devops : Add main-rocm Dockerfile

[berney]

Adds a Dockerfile for a ROCm build.
Based off code from https://github.com/kyuz0/amd-strix-halo-toolboxes/blob/main/toolboxes/Dockerfile.rocm-7.14.

The build and runtime targets:

GFX target	AMD GPU architecture	Devices
gfx1100	RDNA 3 dGPU	Radeon RX 7900 XTX, RX 7900 XT, RX 7900 GRE; Radeon PRO W7900, W7800
gfx1101	RDNA 3 dGPU	Radeon RX 7800 XT, RX 7700 XT; Radeon PRO W7700
gfx1102	RDNA 3 dGPU	Radeon RX 7600 XT, RX 7600
gfx1103	RDNA 3 iGPU	Radeon 780M and related RDNA 3 integrated GPUs
gfx1150	RDNA 3.5 APU	Ryzen AI 300-series / Strix Point APUs, including Radeon 890M-class graphics
gfx1151	RDNA 3.5 APU	Ryzen AI Max/Max+ and PRO 300-series / Strix Halo APUs, including Ryzen AI Max+ PRO 395 and Radeon 8060S
gfx1152	RDNA 3.5 APU	Certain newer Ryzen AI APUs, including Radeon 860M-class graphics
gfx1200	RDNA 4 dGPU	Radeon RX 9060 XT and RX 9060
gfx1201	RDNA 4 dGPU	Radeon RX 9070 XT, RX 9070 and RX 9070 GRE

This pull request introduces a new multi-stage Dockerfile, .devops/main-rocm.Dockerfile, to support building and running the project with AMD ROCm 7.14 for GPU acceleration. The Dockerfile is split into build and runtime stages, ensuring a clean and efficient image for deployment.

ROCm Dockerfile addition:

• Added .devops/main-rocm.Dockerfile with a build stage based on Fedora 44, installing ROCm 7.14 development packages and compiling the project with HIP and AMDGPU targets enabled.
• Added a runtime stage based on Fedora Minimal 44, installing only the necessary ROCm runtime and BLAS packages for supported GPU architectures, and copying the built application from the build stage.
• Configured environment variables (ROCM_PATH, HIP_PATH, PATH, LD_LIBRARY_PATH) for ROCm compatibility in both build and runtime stages.
• Created symbolic links in /opt/rocm to ensure compatibility with ROCm tools and libraries.
• Set the entrypoint to run the application in a Bash shell.

Testing

• Tested on a gfx1151 this works.

In my testing I did notice a transcription difference between running main-vulkan-306c88f4d1286aec1bf96e544632897886af5501 v.1.9.2 docker image and main-rocm. Multiple runs would provide consistent output, but the output of vulkan and rocm differed - I'm not sure why. For the most part it was little differences, where I feel ROCm was more accurate, but there was a chunk near the end that ROCm skipped that Vulkan did not. My testing was with -m ggml-large-v3-turbo-q5_0.bin and --vad --vad-model ggml-silero-v6.2.0.bin

Example small difference - first vulkan (less accurate) then ROCm (more accurate):

192c178,179
<  of the moment, and most importantly, be grateful.
---
>  Honor the moment.
>  And most importantly, be grateful.

The VAD segments are the same in both Vulkan and ROCm.

[danbev]

I tried this locally and it worked for me 👍