CoolFace
Modelpublic

owlninjam/qwen3-tts-1.7b-customvoice-fp16

sourceHugging Faceapache-2.0updated 7mo agoView on Hugging Face
0likes133downloads
Model Card

qwen3-tts-1.7b-customvoice-fp16

FP16 conversion of Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice.

Changes from original

  • —All bfloat16/float32 weights cast to float16
  • —All config.json files updated: torch_dtype: "float16"
  • —Everything else (tokenizer, architecture) unchanged

Why?

GTX 1650 and other Turing GPUs don't support bfloat16 natively. Pre-converted fp16 weights load directly without runtime casting.

Usage

python
from qwen_tts import Qwen3TTSModel
model = Qwen3TTSModel.from_pretrained("owlninjam/qwen3-tts-1.7b-customvoice-fp16", dtype=torch.float16)