fix position embedding computation

stable_speech/modeling_stable_speech.py

@@ -136,7 +136,7 @@ class StableSpeechSinusoidalPositionalEmbedding(nn.Module):
 
     @torch.no_grad()
     def forward(self, input_ids: torch.Tensor, past_key_values_length: int = 0):
-        bsz, codebooks, seq_len = input_ids.size()
+        bsz, seq_len, _ = input_ids.size()
         # Create the position ids from the input token ids.
         position_ids = (torch.arange(seq_len) + past_key_values_length).to(input_ids.device)
         # expand embeddings if needed
@@ -754,7 +754,7 @@ class StableSpeechDecoder(StableSpeechPreTrainedModel):
             )
 
         # embed positions
-        positions = self.embed_positions(input, past_key_values_length)
+        positions = self.embed_positions(inputs_embeds, past_key_values_length)
 
         hidden_states = inputs_embeds + positions.to(inputs_embeds.device)