Add vulkan builds for x86_64

[alaviss]

Vulkan images at https://quay.io/repository/ramalama/vulkan?tab=tags are only built for aarch64.

I've confirmed that the associated Containerfile can be built and works flawlessly with ramalama (using --image and --gpu flags) on x86_64 with very good performance (tested with a RX 580, not supported by ROCm).

[nikAizuddin]

Actually I didn't know Vulkan can do machine learning stuff, so I decided to benchmark my Vega 56 GPU and compare with ROCm. I've found that Vulkan is faster than ROCm.

ROCm benchmark:

Vulkan benchmark:

[azlan]

Sharing my benchmark result for 5700 XT

[rhatdan]

I created a multi-arch image for vulkan and pushed it to quay.io/ramalama/vulkan. Please try it out, reopen if it does not work for you.

[alaviss]

The image is still aarch64-only:

$ ramalama --gpu --image quay.io/ramalama/vulkan serve huggingface://unsloth/DeepSeek-R1-Distill-Qwen-7B-GGUF/DeepSeek-R1-Distill-Qwen-7B-Q5_K_M.gguf
Trying to pull quay.io/ramalama/vulkan:latest...
Getting image source signatures
Copying blob 7aa2efc24f31 done   |
Copying blob f999dbdae714 done   |
Copying blob 17f7af7a37d4 done   |
Copying blob be721eecd673 done   |
Copying config c3ae0ee2e8 done   |
Writing manifest to image destination
WARNING: image platform (linux/arm64) does not match the expected platform (linux/amd64)
{"msg":"exec container process `/usr/bin/llama-server`: Exec format error","level":"error","time":"2025-02-03T22:17:30.870142Z"}

[rhatdan]

Oops, try again.

[alaviss]

I can confirm that it's working. Will the versioned tags be updated as well?

[rhatdan]

Not sure what you mean?

There is a quay.io/ramalama/vulkan:latest and 0.5 version.

And this will be handled just like other images quay.io/ramalama/ramalama:latest and 0.5.