Python
fix for profile 4 with Ace Step 1.5
DeepBeepMeepc597459a1afc
diff --git a/README.md b/README.mdindex e277310..99e053b 100644--- a/README.md+++ b/README.md@@ -38,10 +38,14 @@ WanGP supports the Wan (and derived models) but also Hunyuan Video, Flux, Qwen,- **Ace Step 1.5 Turbo**: this long waited open source project claims to have overthrown *Suno 5*. It lets you generate high multi minutes quality songs. It comes in four flavours: *Vanilla* (No Language Model Preprocessing, **4s Generation Time!!!**) & *3 levels of LM Preprocessing* for a higher Quality (and increasing VRAM requirements)+Please note that when using the *Ace Step LM* variants, this may get very slow with *Memory Profiles 2 or 4* since the LM is an *Autoregressive Model*. It is why I recommed to stick to *Memory Profiles 1/3/3+* unless you have very little VRAM.++- **Kugel Audio 0**: another *TTS* with *Voice Cloning*, this one claims to outperform *ElevenLabs* !!! The nice thing about Kugel Audio is that it can be used to create Dialogues between two cloned voices. Have Fun !-- **LTX-2 Self Refiner**: WanGP exclusive *Self Refiner* has been added to *Distilled/Non Distilled* models, so hopefully this will improve the quality of our Video Gens.+Kugel Audio is entirely an *Autoregressive Model* and quite VRAM Hungry. So either you've got 16GB VRAM and you can run it with *Memory Profile 1/3/3+* or you will have to go the slow way with other Profiles.+- **LTX-2 Self Refiner**: WanGP exclusive *Self Refiner* has been added to *Distilled/Non Distilled* models, so hopefully this will improve the quality of our Video Gens.### February 1st 2026: WanGP v10.61, Upgrade Time !diff --git a/models/TTS/ace_step15/models/modeling_acestep_v15_turbo.py b/models/TTS/ace_step15/models/modeling_acestep_v15_turbo.pyindex 8deefdf..7ab010f 100644--- a/models/TTS/ace_step15/models/modeling_acestep_v15_turbo.py+++ b/models/TTS/ace_step15/models/modeling_acestep_v15_turbo.py@@ -1637,7 +1637,13 @@ class AceStepConditionGenerationModel(AceStepPreTrainedModel):lm_hints_25Hz = precomputed_lm_hints_25Hz[:, :src_latents.shape[1], :]else:if audio_codes is not None:+ restore = False+ if self.tokenizer.quantizer.layers[0].implicit_codebook.device.type=="cpu":+ self.tokenizer.quantizer.layers.to("cuda")+ restore = Truelm_hints_5Hz = self.tokenizer.quantizer.get_output_from_indices(audio_codes)+ if restore:+ self.tokenizer.quantizer.layers.to("cpu")else:lm_hints_5Hz, indices, llm_mask = self.tokenize(hidden_states, silence_latent, attention_mask)lm_hints_25Hz = self.detokenize(lm_hints_5Hz)