req: support Chatterbox-Nano

[knguyen298]

https://huggingface.co/ResembleAI/chatterbox-nano

---

Looks like usage is almost identical to ChatterboxTurboTTS, save for a nano=True flag.

Aside from wiring it up in this repo, the chatterbox fork is missing the model code:

• models/t3/llama_configs.py has no GPT2_SMALL_CONFIG and no "GPT2_small" registry entry.
• tts_turbo.py has no nano argument, no NANO_REPO_ID, and no t3_nano_v1.safetensors branch.

Both exist in resemble-ai/chatterbox@master. Diffing the two, the only divergence in tts_turbo.py is the Nano support plus a Xet download fallback, so upstream's copy of these two files should drop in as-is.

[nomonkeynodeal]

I found this post a few hours ago, asked DeepSeek to do it, and I am now running Chatterbox Nano. Not submitting a PR because I have no idea what I'm doing, but you can pass this to your coding agent. I use Docker myself.

What was done: adding Chatterbox-Nano support
Project: Chatterbox-TTS-Server (Docker, plain docker build/run, onboard/unRAID). Engine pinned to devnen/chatterbox-v2.git@master at image-build time.
The problem
Nano (110M) shares the same ChatterboxTurboTTS class as Turbo, but the pinned engine fork predates Nano — it couldn't load it. Two things were missing:

1. The ChatterboxTurboTTS loader had no nano flag (no ResembleAI/chatterbox-nano repo, no nano=True, no t3_nano_v1.safetensors checkpoint selection).
2. The fork's T3 model only defined LLAMA_CONFIGS = {Llama_520M, GPT2_medium} — no GPT2_small config (Nano's 110M backbone). Selecting Nano caused KeyError: 'GPT2_small'.
   The fix: build-time patch script (docker_patch_nano.py)
   A new script runs only inside the Docker build (a RUN step in Dockerfile, Dockerfile.cu128, Dockerfile.cu130), right after the engine is installed. It rewrites the installed package in two files:
3. chatterbox/tts_turbo.py — backports upstream Nano support:

• Adds NANO_REPO_ID = "ResembleAI/chatterbox-nano".
• Threads a nano flag through init/from_local/from_pretrained.
• When nano=True: uses GPT2_small config, loads t3_nano_v1.safetensors, sets model_label="Nano", and downloads from the Nano HF repo.
• Verified diff matches upstream's Nano implementation exactly.

2. chatterbox/models/t3/llama_configs.py — the actual missing piece:

• Adds the GPT2_SMALL_CONFIG dict.
• Registers "GPT2_small" in LLAMA_CONFIGS.
• The config is byte-for-byte identical to upstream's (hidden_size=768, 12 layers, 12 heads, vocab 50276), and the fork's T3 architecture matches upstream — so the Nano checkpoint loads cleanly.
  The script is idempotent (skips already-patched files) and handles both chatterbox and chatterbox_tts package dir names (same detection pattern start.py uses).
  Server + UI wiring
• engine.py: added nano as a first-class model type — selector map (chatterbox-nano/nano/ResembleAI/chatterbox-nano), class resolution via ChatterboxTurboTTS, from_pretrained(..., nano=True), Nano treated like Turbo for paralinguistic tags and the voice-conditioning cache.
• ui/index.html: added "Chatterbox Nano (Fastest, English)" to the model dropdown.
• ui/script.js: handles type === 'nano' — Nano badge, dropdown sync, paralinguistic tags section, hides exaggeration/CFG (Turbo-style), keeps Turbo presets available.
  Verification
• Patched files compile; GPT2_small config loads through GPT2Config (CPU-only, no VRAM).
• Patched llama_configs.py matches upstream GPT2_SMALL_CONFIG exactly.
• Idempotency confirmed (re-run skips).
  To deploy
  Copy the updated docker_patch_nano.py to the project folder (Dockerfile is already updated), then rebuild the image (docker build + docker run, same as before). Pick "Chatterbox Nano" in the dropdown and apply.