Make it work on mps

TTS/tts/layers/xtts/stream_generator.py

@@ -182,14 +182,44 @@ class NewGenerationMixin(GenerationMixin):
         accepts_attention_mask = "attention_mask" in set(inspect.signature(self.forward).parameters.keys())
         requires_attention_mask = "encoder_outputs" not in model_kwargs
 
-        if model_kwargs.get("attention_mask", None) is None and requires_attention_mask and accepts_attention_mask:
-            pad_token_tensor = torch.tensor([generation_config.pad_token_id], device=inputs_tensor.device) if generation_config.pad_token_id is not None else None
-            eos_token_tensor = torch.tensor([generation_config.eos_token_id], device=inputs_tensor.device) if generation_config.eos_token_id is not None else None
-            model_kwargs["attention_mask"] = self._prepare_attention_mask_for_generation(
-                inputs_tensor,
-                pad_token_tensor,
-                eos_token_tensor,
+        if (
+            model_kwargs.get("attention_mask", None) is None and requires_attention_mask and accepts_attention_mask
+        ):
+            pad_token_tensor = (
+                torch.tensor([generation_config.pad_token_id], device=inputs_tensor.device)
+                if generation_config.pad_token_id is not None
+                else None
             )
+            eos_token_tensor = (
+                torch.tensor([generation_config.eos_token_id], device=inputs_tensor.device)
+                if generation_config.eos_token_id is not None
+                else None
+            )
+
+            # hack to produce attention mask for mps devices since transformers bails but pytorch supports torch.isin on mps now
+            # for this to work, you must run with PYTORCH_ENABLE_MPS_FALLBACK=1 and call model.to(mps_device) on the XttsModel
+            if inputs_tensor.device.type == "mps":
+                default_attention_mask = torch.ones(inputs_tensor.shape[:2], dtype=torch.long, device=inputs_tensor.device)
+
+                is_pad_token_in_inputs = (pad_token_tensor is not None) and (
+                    torch.isin(elements=inputs_tensor, test_elements=pad_token_tensor).any()
+                )
+                is_pad_token_not_equal_to_eos_token_id = (eos_token_tensor is None) or ~(
+                    torch.isin(elements=eos_token_tensor, test_elements=pad_token_tensor).any()
+                )
+                can_infer_attention_mask = is_pad_token_in_inputs * is_pad_token_not_equal_to_eos_token_id
+                attention_mask_from_padding = inputs_tensor.ne(pad_token_tensor).long()
+
+                model_kwargs["attention_mask"] = (
+                    attention_mask_from_padding * can_infer_attention_mask
+                    + default_attention_mask * ~can_infer_attention_mask
+                )
+            else:
+                model_kwargs["attention_mask"] = self._prepare_attention_mask_for_generation(
+                    inputs_tensor,
+                    pad_token_tensor,
+                    eos_token_tensor,
+                )
 
         # decoder-only models should use left-padding for generation
         if not self.config.is_encoder_decoder:

requirements.txt

@@ -3,8 +3,9 @@ numpy==1.22.0;python_version<="3.10"
 numpy>=1.24.3;python_version>"3.10"
 cython>=0.29.30
 scipy>=1.11.2
-torch>=2.1
-torchaudio
+torch==2.3.1
+torchaudio==2.3.1
+torchvision==0.18.1
 soundfile>=0.12.0
 librosa>=0.10.0
 scikit-learn>=1.3.0
@@ -48,7 +49,7 @@ bnnumerizer
 bnunicodenormalizer
 #deps for tortoise
 einops>=0.6.0
-transformers>=4.33.0
+transformers>=4.41.2
 #deps for bark
 encodec>=0.1.1
 # deps for XTTS