Add VibeVoice 7B support and decoder LoRA merging

[justinjohn0306]

Support the VibeVoice 7B model alongside the 1.5B:

• Add the vibevoice_7b model-manager package, reusing the shared Qwen2.5 tokenizer bundle.
• Handle the 7B config.json, tolerating its upstream "acostic_vae_dim" key and reading top-level tie_word_embeddings.
• Load a separate lm_head.weight when word embeddings are untied and bind the decoder logits head to it.

Add optional PEFT LoRA merging for the decoder:

• Merge lora_A/lora_B into the targeted linear weights at load time via a tensor-source decorator, so it composes with the weight-type quantization options at no per-step cost.
• Expose it through the vibevoice.lora / vibevoice.lora_scale load options and document it in the README and docs/tts.md.

[0xShug0]

@justinjohn0306 Thank you for your PR! Could you add extra logs to confirm the LoRA loading is woring with 7B model and share the run log?

Reviewed by Codex:
docs/tts.md:392 and README news say to use --load-option vibevoice.lora, but the implementation only reads vibevoice.lora from SessionOptions in src/models/vibevoice/session.cpp:34. The CLI stores --load-option separately in load_request.options, while session_options.options only receives --session-option. Since the VibeVoice loader does not consume load_request.options, users following the docs will run without any LoRA merge and get no error

[justinjohn0306]

LoRA loading confirmed on VibeVoice-7B + --load-option fix

Thanks for the review — addressed both points, and extended the adapter support to match infer.py.

1. --load-option vibevoice.lora now actually merges (Codex finding)

Previously the adapter was only read from SessionOptions, so the documented --load-option path silently no-op'd. The loader now consumes request.options at load time via a shared apply_vibevoice_finetune_options(...) helper, and the same helper still handles --session-option for backward compatibility. A guard rejects passing it through both at once. So docs/tts.md / README (--load-option vibevoice.lora=<dir>) are now correct.

2. Merge logging + 7B run log

Added --log output confirming each component. Running VibeVoice-7B with the fine-tune via --load-option:

[info][vibevoice] applying fine-tune adapter: ...\finetune_mp1\lora
[info][vibevoice] LM LoRA: merged 196 decoder modules (scale 4.000000)
[info][vibevoice] diffusion head: overrode 26 tensors
[info][vibevoice] acoustic connector: overrode 5 tensors
[info][vibevoice] semantic connector: overrode 5 tensors

3. Confirmed it changes the output toward the fine-tuned speaker

A/B with a fixed seed and identical reference audio — base weights vs the fine-tune adapter — the two outputs correlate 0.956 (not identical), difference RMS ≈ 30% of the signal: same words/prosody, but the voice timbre shifts onto the target speaker. It audibly matches the reference and the Python infer.py output with the same adapter.

Bonus: full fine-tune overlay (not just the LM LoRA)

To mirror infer.py's apply_lora, the overlay now applies all four trained components when present in the adapter dir:

• language-model LoRA → delta-merged into the decoder linears
• fine-tuned diffusion head → replaces model.prediction_head.*
• acoustic + semantic connectors → replace model.{acoustic,semantic}_connector.*

The connector / head .bin files are read by a new pure-C++ Torch-pickle reader (torch_bin.*) — no Python at inference or preprocessing. It's parity-checked bit-for-bit against torch.load (torch_bin_parity), and an overlay check (vibevoice_finetune_overlay_check) confirms the merged tensors equal the adapter's and that the merge reaches the forward pass (verified: set_backend_tensor → require_tensor → require_tensor_data, which the overlay overrides).

Docs and README updated to describe the full fine-tune behavior.

[0xShug0]

Thanks! I will do a final test tomorrow and then merge it.

[0xShug0]

@justinjohn0306 Code merged! Thanks a lot!

[0xShug0]

@justinjohn0306 I did some changes to improve the lora performance. Just pushed acde132.

Metric	Before	After	Improvement
LoRA decoder merge	25,922 ms	11,693 ms	54.9% faster, 2.22x
Decoder weights load	39,001 ms	14,463 ms	62.9% faster, 2.70x
Output WAV	baseline	byte-identical	no change