[Comment-XD]

#6 Run the training backend locally without AWS

Github Issue Number Here: #6

What user problem are we solving?

New contributors can't bring the training backend up on a fresh checkout without an AWS account, and the setup steps aren't written down. On top of that, two issues block a clean local run:

• Unpinned pytorch/torchvision in environment.yml resolve to a build incompatible with Python 3.9 — import torch fails with ImportError: cannot import name 'TypeIs'.
• The training Dockerfiles run poetry install before README.md is present, so the build fails.

What solution does this PR provide?

• Documents a verified local run of the Django + Celery + Redis training backend with no AWS credentials.
• Pins pytorch=2.5.1 / torchvision=0.20.1 in training/environment.yml so the environment resolves against Python 3.9.
• Fixes the training Dockerfiles (training/Dockerfile and training/training/core/celery/Dockerfile) so README.md is copied in before poetry install.

Testing Methodology

How did you test your changes and verify that existing functionality is not broken

Brought the backend up locally (Django + Celery + Redis, no AWS) and verified:

• GET /api/test → 200
• Redis ping → PONG
• pytest → 17 passed / 1 failed. The single failure (CALIFORNIA_HOUSING) is an environment permission gap in the local setup, not a regression in the pipeline.

Any other considerations

The CALIFORNIA_HOUSING test failure is a local environment permission issue rather than a code bug; flagging it so reviewers don't treat it as a regression.

---

Note

Low Risk
Documentation and local Docker/conda build fixes; PyTorch pin narrows dependency resolution but targets known Python 3.9 compatibility.

Overview
Adds a verified no-AWS path for bringing up Django, Celery, and Redis via docker compose up --build, including curl/Redis/pytest checks and notes that full POST /api/tabular still needs AWS for Firebase and S3.

Fixes two blockers called out in the docs: pytorch=2.5.1 / torchvision=0.20.1 in environment.yml so conda resolves a Python 3.9–compatible stack (unpinned builds broke on TypeIs), and both training Dockerfiles now copy README.md before poetry install (required by pyproject.toml) and use poetry install --no-root in the development target.

^{Reviewed by Cursor Bugbot for commit 155695b. Bugbot is set up for automated code reviews on this repo. Configure here.}

[karkir0003]

thought for a followup issue (might not be the most urgent thing):

when i originally created DLP, I used an environment.yml setup to make compatible for mamba, but now that uv has been around for a while and the package manager is faster + simpler to use, have we considered taking a followup to use uv for the environment config

https://docs.astral.sh/uv/ @Comment-XD

[Comment-XD]

sorry for the late response. Yes, this is something we can implement to make package installation faster.

[karkir0003]

sounds good. @Comment-XD can you create a followup issue for that

[alessiopagliarulo]

Thanks Branden! Verified: all checks pass after updating with the latest nextjs (including the backend tests with your pytorch 2.5.1 pin). Merging.