LLM inference in C/C++
C++ ggml
qwen3tts : guard speaker_encoder_config patch for CustomVoice variant (#29179)
SIDDARTHA REDDY48499d2e1c86
diff --git a/conversion/qwen3tts.py b/conversion/qwen3tts.pyindex 2c35799f77..cdb1e0c9b3 100644--- a/conversion/qwen3tts.py+++ b/conversion/qwen3tts.py@@ -218,8 +218,10 @@ class Qwen3TTSSpeakerEncoderModel(MmprojModel):if hparams is None:hparams = ModelBase.load_hparams(dir_model, is_mistral_format=False)hparams["text_config"] = {"hidden_size": hparams["talker_config"]["hidden_size"]}- # ECAPA-TDNN has a fixed 4-stage backbone, but MmprojModel.__init__ needs a n_block_keys- hparams["speaker_encoder_config"]["n_layers"] = 4+ # ECAPA-TDNN has a fixed 4-stage backbone, but MmprojModel.__init__ needs a n_block_keys.+ # The CustomVoice variant ships no speaker encoder, so its config lacks this key entirely.+ if "speaker_encoder_config" in hparams:+ hparams["speaker_encoder_config"]["n_layers"] = 4super().__init__(dir_model, *args, hparams=hparams, **kwargs)self._wav_config_cache = None