model: add IndexTTS-2.5 family

[IIIIIllllIIIIIlllll]

Just for test now.

Add support for IndexTTS-2.5 (GPT + DiT CFM + BigVGAN, zh/en/ja/es/ar) as a new model family alongside index_tts2:

The architecture of IndexTTS-2.5 has many similarities to IndexTTS-2, but I created a new architecture to avoid interfering with the existing code.

GGUF is also supported (GGUF weights are not uploaded).

Test on my 2080TI gfx1151 CPU

The generated content was consistent with the official deployment method.

About convert:

python tools/convert_index_tts2_5.py \
    --model-dir /path/to/IndexTTS-2.5 \
    --output-dir /path/to/staging \
    --native-dir /path/to/IndexTTS-2.5-native \
    --run-converter /path/to/gguf_file --type q8_0

AI usage: Kimi k3 developed this implementation based on the original project.

Side note (not fixed in this PR): the existing index_tts2 family produces noise on the CPU backend, same root cause as the 2.5 fix here — src/models/index_tts2/s2mel.cpp builds the wavenet zero accumulator via sub(ctx, input_bct, input_bct) where input_bct is a permuted view; ggml's CPU binary-op kernels only check src1 row-contiguity and silently misread a permuted src0 (asserts are compiled out under NDEBUG; CUDA/HIP kernels handle it, so only CPU is affected). The same one-line materialize-before-sub fix applies. A more thorough fix belongs in ggml: check row contiguity for both operands in the binary-op path (add/sub/mul/div family).

[IIIIIllllIIIIIlllll]

@0xShug0 This is a real big project. I'd appreciate it if you could do a quick check to see which parts are reusable. Thanks!

[0xShug0]

@IIIIIllllIIIIIlllll Thanks for the work! Based on your code, I think IndexTTS2.5 is close enough to IndexTTS2 that we should extend the existing IndexTTS2 implementation additively instead of copying the whole model into a new family.

The main 2.5 differences seem to be variant-level pieces: multilingual tiktoken/BPE tokenizer + language embedding, CAMPPlus-projected GPT speaker conditioning, changed GPT prefill condition layout, and semantic-codec 2x upsample + up conv. Most other files are identical or near-identical to IndexTTS2 after renaming. Most request/session options seem to be shared with existing IndexTTS2.

Could you refactor this as an IndexTTS2 variant rather than adding copied index_tts2_5 files? Suggested shape:

1. Keep shared audio features, wav2vec2bert, qwen emotion, style encoder, BigVGAN vocoder, and most S2Mel code in the current IndexTTS2 implementation.

2. Add a variant-specific text tokenizer behind one common IndexTTS2 tokenizer interface:

   • v2 keeps the current IndexTTS2 tokenizer behavior.
   • v2.5 uses the multilingual tiktoken/BPE tokenizer, language special tokens, and language-id handling.

3. Split GPT speaker-conditioning into explicit variant modes:

   • v2 mode should preserve the current IndexTTS2 GPT speaker-conditioning path: use speaker semantic features with the existing conditioning_encoder + perceiver_encoder.
   • v2.5 mode loads and uses spk_emb_proj with the CAMPPlus speaker/style embedding, plus lang_embedding during text prefill.
   • Keep the shared GPT decode/cache logic in one place.

4. Extend semantic codec with an explicit variant option:

   • v2 keeps the current code embedding path.
   • v2.5 applies the extra 2x nearest upsample and up convolution after the decoder projection.
   • The branch should be selected from model config/variant, not by probing tensor names in hot paths.

5. Add 2.5 packages/spec entries under the existing IndexTTS2 family/variant surface.

6. Remove debug-only production code such as dump paths, injected noise files, and environment-variable parity hooks before PR is ready.

This should keep the implementation smaller, reduce maintenance cost, and avoid two copies drifting independently.