fix(compose): drop CPU overlay AUDIO_STT_MODEL literal; make AMD memory limit env-driven

[yasinBursali]

What

Two related compose-overlay drift fixes:

• dream-server/docker-compose.cpu.yml: drop the now-empty open-webui: block (was overriding AUDIO_STT_MODEL to turbo)
• dream-server/docker-compose.amd.yml: replace literal memory: 110G with ${LLAMA_SERVER_MEMORY_LIMIT:-110G}

Why

CPU overlay: PR #985 made AUDIO_STT_MODEL env-driven (${AUDIO_STT_MODEL:-Systran/faster-whisper-base} in docker-compose.base.yml). The cpu overlay's residual literal AUDIO_STT_MODEL: "deepdml/faster-whisper-large-v3-turbo-ct2" overrode that interpolation, so on CPU-only Linux installs:

1. Phase 06 of the installer correctly writes AUDIO_STT_MODEL=Systran/faster-whisper-base to .env
2. Phase 12 pre-downloads only the base model
3. Open WebUI's container env is forced to turbo by the overlay literal
4. Whisper 404s on every transcription

The NVIDIA overlay's identical literal is intentional — for NVIDIA, Phase 06 also picks turbo, so the literal is a redundant safety net (per #985's design). Only the cpu overlay drifted.

AMD overlay: 5 of 6 GPU overlays use ${LLAMA_SERVER_MEMORY_LIMIT:-N} (NVIDIA 64G, Apple 32G, Intel/Arc 24G, CPU 6G). .env.schema.json registers the variable as a tunable string. The amd overlay's literal memory: 110G made AMD the only platform where users couldn't tune llama-server's memory cap.

tier0.yml's literal memory: 4G is a deliberate hard cap for <8 GB RAM hosts and was left alone.

How

• cpu.yml: removed the entire open-webui: block (3 + leading-blank lines). Container inherits the base's env-var interpolation.
• amd.yml: 1-line substitution. The 110G default is preserved as the AMD-specific fallback so existing AMD installs without the env var keep current byte-for-byte behavior.

Testing

• YAML parse, make lint, docker compose config --quiet, pre-commit: all PASS
• Functional substitution checks:
  • cpu env unset → resolves Systran/faster-whisper-base
  • cpu env override → propagates correctly
  • amd env unset → resolves 110G (fallback preserved)
  • amd env override LLAMA_SERVER_MEMORY_LIMIT=32G → resolves 32G
• Cross-overlay non-regression sweep (nvidia, intel, arc, apple, multigpu, tier0): all PASS

Manual:

• Linux CPU: docker exec dream-open-webui env | grep AUDIO_STT_MODEL should equal Systran/faster-whisper-base post-install
• Linux AMD: set LLAMA_SERVER_MEMORY_LIMIT=32G in .env, dream restart, then docker inspect dream-llama-server --format '{{.HostConfig.Memory}}' should equal 34359738368 (32 GiB) — was 118111600640 (110 GiB) before

Review

Independent verification confirmed both bugs are real and isolated to these two overlays; cross-overlay scan found no other drift on AUDIO_STT_MODEL or LLAMA_SERVER_MEMORY_LIMIT.

Known Considerations

• Schema default LLAMA_SERVER_MEMORY_LIMIT="64G" and AMD's compose fallback :-110G differ. This matches the codebase's existing pattern: schema default applies to installer env-generation; per-backend overlay fallback applies at compose-time when .env is silent. Same divergence already exists for nvidia/apple/intel/arc/cpu — not a regression.
• A potential follow-up (separate PR, not this one) would have installer phases write LLAMA_SERVER_MEMORY_LIMIT to .env per-backend so the schema default becomes load-bearing. That's a design decision, not a bug.

Platform Impact

• Linux CPU: STT now works on first install (was 404 on every transcription).
• Linux AMD: LLAMA_SERVER_MEMORY_LIMIT env var now honored (was silently ignored).
• Linux NVIDIA / Intel / Arc / multigpu / tier0: no behavior change.
• macOS Apple Silicon: no behavior change (llama-server.replicas: 0, no AUDIO_STT_MODEL override).
• Windows AMD: no behavior change (llama-server.replicas: 0, no override).

[Lightheartdevs]

Two correct surgical fixes: (1) drop CPU overlay's NVIDIA-specific AUDIO_STT_MODEL literal — wrong overlay; (2) make AMD memory limit env-driven via ${LLAMA_SERVER_MEMORY_LIMIT:-110G}. Both compose configs validate. Ship after rebase.

[Lightheartdevs]

Re-audited for merge. CPU overlay now inherits the base AUDIO_STT_MODEL, and AMD memory remains env-driven with the 110G fallback. Local CPU and AMD compose configs passed with required secret placeholders. Approving for squash merge.