Add conda-free Docker setup for the Gradio demo (CUDA 12.1)

[kunci115]

What

Adds a reproducible, conda-free way to run the image-to-3D Gradio demo (app.py) on a Linux server with an NVIDIA GPU, plus a plain venv script for non-Docker users. Aimed at people who hit friction setting the project up today on modern dependency versions.

Added

• Dockerfile — CUDA 12.1 + PyTorch 2.4.0 (cu121), GPU arch pinned to 8.9 (Ada / RTX 40-series, configurable via TORCH_CUDA_ARCH_LIST). Builds all required CUDA extensions (nvdiffrast, diffoctreerast, diff-gaussian-rasterization, kaolin, spconv, xformers) and uses the xformers attention backend (no flash-attn build needed).
• docker-compose.yml — GPU passthrough, port 7860, persistent HF model cache volume, GPU selection.
• .dockerignore
• setup_venv.sh — conda-free venv install mirroring the Docker dependency set.
• README — new "Installation without Conda (WSL2)", "Installation with Docker", and "Choosing a model for limited VRAM" sections.

Fixes folded in (needed to run on current dep versions)

• Install CUDA extensions with --no-build-isolation so they compile against the installed PyTorch.
• Pin transformers / huggingface_hub to versions compatible with gradio 4.44.1 (HfFolder import).
• Pin the gradio web stack (fastapi, starlette, pydantic) to gradio-4.44-era versions and patch gradio_client for tolerant JSON-schema parsing (fixes the APIInfoParseError / unhashable type: 'dict' startup crashes).
• Bind Gradio to 0.0.0.0:7860 so it is reachable from outside the container.

Verification

Built and verified serving on 2× RTX 4090 (CUDA 13.x driver):

docker compose up --build
# -> Running on local URL: http://0.0.0.0:7860
curl -s -o /dev/null -w "%{http_code}\n" http://localhost:7860   # 200

The existing conda setup.sh is left untouched; this is purely additive.

[kunci115]

@microsoft-github-policy-service agree