feat(docker): add CUDA 12.8 support for RTX 50-series GPUs

[sh3ll3x3c]

Description

Add CUDA 12.8 support for NVIDIA RTX 50-series (Blackwell/sm_120) GPUs.

The current Dockerfile.onnx.gpu uses CUDA 12.4 which doesn't support the new RTX 50-series architecture (sm_120). This PR adds a new Dockerfile that enables GPU inference on RTX 5090, 5080, 5070 Ti, and 5070 cards.

Key changes:

• CUDA 12.8.1 base image (required for sm_120 architecture)
• PyTorch nightly with cu128 support (stable PyTorch doesn't support sm_120 yet)
• onnxruntime-gpu from Microsoft's CUDA 12 index (default PyPI package lacks CUDAExecutionProvider for CUDA 12)
• flash_attn build skipped by default (optional, significantly reduces build time)

Related issue: Users with RTX 50-series GPUs cannot use GPU acceleration with the current Docker images.

Type of change

• New feature (non-breaking change which adds functionality)

How has this change been tested, please provide a testcase or example of how you tested the change?

Tested on NVIDIA GeForce RTX 5090:

1. Built the image:

   docker build -f docker/dockerfiles/Dockerfile.onnx.gpu.cuda128 \
     -t roboflow/roboflow-inference-server-gpu-cuda128 .

2. Verified CUDA provider is available:

   docker exec <container> python -c "import onnxruntime as ort; print(ort.get_available_providers())"
   # Output: ['TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider']

3. Tested inference speed (1080p image, object detection):

   • First request: ~54s (model download + load)
   • Subsequent requests: 65-100ms (GPU inference)

4. Verified GPU memory allocation via nvidia-smi

Any specific deployment considerations

• This Dockerfile uses PyTorch nightly builds (cu128) since stable PyTorch doesn't yet support sm_120
• Once PyTorch stable releases with CUDA 12.8/sm_120 support, the Dockerfile can be updated to use stable builds
• Users who need Paligemma/Florence2 support can uncomment the flash_attn build step (adds significant build time)

Docs

• Docs updated? What were the changes:

Documentation update suggested: Add a note in the Docker deployment docs mentioning Dockerfile.onnx.gpu.cuda128 for RTX 50-series GPU users.

[CLAassistant]

All committers have signed the CLA.