add containerfile that use native fedora-43 rocm (6.4.4)

[Djip007]

This is 2 containerfile:
I build and pre configure llama.cpp.
llama.cpp is installed in /opt/llama.cpp

• Containerfile.rocm-6.4.4
  this is a simple image, to use with podman. It run the llama-server by default, can be use:

# build:
podman build -t rocm-serve:6.4.4 -f Containerfile.rocm-6.4.4

# run llama-server with podman:
#  - mount simple model file (the toolbox start by default "/models/default.gguf":
podman run --rm --device /dev/dri --device /dev/kfd --privileged --network=host -v ~/LLM/openai_gpt-oss-120b/MXFP4.gguf:/models/default.gguf rocm-serve:6.4.4 --ctx-size 0 --jinja

#  - mount models path you need to define the "-m <model path>":
podman run --rm --device /dev/dri --device /dev/kfd --privileged --network=host -v ~/LLM:/models rocm-serve:6.4.4 --ctx-size 0 -m /models/openai_gpt-oss-120b/MXFP4.gguf --jinja

# run other llama exec:
podman run --rm --device /dev/dri --device /dev/kfd --privileged --network=host -v ~/LLM:/models --entrypoint="/opt/llama.cpp/bin/llama-bench" rocm-serve:6.4.4 -m /models/openai_gpt-oss-120b/MXFP4.gguf -b 8192 -ub 2048 -fa 1 -ngl 999 --mmap 0 -p "1,1,2,4,8,16,32,64,128,256,512,1024,2048,4096,8192"

# for interactive run you can change the entrypoint to bash:
podman run --rm -it --device /dev/dri --device /dev/kfd --privileged --network=host -v ~/LLM:/models --entrypoint="/usr/bin/bash" rocm-serve:6.4.4
# - next in the container:
/opt/llama.cpp/bin/llama-cli -m /models/openai_gpt-oss-120b/MXFP4.gguf --jinja --ctx-size 0

• Toolboxfile.rocm-6.4.4
  it is a true toolbox to use with ... toolbox with no config need on fedora:

# build:
podman build -t rocm-toolbox:6.4.4 -f Toolboxfile.rocm-6.4.4

# run (no need for "--device" .. etc.:
toolbox create -i rocm-toolbox:6.4.4 rocm-toolbox-6.4.4
toolbox enter rocm-toolbox-6.4.4

# example use in the toolbox:
/opt/llama.cpp/bin/llama-cli -m openai_gpt-oss-120b/MXFP4.gguf --jinja --ctx-size 0

[Djip007]

to remove when published on updates repos.

[Djip007]

for now pick the testing package, follow : https://src.fedoraproject.org/rpms/rocm-runtime
patch is in rocm-runtime-6.4.2-3.fc43

[Djip007]

look fedora-43 patch is now up-stream => no need for testing.
fedora-43+rocm-6.4.4 today update (23/01/2026) is back stable/working. With all latest:

• kernel: 6.18.5-200.fc43.x86_64
• rocm: 6.4.4 (with rocm-runtime-6.4.2-3.fc43)
• amd-gpu-firmware: 20260110-1.fc43

[kyuz0]

Thanks for the work on the PR. I might not merge this at the moment, as I want to update the Dockerfiles I already have, which are part of my automated build and test pipeline.

[Djip007]

Take your time. Create this PR only to share it.

Note: in this PR I install llama.cpp in /opt/llama.cpp did not take time to figure what to do to install on /usr/...

And thanks for all you work/time on this nice stric-halo !

[Djip007]

• add fedora-44/rocm-7.1.1
• for podman/Containerfile install llama.cpp in /usr/local (live it in /op/llama.cpp for Toolboxfile ...)