fix(extensions): create AMD GPU overlay files for 5 services

[yasinBursali]

What

Create compose.amd.yaml overlay files for 5 services that support AMD GPUs per their manifests, providing ROCm device passthrough.

Why

These services declare gpu_backends: [amd, nvidia] in their manifests but had no AMD overlay — meaning AMD GPU users could not utilize GPU acceleration.

How

Created compose.amd.yaml for each service with:

• /dev/dri and /dev/kfd device passthrough (ROCm)
• group_add for video/render groups
• HSA_OVERRIDE_GFX_VERSION environment variable

Services

Service	Image swap	Notes
ollama	No	Official image supports AMD natively
text-generation-webui	Yes → default-rocm-v4.0 (SHA pinned)	Base uses NVIDIA-specific image tag
xtts	No	Device passthrough only; ROCm compatibility unverified
invokeai	No	Official image has ROCm support
rvc	No	Device passthrough only; ROCm compatibility unverified

Excluded services (with rationale)

• fooocus: runpod/fooocus:2.5.3 is CUDA-only, no ROCm variant exists
• immich: GPU accel requires immich-machine-learning sidecar which is not yet in the extensions library compose

Known limitations

• rvc and xtts: Base Docker images may be CUDA-built. AMD device passthrough is provided but GPU acceleration may silently fall back to CPU if the image lacks ROCm PyTorch. Should be verified before production promotion.

Merge Order

#379-#383 (manifest fixes) → #385 (NVIDIA extraction) → THIS PR → #387 (structural fixes)

This PR MUST merge after #385 (NVIDIA extraction). Three base compose files (ollama, text-gen-webui, rvc) currently contain driver: nvidia device reservations. #385 removes those. If this PR merges first, Docker Compose would merge both NVIDIA and AMD device entries, causing runtime errors on AMD systems.

Testing

• docker compose config validation passed for all 5 services (base + AMD overlay)
• Service IDs verified matching base compose.yaml keys
• text-gen-webui uses correct default-rocm-v4.0 tag (verified exists on Docker Hub) with SHA pin
• Critique Guardian: APPROVED WITH WARNINGS (image compatibility for rvc/xtts noted above)

Platform Impact

• AMD: GPU acceleration now available for 5 services
• NVIDIA: No change (separate overlay)
• macOS / CPU-only: No change (overlay not applied)

[Lightheartdevs]

Approve — Correct AMD GPU overlay files.

What's good

• Creates compose.amd.yaml for 5 services: invokeai, ollama, rvc, text-generation-webui, xtts
• AMD pattern is correct: /dev/dri + /dev/kfd device passthrough, VIDEO_GID/RENDER_GID group adds, HSA_OVERRIDE_GFX_VERSION env var
• text-generation-webui AMD overlay correctly swaps to the ROCm image variant (atinoda/text-generation-webui:default-rocm-v4.0)
• Follows established pattern from core stack's docker-compose.amd.yml

Merge after #385 (which extracts NVIDIA overlays from the same base compose files).

[Lightheartdevs]

Approve — Correct AMD GPU overlay files.

Creates compose.amd.yaml for 5 services (invokeai, ollama, rvc, text-generation-webui, xtts) with the correct AMD pattern:

• /dev/dri + /dev/kfd device passthrough
• VIDEO_GID/RENDER_GID group adds with sensible defaults
• HSA_OVERRIDE_GFX_VERSION env var for GPU compatibility
• text-generation-webui correctly swaps to the ROCm image variant

Follows established pattern from core stack's docker-compose.amd.yml. LGTM. Merge after #385.