acestep t2m

diffsynth/models/ace_step_conditioner.py

@@ -540,17 +540,9 @@ class AceStepTimbreEncoder(nn.Module):
     ) -> BaseModelOutput:
         inputs_embeds = refer_audio_acoustic_hidden_states_packed
         inputs_embeds = self.embed_tokens(inputs_embeds)
-        # Handle 2D (packed) or 3D (batched) input
-        is_packed = inputs_embeds.dim() == 2
-        if is_packed:
-            seq_len = inputs_embeds.shape[0]
-            cache_position = torch.arange(0, seq_len, device=inputs_embeds.device)
-            position_ids = cache_position.unsqueeze(0)
-            inputs_embeds = inputs_embeds.unsqueeze(0)
-        else:
-            seq_len = inputs_embeds.shape[1]
-            cache_position = torch.arange(0, seq_len, device=inputs_embeds.device)
-            position_ids = cache_position.unsqueeze(0)
+        seq_len = inputs_embeds.shape[1]
+        cache_position = torch.arange(0, seq_len, device=inputs_embeds.device)
+        position_ids = cache_position.unsqueeze(0)
 
         dtype = inputs_embeds.dtype
         device = inputs_embeds.device
@@ -586,9 +578,8 @@ class AceStepTimbreEncoder(nn.Module):
             hidden_states = layer_outputs[0]
 
         hidden_states = self.norm(hidden_states)
+        hidden_states = hidden_states[:, 0, :]
         # For packed input: reshape [1, T, D] -> [T, D] for unpacking
-        if is_packed:
-            hidden_states = hidden_states.squeeze(0)
         timbre_embs_unpack, timbre_embs_mask = self.unpack_timbre_embeddings(hidden_states, refer_audio_order_mask)
         return timbre_embs_unpack, timbre_embs_mask
 
@@ -686,7 +677,7 @@ class AceStepConditionEncoder(nn.Module):
         text_attention_mask: Optional[torch.Tensor] = None,
         lyric_hidden_states: Optional[torch.LongTensor] = None,
         lyric_attention_mask: Optional[torch.Tensor] = None,
-        refer_audio_acoustic_hidden_states_packed: Optional[torch.Tensor] = None,
+        reference_latents: Optional[torch.Tensor] = None,
         refer_audio_order_mask: Optional[torch.LongTensor] = None,
     ):
         text_hidden_states = self.text_projector(text_hidden_states)
@@ -695,11 +686,7 @@ class AceStepConditionEncoder(nn.Module):
             attention_mask=lyric_attention_mask,
         )
         lyric_hidden_states = lyric_encoder_outputs.last_hidden_state
-        timbre_embs_unpack, timbre_embs_mask = self.timbre_encoder(
-            refer_audio_acoustic_hidden_states_packed,
-            refer_audio_order_mask
-        )
-
+        timbre_embs_unpack, timbre_embs_mask = self.timbre_encoder(reference_latents, refer_audio_order_mask)
         encoder_hidden_states, encoder_attention_mask = pack_sequences(
             lyric_hidden_states, timbre_embs_unpack, lyric_attention_mask, timbre_embs_mask
         )

diffsynth/models/ace_step_dit.py

@@ -165,7 +165,7 @@ class TimestepEmbedding(nn.Module):
         self,
         in_channels: int,
         time_embed_dim: int,
-        scale: float = 1000,
+        scale: float = 1,
     ):
         super().__init__()
 
@@ -711,7 +711,7 @@ class AceStepDiTModel(nn.Module):
         encoder_hidden_states: torch.Tensor,
         encoder_attention_mask: torch.Tensor,
         context_latents: torch.Tensor,
-        use_cache: Optional[bool] = None,
+        use_cache: Optional[bool] = False,
         past_key_values: Optional[EncoderDecoderCache] = None,
         cache_position: Optional[torch.LongTensor] = None,
         position_ids: Optional[torch.LongTensor] = None,

diffsynth/models/ace_step_text_encoder.py

@@ -2,17 +2,6 @@ import torch
 
 
 class AceStepTextEncoder(torch.nn.Module):
-    """
-    Text encoder for ACE-Step using Qwen3-Embedding-0.6B.
-
-    Converts text/lyric tokens to hidden state embeddings that are
-    further processed by the ACE-Step ConditionEncoder.
-
-    Wraps a Qwen3Model transformers model. Config is manually
-    constructed, and model weights are loaded via DiffSynth's
-    standard mechanism from safetensors files.
-    """
-
     def __init__(
         self,
     ):
@@ -49,8 +38,6 @@ class AceStepTextEncoder(torch.nn.Module):
         )
 
         self.model = Qwen3Model(config)
-        self.config = config
-        self.hidden_size = config.hidden_size
 
     @torch.no_grad()
     def forward(
@@ -58,23 +45,9 @@ class AceStepTextEncoder(torch.nn.Module):
         input_ids: torch.LongTensor,
         attention_mask: torch.Tensor,
     ):
-        """
-        Encode text/lyric tokens to hidden states.
-
-        Args:
-            input_ids: [B, T] token IDs
-            attention_mask: [B, T] attention mask
-
-        Returns:
-            last_hidden_state: [B, T, hidden_size]
-        """
         outputs = self.model(
             input_ids=input_ids,
             attention_mask=attention_mask,
             return_dict=True,
         )
         return outputs.last_hidden_state
-
-    def to(self, *args, **kwargs):
-        self.model.to(*args, **kwargs)
-        return self

diffsynth/models/ace_step_vae.py

@@ -226,6 +226,7 @@ class AceStepVAE(nn.Module):
             upsampling_ratios=upsampling_ratios,
             channel_multiples=channel_multiples,
         )
+        self.sampling_rate = sampling_rate
 
     def encode(self, x: torch.Tensor) -> torch.Tensor:
         """Audio waveform [B, audio_channels, T] → latent [B, encoder_hidden_size, T']."""