docker-compose.ml-worker.gpu.yml

# Hand ml-worker an NVIDIA card. Overlay, never used alone:
#
#   docker compose -f docker-compose.yml -f docker-compose.ml-worker.gpu.yml \
#     up -d ml-worker
#
# Inert scaffolding right now, not a feature this stack turns on. Per
# ml-gpu-coordinated-roadmap.md §1 decision 4: "Isolation Forest and HBOS
# stay on CPU. Only LSTM and embeddings may use CUDA" -- and both of those
# are gated behind Milestone D (temporal quality) and Milestone H (GPU
# acceleration, #67), neither of which is done. Milestone H's own first
# step requires a measured CPU baseline from #62 before touching CUDA at
# all. Applying this overlay today reserves a device that nothing in the
# worker's code path will use yet.
#
# Whether ml-worker gets its own model runtime here or shares the ghidra
# stack's ollama (analysis/ghidra/docker-compose.ghidra.yml) for the later
# embeddings phase is explicitly #67's decision, not assumed by this file.

services:
  ml-worker:
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]