Add Nvidia Spark GB10 Docker image

[ich777]

• Add Dockerfile Dockerfile.cuda-spark with optimized build flags
• Modify build/publish matrix to support arm64 build for Nvidia Spark GB10

[ich777]

@leejet is that something that can be implemented?
It would be super cool to also have an image for the Nvidia Spark GB10

[leejet]

I think the idea makes sense, but I don’t think this needs a separate Dockerfile.cuda-spark in its current form. From the diff, the Spark variant is still using the same CUDA backend and build flow as the existing CUDA image; the main differences appear to be build parameters:

• platform=linux/arm64
• CUDA_VERSION=13.0.0
• UBUNTU_VERSION=24.04
• CMAKE_CUDA_ARCHITECTURES=121
• GGML_CUDA_FA_ALL_QUANTS=ON

So this looks more like a CUDA build variant for DGX Spark / GB10 rather than a distinct Dockerfile target that needs to be maintained separately.

I would suggest parameterizing the existing Dockerfile.cuda instead of duplicating most of it into Dockerfile.cuda-spark. For example, Dockerfile.cuda could expose build args such as:

ARG CUDA_VERSION=...
ARG UBUNTU_VERSION=...
ARG CUDA_ARCHITECTURES=...
ARG GGML_CUDA_FA_ALL_QUANTS=...

[ich777]

So this looks more like a CUDA build variant for DGX Spark / GB10 rather than a distinct Dockerfile target that needs to be maintained separately.

Oh okay, I can change that.
My initial thought for a dedicated file was that the Spark uses a somewhat unique combination (SM 121 + CUDA 13.0, where other targets use either SM 120 or SM 121 with CUDA 13.1+) but you are probably right, I'll change the PR if that would the repo more nicely.