frozen-lockfile
git clone https://github.com/vavo/LTX2.5-serverless.git
cd LTX2.5-serverless

Tests

Run the same host-safe validation used by CI:

bash -n src/*.sh scripts/*.sh tests/*.sh
bash tests/test_restore_snapshot.sh
bash tests/test_bootstrap_workspace.sh
bash tests/test_bootstrap_ltx25.sh

python3 -m unittest \
  tests.test_ltx_payload_builder \
  tests.test_ltx25_workflow \
  tests.test_verify_comfy_models \
  tests.test_workflow_support \
  -v

python3 -m json.tool .runpod/hub.json >/dev/null
python3 -m json.tool video_ltx2_5_i2v_API.json >/dev/null
docker buildx bake --print ltx2-5-distilled-int8 >/dev/null
docker buildx bake --print ltx2-5-distilled-int8-cu128 >/dev/null

tests/test_handler.py and tests/test_frontend_app.py are legacy integration-oriented modules and are not part of the current CI gate. Do not claim the full unittest discover suite passes unless those dependencies and mocks have been repaired.

Local GPU stack

All runtime source files are copied into the image; there is no source bind mount. Rebuild after changing Python or shell code:

docker compose down
docker build --platform linux/amd64 --target base -t ltx25-worker:dev .
docker compose up

The compose service uses RUN_MODE=local-api and exposes:

  • RunPod-compatible local API: http://localhost:8000
  • ComfyUI: http://localhost:8188
  • Bundled frontend: http://localhost:7777

Its persistent data is stored under ./data/runpod-volume. Set the LTX preload variables in an override file or environment if the local volume does not already contain the model stack.

GPU release gate

Before publishing a release:

  1. Build the exact bake target for linux/amd64.
  2. Boot it on the intended Blackwell GPU and CUDA-compatible driver.
  3. Confirm the preload is reused after restart.
  4. Run the checked-in I2V workflow through the handler.
  5. Verify the returned video and S3 mode if enabled.
  6. Repeat on the CUDA 12.8 target only if that fallback will be published.

Passing host tests proves the plumbing is coherent. It does not prove a 22B video model fits, starts, or renders on a GPU that was never involved.