Skip to content

v1.3.2 — NER image optimization (−79%)

Latest
Latest

Choose a tag to compare

View all tags
@josortmel josortmel released this 01 Jul 13:05
· 3 commits to main since this release
v1.3.2
0d5da0a

NER image optimization

ecodb-ner image: 8.36 GB → 1.77 GB (−79%, −6.59 GB)

GLiNER runs CPU-only inference — the full CUDA stack (~5.3 GB of nvidia-cublas, nvidia-cuda-runtime, triton, etc.) was dead weight pulled as a transitive dependency of torch. This release pins torch==2.12.1+cpu via the official PyTorch CPU wheel index.

What changed

  • ner/Dockerfile: installs from requirements.txt with --extra-index-url https://download.pytorch.org/whl/cpu
  • ner/requirements.txt: explicit torch==2.12.1+cpu pin

Verified

  • Full regression: 7/7 containers healthy, 0 regressions
  • Entity extraction (GLiNER): end-to-end verified (save_memory → entity extraction → graph)
  • Semantic search (GAMR 8-stage): verified
  • Graph traversal: 1270 nodes, 3776 triples, 181 predicates
  • 5 MCP tools confirmed operational
  • Revert anchor: sha256:7504899c22d3 (pre-pin image)

Upgrade

docker compose build ner
docker compose up -d ner
# Wait ~30s for GLiNER model load, then verify:
docker compose ps  # ner should be healthy

No schema changes. No configuration changes. Drop-in replacement.

Assets 2
Loading