[0xShug0]

Summary

• Add session LoRA loading for Qwen3-TTS, Qwen3-ASR, VoxCPM1/2, Fish Audio S2 Pro, CosyVoice3, F5-TTS, and OuteTTS.
• Add model-aware GGUF adapter export for these families plus YuE2, ACE-Step, and VibeVoice.
• Separate checkpoint containers from adapter layouts using shared tensor I/O and checkpoint discovery; keep model-specific bindings and sidecars local.
• Document adapter scaling, required sidecars, and conversion. F5-TTS requires reconverting the matching original checkpoint with the current converter rather than using the currently published GGUF packages.

Validation

• Debug C++17 build and focused model-spec, weight-adapter, GGUF tensor-source, and Torch checkpoint tests.
• Python merge-tool tests, including fused-row coverage, native/PEFT package selection, and configured-rank validation.
• Local runtime and merge checks; malformed ranks and missing required native sidecars are rejected.

Adapters are opt-in; base weight files are not modified.