hydraedge 0.1.1

HydraEdge: cross-document linker (VSA × RL) – EMNLP 2025 system demo

HydraEdge Rapid-Startup Environment

(Python 3.10 · CUDA 12.1 · FAISS GPU 1.11 · VS Code ready)

*Goal *: give reviewers / team-mates a one-shot, GPU-enabled dev box with the exact stack we used for the EMNLP-2025 system-demo paper—no local Python conflicts, no CUDA hell.

---

0 · Prerequisites

Tool	Minimum version	Notes
Docker	20.10+	Docker Desktop is fine (enable GPU support in Settings → Resources → GPU).
NVIDIA Driver	R525+	Must support CUDA ≥ 12.1.
VS Code	1.85+	Plus the Remote - Containers extension.

---

1 · Build the image (one-off)

We pin Python 3.10 on top of a small CUDA runtime base and add FAISS-GPU.

# inside the repo root (contains Dockerfile)
docker build -t py310-cuda-faiss:12.1 .

If you prefer to pull rather than build, grab our pre-built image:

docker pull ghcr.io/your-org/py310-cuda-faiss:12.1

---

2 · Run the container

docker run --gpus all \
  -v "C:/Documents/All_github_repo/hydraedge:/workspace" \
  -p 8888:8888 -p 6006:6006 \
  --name hydraedge-dev \
  -it py310-cuda-faiss:12.1 bash

Flag	Why
--gpus all	exposes every GPU to the container.
-v local: /workspace	your code lives in the container.
-p 8888:8888 -p 6006:6006	ready for Jupyter & TensorBoard.
-it … bash	drop into a shell immediately; feel free to swap for python or zsh.

---

3 · GPU / FAISS sanity check (optional)

Inside the container run:

python - <<'PY'
import faiss, numpy as np, torch, os
print("Python  :",  os.sys.version)
print("CUDA    :",  torch.version.cuda)
print("GPUs    :",  faiss.get_num_gpus())
xb = np.random.rand(2,128).astype('float32')
index = faiss.index_cpu_to_all_gpus(faiss.index_factory(128, "Flat"))
index.add(xb)
print("Distances", index.search(xb, 1)[0].ravel())
PY

You should see GPUs : 1 (or more) and finite self-distances.

---

4 · Attach VS Code (Remote-Containers)

1. Install Remote - Containers extension (once).
2. Command Palette → “Remote-Containers: Attach to Running Container…” choose hydraedge-dev → VS Code reloads into /workspace.

4.1 Install the Python extension in the container

The first reload shows “Extensions (in Container)” at the top of the sidebar.

• Search Python → Install (again – this time inside the container).
• Reload when prompted.

4.2 Pick the interpreter

Command Palette → Python: Select Interpreter → /usr/bin/python The green ▶︎ Run / Debug buttons now appear.

---

5 · Dev-container auto-setup (optional but recommended)

Put this in .devcontainer/devcontainer.json so newcomers get the tools automatically:

{
  "name": "hydraedge-dev",
  "image": "py310-cuda-faiss:12.1",
  "extensions": [
    "ms-python.python",
    "ms-toolsai.jupyter"
  ],
  "settings": {
    "python.defaultInterpreterPath": "/usr/bin/python"
  },
  "postCreateCommand": "python -m pip install -r requirements.txt || true"
}

---

6 · Shut-down & cleanup

# inside the container
exit               # or Ctrl-D

# on host
docker stop hydraedge-dev
docker rm   hydraedge-dev     # optional

---

7 · Troubleshooting

Symptom	Fix
faiss.get_num_gpus() -> 0	The Docker runtime cannot see your GPU.• Check `docker info --format '{{json .Runtimes}}'` shows *nvidia*.• On Windows / WSL: enable “Use the WSL 2 based engine” + GPU in Docker Desktop.
No ▶︎ Run button in VS Code	Install thePython extension inside the container (see § 4.1).
AssertionError during FAISS test	Means faiss.index_cpu_to_all_gpus got 0 GPUs. Fix GPU visibility then rerun.
Slow build (downloads huge cuBLAS wheel)	First build always pulls ≥ 500 MB CUDA libraries. They are cached afterwards.