Stanza service: CPU-only torch, pinned dependencies

[mircealungu]

Tonight's Stanza rebuild installed PyPI's default torch, which brings about 2.5 GB of CUDA libraries (nvidia-cudnn, cublas, nccl, cusparse, ...). The server has no GPU, so they never load. They only made the image's main layer 3.09 GB and took up most of a 3-minute build.

• torch from PyTorch's CPU index: torch==2.14.0+cpu, the same version as tonight's build without CUDA.
• pinned to what that build installed and ran with: flask 3.1.3, gunicorn 26.2.0, psutil 7.2.2, numpy 2.5.3, protobuf 7.36.2. Before, each rebuild could silently pick up new major versions.

Checked without building (no Docker here): pip install --dry-run --platform manylinux_2_28_x86_64 --python-version 3.12 --only-binary=:all: resolves to 27 packages instead of 46, with no nvidia, CUDA or triton packages.

Runtime: tokenization speed should be unchanged, since both builds run the same CPU code. Container start and image builds get faster; memory may drop slightly.

Takes effect on the next Stanza rebuild. Do it at a quiet moment, because a Stanza restart briefly breaks uncached tokenization. Recreate stanza_crawl when no crawler_* container is running:

ssh zeeguu@$ZORGIP 'cd ~/ops/running/api && git -C ~/src/api pull --ff-only && docker compose build stanza && docker compose up -d --no-deps stanza'
ssh zeeguu@$ZORGIP 'docker ps --filter name=crawler_ --format "{{.Names}}"'   # wait until empty
ssh zeeguu@$ZORGIP 'cd ~/ops/running/api && docker compose up -d --no-deps stanza_crawl'

After the restart, /health should report pipelines_loaded: 30.

🤖 Generated with Claude Code

[github-actions]

ArchLens - No architecturally relevant changes to the existing views