feature(docker): docker + CI Integrated

Author: Julina Maharjan

memory.md

@@ -1,11 +1,29 @@
 # memory.md — AtlasAI (active state)
 
-**Updated:** 2026-07-20 · **Milestone:** Week 2 done + real backend added → Week 3 next
+**Updated:** 2026-07-21 · **Milestone:** Week 3 (Docker Platform) in progress — CPU path built, GPU path written (user-verified)
 
 ## Progress
 - Wk1 API Gateway ✅ — auth, validation, JSON logging, rate limiting; 17 tests.
 - Wk2 Inference Engine ✅ — pluggable backends, BatchManager, SSE streaming,
   engine-gated `/ready`; gateway services async over engine.
+- Wk3 Docker Platform 🚧 — `deployment/docker/`: two separate Dockerfiles
+  (`Dockerfile.cpu` mock+hf, multi-stage slim non-root venv; `Dockerfile.gpu`
+  CUDA 12.1 + py3.12/deadsnakes + vllm) and `docker-compose.yml` (CPU api
+  service + hf-cache volume + healthcheck; commented Postgres/Redis/MLflow
+  stubs). Root `.dockerignore` trims context to source. Images COPY only
+  api/+shared/+inference/ (verified: no imports of future-week packages);
+  HEALTHCHECK curls `/health`. Chosen scope: two Dockerfiles + API-only compose
+  with commented stubs. GPU build NOT run in Claude env — user verifies on GPU.
+  CPU+hf image verified running in container by user. Volume-perms fix: image
+  now `mkdir -p` + `chown atlas` on `/home/atlas/.cache` so the named hf-cache
+  volume initializes writable for the non-root user (was root-owned → Errno 13).
+- Wk3 CI ✅ — `.github/workflows/ci.yml`: job `lint-test` (ruff + pytest, py3.12,
+  mock backend) → job `docker-cpu` (build CPU image `INCLUDE_HF=0` for speed,
+  run container, smoke-test `/health`=ok and `/ready`=mock). First feature run
+  through the pipeline: **Stop button** on `/tester` — AbortController cancels
+  the in-flight Generate/Stream fetch (`stopRun()`, buttons enable/disable,
+  AbortError → "stopped"). New test `test_tester_page_has_stop_control` (→ 40
+  tests). Verified: tester.py compiles + Stop tokens present (mock check).
 - Real backend ✅ — added **HFBackend** (`inference/backends/transformers_backend.py`):
   real transformers model on CPU/MPS/CUDA, async (to_thread), streaming via
   TextIteratorStreamer, lazy torch import. Backends now: `mock` | `hf` | `vllm`,
@@ -18,6 +36,11 @@
 - GPU server: `requirements-inference.txt` → `ATLAS_INFERENCE_BACKEND=vllm uvicorn api.main:app --host 0.0.0.0`
 - Browser test: `/docs` (Swagger, Authorize `dev-key-local`) OR `/tester` (custom
   page, supports Generate + live Stream).
+- Docker (CPU): from repo root `docker build -f deployment/docker/Dockerfile.cpu
+  -t atlasai:cpu .` then `docker run --rm -p 8000:8000 atlasai:cpu` (mock). Or
+  `cd deployment/docker && docker compose up --build` (add
+  `ATLAS_INFERENCE_BACKEND=hf` for real gen). GPU: `Dockerfile.gpu` on a CUDA
+  host with `--gpus all` (user-verified only).
 
 ## Notes / fixes
 - requirements-hf.txt uses version FLOORS (torch>=2.2.0 …): macOS-x86_64 has no
@@ -44,18 +67,22 @@
   excluded from OpenAPI + rate limit). 39 tests pass.
 
 ## Right now
-Real-backend workflow implemented + documented. Next: user browser-tests `hf`
-locally. Then resume **Week 3 — Docker Platform** (image strategy/compose still TBD).
+Week 3 Docker Platform + CI done. CPU+hf image runs in container; CI workflow
+lints, tests, builds the CPU image, and smoke-tests the running container. Stop
+button shipped as the first change through the pipeline. Next: user pushes so CI
+runs green on GitHub, runs `pytest` locally (expect 40), and validates the GPU
+image on CUDA hardware. Then checkpoint Wk3 (tag `v0.3.0-week3`).
 
 ## Open follow-ups
 - First `hf` request downloads model (~1GB) → cached; `/ready` 503 until loaded.
 - HF/vLLM real runs not testable in Claude's env (no torch/GPU).
 - Later: dedicated embedding backend; LoRA via `GenerationRequest.adapter` (Wk5).
 
 ## Immediate next steps
-1. Checkpoint Wk2: branch `week-2-inference-engine` → PR → tag `v0.2.0-week2`.
-2. Start **Week 3 — Docker Platform**: Dockerfile(s) + docker-compose, reproducible
-   builds; likely split CPU (mock) vs CUDA (vllm) images. Confirm scope first.
+1. User builds/runs CPU image (`docker compose up --build`), verifies `/health`,
+   `/ready`, `/tester`; then verifies GPU image (`Dockerfile.gpu`) on CUDA host.
+2. Checkpoint Wk3: conventional commit `feat(docker): ...`, tag `v0.3.0-week3`.
+3. Ask before advancing to Week 4.
 
 ## Open follow-ups
 - Cleanup pending: remove superseded Wk1 files (deletion was declined earlier):