CUDA support

[dvcdsys]

Title: Add CUDA GPU support for inference

---

What

Adds a separate CUDA Docker image that runs embedding inference on NVIDIA GPUs.

Why

CPU inference is significantly slower on large codebases. Servers with NVIDIA GPUs can run inference on GPU automatically — embeddings.py already has CUDA auto-detection in place, it just needed a proper CUDA
image to run in.

How

• api/Dockerfile.cuda — Dockerfile based on nvidia/cuda:12.6.3-devel-ubuntu22.04 with Python 3.12 via deadsnakes PPA
• api/requirements-cuda.txt — copy of dependencies with a note (torch 2.6+ includes CUDA wheels on linux/amd64 by default)
• docker-compose.cuda.yml — Compose file with GPU reservation via NVIDIA Container Toolkit
• portainer-stack-cuda.yml — stack for Portainer deployment
• Makefile — new targets: server-docker-cuda, docker-build-cuda, docker-push-cuda
• Removed onnxruntime / optimum from both requirements files and onnx_model_path from config — ONNX was never used
• README — added Experimental section describing the CUDA image and host requirements

Published as dvcdsys/code-index:cuda. The CPU image :latest is unchanged.

Type of change

• New feature
• CI / infra

Checklist

• Tested against a running API server
• go vet ./... passes (CLI changes)
• pytest tests/ passes (API changes)
• No secrets or API keys committed

---

Note

Medium Risk
Adds a new CUDA-based container build/deploy path and changes Python dependencies, which could impact runtime packaging and deployment on GPU hosts. Core API logic is mostly unchanged but build/infra changes can break installs if misconfigured.

Overview
Adds an experimental NVIDIA GPU (CUDA) deployment option by introducing a new api/Dockerfile.cuda, docker-compose.cuda.yml, and portainer-stack-cuda.yml to run embedding inference on GPUs via the NVIDIA Container Toolkit.

Updates the Makefile with server-docker-cuda plus CUDA image build/push targets, and documents host requirements and usage in README.md. Also removes unused ONNX-related configuration/dependencies by dropping onnx_model_path from api/app/config.py and removing optimum/onnxruntime from api/requirements.txt.

^{Written by Cursor Bugbot for commit d725614. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

[cursor]

Duplicated requirements file risks silent dependency drift

Medium Severity

requirements-cuda.txt is an exact copy of requirements.txt (all 11 dependency lines are identical). If a future change adds or updates a dependency in one file but not the other, the CUDA image silently breaks. The Dockerfile.cuda could simply COPY api/requirements.txt requirements.txt instead, eliminating the duplicate entirely. Any CUDA-specific packages could be added via a separate RUN pip install step or by using -r requirements.txt inside a minimal CUDA-only file.

Additional Locations (1)

• api/Dockerfile.cuda#L26-L28