feat: add AMD ROCm support for RX 5700 XT (gfx1010 / RDNA 1)

[PelaGB17]

• Dockerfile.rocm: new multi-stage image based on rocm/dev-ubuntu-24.04:6.3.1-complete
  • Sets PYTORCH_ROCM_VERSION=6.3 so uv installs ROCm-backed PyTorch wheels
  • Sets HSA_OVERRIDE_GFX_VERSION=10.3.0 to map RDNA1 (gfx1010) to RDNA2 kernels
  • Adds container user to 'video' and 'render' groups for /dev/kfd + /dev/dri access
• docker-compose.rocm.yml: pre-built image compose file
  • Mounts /dev/kfd and /dev/dri (no NVIDIA plugin required)
  • group_add: video, render for GPU device node access
  • Documents per-generation HSA_OVERRIDE_GFX_VERSION values
• docker-compose.build.rocm.yml: local build variant using Dockerfile.rocm
• base_adapter.go: add GetPyTorchROCmVersion() and update GetPyTorchWheelURL()
  • PYTORCH_ROCM_VERSION env var selects ROCm wheel index (rocm6.3, etc.)
  • ROCm takes priority over CUDA; falls back to cu126 if neither is set
  • All existing adapters (WhisperX, PyAnnote, Voxtral, Parakeet, Canary) inherit ROCm wheel selection automatically via GetPyTorchWheelURL()

[icedream]

Was trying to run this (my system runs a gfx1100 card, the 7900 XTX, so I removed the HSA_OVERRIDE_GFX_VERSION setting), got following errors:

[+] Preparing Python environment
time=22:08:36 level="INFO " msg="Initializing unified transcription service"
time=22:08:36 level="INFO " msg="Initializing registered models in parallel..."
time=22:08:36 level="INFO " msg="Preparing NVIDIA Parakeet environment" env_path=/app/whisperx-env/parakeet
time=22:08:36 level="INFO " msg="Preparing NVIDIA Sortformer environment" env_path=/app/whisperx-env/parakeet
time=22:08:36 level="INFO " msg="Preparing WhisperX environment" env_path=/app/whisperx-env
time=22:08:36 level="INFO " msg="Preparing NVIDIA Canary environment" env_path=/app/whisperx-env/parakeet
time=22:08:36 level="INFO " msg="transcription model initialized" model_id=openai_whisper
time=22:08:36 level="INFO " msg="Preparing PyAnnote environment" env_path=/app/whisperx-env/pyannote
time=22:08:36 level="INFO " msg="Preparing Voxtral environment" env_path=/app/whisperx-env/voxtral
time=22:08:36 level="INFO " msg="Created buffered transcription script" path=/app/whisperx-env/parakeet/parakeet_transcribe_buffered.py
time=22:08:36 level="INFO " msg="Installing Sortformer dependencies"
time=22:08:36 level="INFO " msg="Installing PyAnnote dependencies"
time=22:08:36 level="INFO " msg="Installing Voxtral dependencies"
time=22:08:36 level="INFO " msg="Parakeet environment not ready, setting up"
time=22:08:36 level="INFO " msg="Environment already configured by Parakeet"
time=22:08:36 level="INFO " msg="Downloading Canary model" path=/app/whisperx-env/parakeet/canary-1b-v2.nemo
time=22:08:36 level="INFO " msg="Installing Parakeet dependencies"
time=22:08:51 level=ERROR msg="Failed to initialize diarization model" model_id=pyannote error="failed to setup PyAnnote environment: uv sync failed: exit status 1: Using CPython 3.12.3 interpreter at: /usr/bin/python3\nCreating virtual environment at: .venv\n  × No solution found when resolving dependencies for split (markers:\n  │ python_full_version >= '3.12' and platform_machine == 'x86_64' and\n  │ sys_platform == 'linux'):\n  ╰─▶ Because pyannote-audio==4.0.2 depends on torch==2.8.0 and only the\n      following versions of torch are available:\n          torch<2.8.0\n          torch>=2.8.0+rocm6.3\n      we can conclude that pyannote-audio==4.0.2 depends on\n      torch==2.8.0+rocm6.3. (1)\n\n      Because there is no version of pytorch-triton-rocm{platform_machine ==\n      'x86_64' and sys_platform == 'linux'}==3.4.0 and torch==2.8.0+rocm6.3\n      depends on pytorch-triton-rocm{platform_machine == 'x86_64'\n      and sys_platform == 'linux'}==3.4.0, we can conclude that\n      torch==2.8.0+rocm6.3 cannot be used.\n      And because we know from (1) that pyannote-audio==4.0.2 depends on\n      torch==2.8.0+rocm6.3, we can conclude that pyannote-audio==4.0.2 cannot\n      be used.\n      And because your project depends on pyannote-audio==4.0.2 and your\n      project requires pyannote-diarization[dev], we can conclude that your\n      project's requirements are unsatisfiable."
time=22:08:52 level=ERROR msg="Failed to initialize transcription model" model_id=voxtral error="failed to setup Voxtral environment: uv sync failed: exit status 2: Using CPython 3.12.3 interpreter at: /usr/bin/python3\nCreating virtual environment at: .venv\nResolved 84 packages in 16.09s\nerror: Distribution `torch==2.0.1 @ registry+https://download.pytorch.org/whl/rocm6.3` can't be installed because it doesn't have a source distribution or wheel for the current platform"
time=22:08:55 level=ERROR msg="Failed to initialize transcription model" model_id=whisperx error="failed to sync WhisperX: uv sync failed: exit status 1: Downloading cpython-3.10.20-linux-x86_64-gnu (download) (28.5MiB)\n Downloaded cpython-3.10.20-linux-x86_64-gnu (download)\nUsing CPython 3.10.20\nCreating virtual environment at: .venv\n  × No solution found when resolving dependencies for split (markers:\n  │ python_full_version == '3.13.*' and platform_machine == 'x86_64' and\n  │ sys_platform != 'darwin'):\n  ╰─▶ Because there is no version of pytorch-triton-rocm{platform_machine ==\n      'x86_64' and sys_platform == 'linux'}==3.4.0 and torch==2.8.0+rocm6.3\n      depends on pytorch-triton-rocm{platform_machine == 'x86_64'\n      and sys_platform == 'linux'}==3.4.0, we can conclude that\n      torch==2.8.0+rocm6.3 cannot be used.\n      And because only the following versions of torch{platform_machine ==\n      'x86_64' and sys_platform != 'darwin'} are available:\n          torch{platform_machine == 'x86_64' and sys_platform !=\n      'darwin'}<2.8.0\n          torch{platform_machine == 'x86_64' and sys_platform !=\n      'darwin'}==2.8.0+rocm6.3\n          torch{platform_machine == 'x86_64' and sys_platform !=\n      'darwin'}>2.9.dev0\n      and your project depends on torch{platform_machine == 'x86_64' and\n      sys_platform != 'darwin'}>=2.8.0,<2.9.dev0, we can conclude that your\n      project's requirements are unsatisfiable.\n\n      hint: While the active Python version is 3.10, the resolution failed for\n      other Python versions supported by your project. Consider limiting your\n      project's supported Python versions using `requires-python`.\n\n      hint: `torch` was requested with a pre-release marker (e.g., all of:\n          torch>=2.8.0,<2.8.0+rocm6.3\n          torch>2.8.0+rocm6.3,<2.9.dev0\n      ), but pre-releases weren't enabled (try: `--prerelease=allow`)"

Seems to be python dependency version conflicts.